Read also : Recap NetApp Insight 2026 Key news, strategic for the future of NetApp Top News | By Philippe Nicolas | October 6, 2026 | News
GL iNet Remote KVM GL-RMQ1-Screen 1 The GL.iNet Comet Q GL-RMQ1 is a new class of remote KVM devices, and it has something really neat. It is designed to be a single-cable solution. Plug a USB Type-C cable into a device with a Type-C port that supports video, and that is the only cable you need. WiFi handles network connectivity, and the built-in Type-C cable powers the unit, carries the video feed, and sends keyboard, mouse, and remote media to the remote system. One of the neat functions is that, along with being able to control many PCs, you can also connect phones and tablets using this solution. Here is an Amazon affiliate link to where we bought ours. GL.iNet Comet Q GL-RMQ1 Hardware Overview The unit itself has a large 1.8″ touchscreen on its face, but it is like a puck design rather than many of the KVM units that are very boxy. What it is not, however, is small. GL iNet Remote KVM GL-RMQ1-Screen 1 On the rear of the unit, you can see cooling vents, rubber feet, and an integrated USB Type-C cable. GL iNet Remote KVM GL-RMQ1-Bottom 1 That integrated Type-C cable is not huge, but it is nice to grab the unit and have the cable ready. It does, however, mean that if you wanted a Type-A connection, or a Type-A and HDMI, for example, the convenience is not there. GL iNet Remote KVM GL-RMQ1-USB Type-C Connector Cable 1 We plugged it into the Bosgame M5 since that was next to the step, and it is something you can see that is really important here. If you plug it into a system, then it has to boot. If you go from off to on, some systems boot faster than the GL-RMQ1, so that makes it hard to get into the BIOS. GL iNet Remote KVM GL-RMQ1-Screen 2 Once it boots, the little touchscreen is very useful. You can see the IP address, along with the KVM and video connectivity status. You can also set up the unit, for example, by selecting a WiFi SSID and entering a password. Once the initial setup is done, you can then manage it via a browser. GL iNet Remote KVM GL-RMQ1-Screen 3 Something important is that there is another Type-C port. This is a power input, which makes sense since many devices you connect to only have a single Type-C port. GL iNet Remote KVM GL-RMQ1-USB Type-C Port 1 Then there is one more accessory in the box, a little protector that is bright orange. We bought two of these and got the same color. Our guess is that this will eventually get a Sharpie marker on it to help quickly label these with a hostname so we know which is which. GL iNet Remote KVM GL-RMQ1-Rubber Case 1 Here is the protector installed. GL iNet Remote KVM GL-RMQ1-Rubber Case 2 Next, let us get to using the GL.iNet Comet Q.
Microsoft Surface Laptop Ultra-Front Angled 3 Amid the many trends and stories shaping the hardware world in 2026, the final quarter shows a confluence of several influences. AI continues to loom over everything this year, both client and server, and 2026 picks up where 2025 left off, with an even heavier emphasis on inference. Meanwhile, Microsoft’s latest efforts to bootstrap its Windows-on-Arm ecosystem are making good progress, thanks to another generation of improvements to its x86-on-Arm Prism emulator. And finally, amid all this, the client hardware market is trying to make sense (and cents) of a run-up in DRAM and NAND prices. As we enter the fall, these trends conveniently converge in a single device: Microsoft’s Surface Laptop Ultra. The company’s latest-generation flagship Surface laptop, the Ultra, is unlike anything that came before it, with Microsoft going above and beyond to develop a high-end laptop that, well, can go above and beyond itself. All the while, they are staking the fate of their new flagship laptop on the Arm architecture, which is still finding its footing in the Windows world. As a result, the Surface Laptop Ultra is a culmination of several firsts for Microsoft. As a complete product, it is the company’s most direct take on (or competitor to) Apple’s MacBook Pro lineup yet, making a direct play on the high-end market that Apple has captured so much of over the last several years. Meanwhile, this is one of the first-generation Windows laptops to ship with NVIDIA’s RTX Spark SoC, the consumer branding for their high-performance GB10 SoC that we have seen in numerous developer-focused mini-PCs over the last year. Finally, it is one of Microsoft’s biggest hardware releases for the developer market in at least the last several years, as Microsoft seeks to further bootstrap developers with all the tools they need to make the next generation of Windows applications. Microsoft Surface Laptop Ultra Key Specs (As Configured) Processor NVIDIA RTX Spark N1X 20-core Operating System Windows 11 Pro Memory 128GB LPDDR5X Storage 1TB SSD (M.2 2280, PCIe Gen5) GPU Options NVIDIA Integrated Blackwell RTX 5120 GPU core Display 15-Inch PixelSense, 3270×2180 120Hz, 2000nits Max Brightness, 100% DCI, Touchscreen Battery 92Wh Form Factor Laptop Dimensions 328.8 x 238.7 x 19.16 mm (12.95 x 9.40 x 0.75 in) Weight 2.0 kg (4.41 lbs) Wireless Wi-Fi 7 + Bluetooth 5.4 Colors Platinum (Gray) Nightfall (Black) Ports Left: 1x HDMI 2.1b, 2x USB-C 40Gbps, 3.5mm Combo Audio Jack Right: 1x USB-C 40Gbps (Magnetic Connect), 1x USB-A 10Gbps, SD Card Reader As a result, although this is far from Microsoft’s first Surface device, first Arm-on-Windows device, or even their first productivity laptop, the Surface Laptop Ultra is potentially the most important device launch from them in years. If it succeeds, it will shape the Surface ecosystem for years to come. If it does not, it will end up in the footnote pile next to 2012’s Surface RT (which we used at STH). The Surface Laptop Ultra is not meant to (nor will it) be a game changer for the broader laptop market, but it is still a significant launch for Microsoft’s hardware division. Meanwhile, for partner/supplier NVIDIA, although this is far from the first GB10 device on the market, the Ultra is one of a wave of laptops that marks the first time the chip has made it out of a mini-PC and into a laptop. NVIDIA has wanted to break into the market as a CPU/SoC supplier for years. This is their first chance in several years to do that. But how well will the N1X sibling to the GB10 translate into powering a Windows laptop? That is one of the questions we will have to answer today. Quick note: You may notice Ryan Smith wrote the first half of this review, and Patrick wrote the second half. Again, we had a week, and we were at OCP 2026 all week covering that show. So without any further ado, let us jump right in. Microsoft Surface Laptop Ultra External Hardware Overview When it comes to the Surface Laptop Ultra’s design, there is no dancing around the fact that this is meant to be Microsoft’s answer to the MacBook Pro. So we can only address it head-on. Microsoft wants to compete with Apple’s flagship laptops, and this is what they think a Windows equivalent should look and behave like. Microsoft Surface Laptop Ultra-Front 1 Microsoft is no stranger to designing laptops, of course. Their mainstream Surface laptops are already up to their 8th generation, and the Surface Pro 2-in-1s are too. But the company needs to step up its game with the Ultra. They are building a premium laptop, packing a more powerful SoC, and charging a higher price than ever. The 15-inch laptop is meant to straddle the line between being a productivity laptop and a portable laptop. It is a bit too small to be a desktop replacement-class laptop, which means it is reasonably portable. But it is not a thin and light ultraportable like the smaller mainstream Surface Laptops. Instead, at 19mm thick and weighing 2kg on the dot, this laptop fits the developer-on-the-go mindset. Microsoft Surface Laptop Ultra-Rear 1 Microsoft is offering a few different configurations of the Surface Laptop Ultra. Most notably, a choice between 18- and 20-core versions of NVIDIA’s RTX Spark N1X processor. Regardless of configuration, there is only one body and one size available. This is not a laptop lineup bifurcated by size, at least not yet. Surface-Laptop-Ultra-Ports-Side-by-side The body itself is made from aluminum, which is pretty standard fare these days for premium laptops. Microsoft offers it in either a “Platinum” (grey) finish, as seen with our review sample, or a “Nightfall” (black) finish. Many configurations are available in both colors, but on Microsoft’s pre-order website, Nightfall seems to be treated as the more premium option, since you cannot currently order a laptop with more than 32GB of RAM in the Platinum finish. Microsoft Surface Laptop Ultra-Front 2 All Surface Laptop Ultras ship with the same display. This is a 15-inch high-DPI 3270×2180 “PixelSense” touchscreen LCD display. According to Microsoft, the HDR-capable display can reach an eye-searing max brightness of 2000 nits and cover the full DCI-P3 color space. For developers who need a bit more fluidity (or want to sneak in some gaming), the display also supports variable refresh rates up to 120Hz. Since this is a premium laptop, Microsoft color-calibrates the panels individually. It is not an OLED, but our STH studio team members who usually use MacBook Pros all thought the display looked sharp. If you really want an OLED for those deep blacks, other manufacturers are making RTX Spark laptops with other screens. Still, this is a touchscreen, which can be useful for many. On the bottom deck of the laptop, we have Microsoft’s keyboard and haptic trackpad. As is typical for anything above budget laptops these days, this is a fully backlit keyboard. Microsoft is not duplicating Apple’s fingerprint sensor functionality here (relying instead on Windows Hello), so it is a keyboard that is just a keyboard. Meanwhile, according to Microsoft, the Ultra’s touchpad is their largest to date, larger than what is on the mainstream 15-inch Surface Laptop. And while you can just barely spot it due to the very thin bezel around the LCD panel, Microsoft has embedded a full camera suite along the top edge of the display. At the center is a 1080p “Surface Studio” camera. Meanwhile, IR cameras flank it on either side to support Windows Hello. For audio, Microsoft built in four downward-firing “OmniSonic” speakers, with two on the left front and two on the right front. Microsoft supports Dolby Atmos here. For users who need more enclosed audio, Microsoft explicitly confirms the laptop supports Bluetooth LE Audio, a feature that is still pretty rare on Windows laptops. Meanwhile, for audio input, there are a pair of microphones built into the body to allow for voice focusing. Microsoft Surface Laptop Ultra-Side 1 Meanwhile, when it comes time to plug something into the Surface Laptop Ultra, Microsoft has taken a somewhat restrained approach to the number of available ports. On the left side of the laptop, you will find an HDMI 2.1b port, two 40Gbps USB-C ports, and a 3.5mm combo audio jack. Surprisingly, Microsoft has not opted to secure Thunderbolt 4 certification despite the laptop’s premium focus. Still, at 40Gbps with full DP Alt Mode and 140W charging, these are highly capable USB-C ports. At present, we are still trying to suss out who is providing the controller driving them, as the RTX Spark SoC itself does not support 40Gbps USB speeds. Microsoft Surface Laptop Ultra-Side 2 As for the right side of the laptop, things get a bit more interesting. On this side we find 3 ports: a full-size SD card reader, a 10Gbps USB-A port, and a special 40Gbps USB-C port that Microsoft calls its Magnetic Connect port. Microsoft Surface Laptop Ultra-Magnetic USB Type-C Charging Cable 2 Replacing Microsoft’s proprietary Surface Connect port from past designs, the company has developed a magnetic USB-C port; Microsoft’s take on Apple’s MagSafe. Whereas MagSafe is only a charging port, the Magnetic Connect port is an upgraded USB-C port, meaning that it can be used for data as well and is otherwise a fully capable USB-C port. The magnetic aspect of the port is technically not the connector itself, but rather it is magnets that Microsoft has placed on either side of the port to help keep the charging cable firmly attached. Microsoft-Surface-Laptop-Ultra-USB-C-vs-Magnetic-Connect-Ports To further reinforce the port (or rather, to reduce the chance of breakage), the actual port is larger than a standard USB-C port. The connector inside is still USB-C through and through, but the body is larger to allow it to break away more safely from the intricate 24-pin USB-C connector. While the Ultra can accept a charge on any of its USB-C ports, you would need a standard USB-C cable to plug in to the
Microsoft Surface Laptop Ultra-Front Angled 3 Amid the many trends and stories shaping the hardware world in 2026, the final quarter shows a confluence of several influences. AI continues to loom over everything this year, both client and server, and 2026 picks up where 2025 left off, with an even heavier emphasis on inference. Meanwhile, Microsoft’s latest efforts to bootstrap its Windows-on-Arm ecosystem are making good progress, thanks to another generation of improvements to its x86-on-Arm Prism emulator. And finally, amid all this, the client hardware market is trying to make sense (and cents) of a run-up in DRAM and NAND prices. As we enter the fall, these trends conveniently converge in a single device: Microsoft’s Surface Laptop Ultra. The company’s latest-generation flagship Surface laptop, the Ultra, is unlike anything that came before it, with Microsoft going above and beyond to develop a high-end laptop that, well, can go above and beyond itself. All the while, they are staking the fate of their new flagship laptop on the Arm architecture, which is still finding its footing in the Windows world. As a result, the Surface Laptop Ultra is a culmination of several firsts for Microsoft. As a complete product, it is the company’s most direct take on (or competitor to) Apple’s MacBook Pro lineup yet, making a direct play on the high-end market that Apple has captured so much of over the last several years. Meanwhile, this is one of the first-generation Windows laptops to ship with NVIDIA’s RTX Spark SoC, the consumer branding for their high-performance GB10 SoC that we have seen in numerous developer-focused mini-PCs over the last year. Finally, it is one of Microsoft’s biggest hardware releases for the developer market in at least the last several years, as Microsoft seeks to further bootstrap developers with all the tools they need to make the next generation of Windows applications. Microsoft Surface Laptop Ultra Key Specs (As Configured) Processor NVIDIA RTX Spark N1X 20-core Operating System Windows 11 Pro Memory 128GB LPDDR5X Storage 1TB SSD (M.2 2280, PCIe Gen5) GPU Options NVIDIA Integrated Blackwell RTX 5120 GPU core Display 15-Inch PixelSense, 3270×2180 120Hz, 2000nits Max Brightness, 100% DCI, Touchscreen Battery 92Wh Form Factor Laptop Dimensions 328.8 x 238.7 x 19.16 mm (12.95 x 9.40 x 0.75 in) Weight 2.0 kg (4.41 lbs) Wireless Wi-Fi 7 + Bluetooth 5.4 Colors Platinum (Gray) Nightfall (Black) Ports Left: 1x HDMI 2.1b, 2x USB-C 40Gbps, 3.5mm Combo Audio Jack Right: 1x USB-C 40Gbps (Magnetic Connect), 1x USB-A 10Gbps, SD Card Reader As a result, although this is far from Microsoft’s first Surface device, first Arm-on-Windows device, or even their first productivity laptop, the Surface Laptop Ultra is potentially the most important device launch from them in years. If it succeeds, it will shape the Surface ecosystem for years to come. If it does not, it will end up in the footnote pile next to 2012’s Surface RT (which we used at STH). The Surface Laptop Ultra is not meant to (nor will it) be a game changer for the broader laptop market, but it is still a significant launch for Microsoft’s hardware division. Meanwhile, for partner/supplier NVIDIA, although this is far from the first GB10 device on the market, the Ultra is one of a wave of laptops that marks the first time the chip has made it out of a mini-PC and into a laptop. NVIDIA has wanted to break into the market as a CPU/SoC supplier for years. This is their first chance in several years to do that. But how well will the N1X sibling to the GB10 translate into powering a Windows laptop? That is one of the questions we will have to answer today. Quick note: You may notice Ryan Smith wrote the first half of this review, and Patrick wrote the second half. Again, we had a week, and we were at OCP 2026 all week covering that show. So without any further ado, let us jump right in. Microsoft Surface Laptop Ultra External Hardware Overview When it comes to the Surface Laptop Ultra’s design, there is no dancing around the fact that this is meant to be Microsoft’s answer to the MacBook Pro. So we can only address it head-on. Microsoft wants to compete with Apple’s flagship laptops, and this is what they think a Windows equivalent should look and behave like. Microsoft Surface Laptop Ultra-Front 1 Microsoft is no stranger to designing laptops, of course. Their mainstream Surface laptops are already up to their 8th generation, and the Surface Pro 2-in-1s are too. But the company needs to step up its game with the Ultra. They are building a premium laptop, packing a more powerful SoC, and charging a higher price than ever. The 15-inch laptop is meant to straddle the line between being a productivity laptop and a portable laptop. It is a bit too small to be a desktop replacement-class laptop, which means it is reasonably portable. But it is not a thin and light ultraportable like the smaller mainstream Surface Laptops. Instead, at 19mm thick and weighing 2kg on the dot, this laptop fits the developer-on-the-go mindset. Microsoft Surface Laptop Ultra-Rear 1 Microsoft is offering a few different configurations of the Surface Laptop Ultra. Most notably, a choice between 18- and 20-core versions of NVIDIA’s RTX Spark N1X processor. Regardless of configuration, there is only one body and one size available. This is not a laptop lineup bifurcated by size, at least not yet. Surface-Laptop-Ultra-Ports-Side-by-side The body itself is made from aluminum, which is pretty standard fare these days for premium laptops. Microsoft offers it in either a “Platinum” (grey) finish, as seen with our review sample, or a “Nightfall” (black) finish. Many configurations are available in both colors, but on Microsoft’s pre-order website, Nightfall seems to be treated as the more premium option, since you cannot currently order a laptop with more than 32GB of RAM in the Platinum finish. Microsoft Surface Laptop Ultra-Front 2 All Surface Laptop Ultras ship with the same display. This is a 15-inch high-DPI 3270×2180 “PixelSense” touchscreen LCD display. According to Microsoft, the HDR-capable display can reach an eye-searing max brightness of 2000 nits and cover the full DCI-P3 color space. For developers who need a bit more fluidity (or want to sneak in some gaming), the display also supports variable refresh rates up to 120Hz. Since this is a premium laptop, Microsoft color-calibrates the panels individually. It is not an OLED, but our STH studio team members who usually use MacBook Pros all thought the display looked sharp. If you really want an OLED for those deep blacks, other manufacturers are making RTX Spark laptops with other screens. Still, this is a touchscreen, which can be useful for many. On the bottom deck of the laptop, we have Microsoft’s keyboard and haptic trackpad. As is typical for anything above budget laptops these days, this is a fully backlit keyboard. Microsoft is not duplicating Apple’s fingerprint sensor functionality here (relying instead on Windows Hello), so it is a keyboard that is just a keyboard. Meanwhile, according to Microsoft, the Ultra’s touchpad is their largest to date, larger than what is on the mainstream 15-inch Surface Laptop. And while you can just barely spot it due to the very thin bezel around the LCD panel, Microsoft has embedded a full camera suite along the top edge of the display. At the center is a 1080p “Surface Studio” camera. Meanwhile, IR cameras flank it on either side to support Windows Hello. For audio, Microsoft built in four downward-firing “OmniSonic” speakers, with two on the left front and two on the right front. Microsoft supports Dolby Atmos here. For users who need more enclosed audio, Microsoft explicitly confirms the laptop supports Bluetooth LE Audio, a feature that is still pretty rare on Windows laptops. Meanwhile, for audio input, there are a pair of microphones built into the body to allow for voice focusing. Microsoft Surface Laptop Ultra-Side 1 Meanwhile, when it comes time to plug something into the Surface Laptop Ultra, Microsoft has taken a somewhat restrained approach to the number of available ports. On the left side of the laptop, you will find an HDMI 2.1b port, two 40Gbps USB-C ports, and a 3.5mm combo audio jack. Surprisingly, Microsoft has not opted to secure Thunderbolt 4 certification despite the laptop’s premium focus. Still, at 40Gbps with full DP Alt Mode and 140W charging, these are highly capable USB-C ports. At present, we are still trying to suss out who is providing the controller driving them, as the RTX Spark SoC itself does not support 40Gbps USB speeds. Microsoft Surface Laptop Ultra-Side 2 As for the right side of the laptop, things get a bit more interesting. On this side we find 3 ports: a full-size SD card reader, a 10Gbps USB-A port, and a special 40Gbps USB-C port that Microsoft calls its Magnetic Connect port. Microsoft Surface Laptop Ultra-Magnetic USB Type-C Charging Cable 2 Replacing Microsoft’s proprietary Surface Connect port from past designs, the company has developed a magnetic USB-C port; Microsoft’s take on Apple’s MagSafe. Whereas MagSafe is only a charging port, the Magnetic Connect port is an upgraded USB-C port, meaning that it can be used for data as well and is otherwise a fully capable USB-C port. The magnetic aspect of the port is technically not the connector itself, but rather it is magnets that Microsoft has placed on either side of the port to help keep the charging cable firmly attached. Microsoft-Surface-Laptop-Ultra-USB-C-vs-Magnetic-Connect-Ports To further reinforce the port (or rather, to reduce the chance of breakage), the actual port is larger than a standard USB-C port. The connector inside is still USB-C through and through, but the body is larger to allow it to break away more safely from the intricate 24-pin USB-C connector. While the Ultra can accept a charge on any of its USB-C ports, you would need a standard USB-C cable to plug in to the
USB4 has been out for some time now and, to say the least, it has been a breath of fresh air to storage enthusiasts and media professionals worldwide. Newest on that USB4 storage block is the Samsung P9 USB4 Portable SSD which is capable of 4GB/s read and 3.8GB/s write speeds. What’s is an even more important metric is that Samsung has put the effort into this storage device and ensured that it contains great sustained write speeds. After all, transferring media means nothing if our intial speedy data transfer speed drops to anything less that 1GB/s in today’s media world. Better yet, we have the Samsung P9 on our test bench and will be demonstrating these speeds first hand in this report! The P9 doubles that speed with an incredible 2GB/s sustained write performance. The Samsung P9 is brand spanking new to the storage space. We are very fortunate to have one of the first in hand for review, and in fact, Samsung’s P9/P7 Launch Event was only three weeks ago. Check it out here: The Samsung P9 is a USB4 Gen3 x4 portable SSD and is available in capacities of 1, 2, 4 and 8TB. Samsung claims the 8TB is the industry’s first of its kind with these speeds and this capacity. Listed performance is 4GB/s read and 3.8GB/s write and Samsung speaks to sustained write performance of 2GB/s in their preliminary literature. The Samsung P9 comes with the P9 itself and a USB4 rated cable. This cable is rather important as it is a USB 4 cable rated at 40Gbps and, anything with a lesser rating, will not be able to attain full performance. Speaking of which, this portable SSD is backward compatible but, in order to attain top performance, your PC or laptop must have a USB Type-C port capable of performance up to 40Gbps. This will be listed somewhere in your system literature. The Samsung P9 is constructed of a flat black metal alloy with a LED top left of the front, and also has a thin rubber base which makes it convenient for desk placement. It is rather thin, measuring 90x65x8.3mm and weighing only 96 grams. MSRP for the P9 is listed at $339.99 (1TB), $649.99 (2TB), $1349.99 (4TB) and $2289.99 (8TB) but CHECK AMAZON for availability and more accurate street pricing. Surprisingly, there is no IP-rating for the P9 but it does speak to advanced thermal protection and this SSD does have AES 256-bit hardware encryption. In fact, this hardware encryption gets put to use right off if you choose to run Samsung Magician and secure this portable SSD by setting a 4-16 character password. Samsung Magician is absolutely free and one of the best SSD maintenance software packages we have seen to date. Not only does it allow you to activate the SSDs security mode, but also, one can secure erase the SSD, validate that it is Samsung Authenticated, change partition sizes, as well as confirm the performance as we have done here: Let’s jump to the next page and validate the Samsung P9 portable SSD performance…
AMD Helios HPE Juniper QFX5252 OCP 2026-2 At OCP 2026, we saw something that looked absolutely awesome. Walking by the HPE booth, we saw an AMD Helios tray set up and open on the show floor. It stopped me in my tracks, so I thought folks would like to see it. This is the AMD Helios HPE Juniper QFX5252 Switch Tray at OCP 2026 Oftentimes, when we open up huge switches, we see a lot of internal PCB. This is totally different since it is a double-wide tray with side-by-side switches and lots of copper. AMD Helios HPE Juniper QFX5252 OCP 2026-1 You can see the HPE QFX5252 label on the front latch, which is huge. Because there is so much connectivity on the back, and the back panel is long and needs to blind mate, the insertion force required for a switch tray like this into the Helios rack is significant. As a result, these levers are very long. Since this is two big switch chips in a 1OU tray, it needs to be liquid cooled. You can see the internal liquid cooling manifolds distributing liquid to the various blocks. AMD Helios HPE Juniper QFX5252 OCP 2026-3 There are two switch chips inside, which are Broadcom Tomahawk 6 102.6T switches. At first, you might wonder why these are not Broadcom Tomahawk Ultra chips, since that might make sense for a scale-up architecture. The current generation Tomahawk Ultra is a 112G PAM4 SerDes switch (so 512 lanes of PAM4-112G.) That is not enough for Helios which is using 224G PAM4 SerDes. As a result, these need to be Tomahawk 6. AMD Helios HPE Juniper QFX5252 OCP 2026-2 Now, the next part is very different. AMD and HPE-Juniper are not just slinging Tomahawk 6 on some PCBs and calling it a day. You will notice that the cabled connections to the rear high-density fabric connectors for the Helios rack cable cartridges are quite far from the Tomahawk 6 switch chips. Some of them in particular are very far from the switch chips. The liquid cooling blocks between those cable connectors and the switch chip are apparently retimers. AMD is big on retimers while NVIDIA has worked to remove them. AMD Helios HPE Juniper QFX5252 OCP 2026-6 Aside from the switches and retimers, let us talk about the cables. AMD Helios HPE Juniper QFX5252 OCP 2026-7 When we say copper inside a rack, and why we will eventually go to co-packaged optics, this is a great illustration. In the QFX5252, retimers extend reach so we can reliably scale up within the rack. Then we need an enormous array of copper cables. CPO would simplify this design. I will probably use these photos when explaining the “why” for CPO in the future. AMD Helios HPE Juniper QFX5252 OCP 2026-4 You can see the fans and airflow guides over heatsinks for some other components. Unlike NVIDIA’s push to move fans out of the AI rack, AMD Helios still uses fans and air cooling for some other bits in the tray. AMD Helios HPE Juniper QFX5252 OCP 2026-5 Overall, very cool. Final Words We previously covered the QFX5252 at HPE Discover 2026, but seeing the switch tray assembled and on the show floor was really cool. A lot is going on here, and it is very different from the normal rackmount switches we see. One of the biggest differences, of course, is that it is designed to scale up within a rack instead of out of front-panel connectors. HPE Discover 2026 HPE Juniper Networking QFS5252 Of course, if you are into networking, check out our Inside the Secretive AWS Networking Lab tour that we were finally able to publish yesterday. VIDEO As a fun aside, I was at HPE’s booth looking at this switch tray when Pradeep Sindhu walked over and started looking at the switch with the HPE-Juniper team and me. For me, it was neat because the last time I saw Pradeep was when we did this Fungible DPU-based Storage Cluster Hands-on Look back in the pandemic era in 2020. It was neat having Juniper’s co-founder and former CTO look at the big switch tray while I was there.
The Open Compute Project (OCP) Foundation announced a sharp increase in IDCs forecast for worldwide spending on OCP-recognized IT equipment. The 2026 IDC Market Impact Study projects annual spending of approximately $800 billion in 2029, up from the $295 billion forecast for the same year in last year’s study—an increase of approximately 2.7x.The new forecast reaches $892 billion in 2030, when OCP-recognized IT equipment is projected to account for approximately 51% of all spending across the infrastructure categories covered by the study. Commissioned annually by the OCP Foundation, the study assesses the market impact of OCP-recognized equipment. This year’s findings reflect both the rapid expansion of AI infrastructure investment and broader adoption of OCP designs. “AI is changing what the industry needs from its infrastructure, and these findings show the growing role of open designs in meeting that demand,” said Cliff Grossner, Ph.D., Chief Innovation Officer, OCP Foundation. “Our Community is bringing companies together to solve challenges that span the entire data center, from silicon and servers to cooling, power and the grid. That collaboration helps organizations put new technology to work at scale.” AI Investment Drives Growth IDC projects spending on OCP-recognized IT equipment will grow from approximately $154 billion in 2025 to $892 billion in 2030, representing a compound annual growth rate of approximately 42%. Servers represent the largest spending category, with GPU-accelerated systems driving much of the growth. IDC also forecasts continued growth in switching equipment, rack infrastructure, power and cooling as organizations build the systems needed to support increasingly demanding workloads. Key findings include: $800 billion in projected annual spending in 2029, approximately 2.7 times the prior forecast for that year $892 billion in projected annual spending in 2030, including approximately $814 billion in servers Approximately 51% of spending across the infrastructure categories measured going to OCP-recognized infrastructure by 2030 Approximately $628 billion in AI-related spending on OCP-recognized infrastructure in 2030, accounting for about 70% of the forecast total Click to enlarge IDC also points to the growing ecosystem of solution providers and the availability of openly shared designs as factors supporting adoption. Use of OCP-recognized IT equipment extends beyond the Foundation’s formal membership to organizations that access and implement those designs worldwide. Advancing Open Infrastructure from Chip to Grid The findings reinforce OCP’s expanded Open Data Center Ecosystem Vision, which addresses the interconnected demands of AI infrastructure from chip to grid. Over the past year, the OCP Community has advanced higher-power rack designs, liquid cooling, Ethernet-based AI networking and open chiplet specifications. Its facilities work includes DC power distribution, energy storage, and coordination between data centers and the utility grid. The Foundation’s Open Data Center for AI Strategic Initiative brings these efforts together across facility power, cooling and management. The OCP Energy Project extends collaboration into energy generation and storage, while the AI Computing Continuum project supports deployment beyond hyperscale data centers into enterprise, regional and edge environments. As AI infrastructure evolves, OCP will focus on supporting diverse AI silicon, coordinating design across the data center, and developing open building blocks that organizations can deploy across a wider range of environments.
Summary: Expanding memory capacity for AI: Second-gen Marvell Structera X 3406 and 3506 CXL 3.2 memory expanders add six DDR4/DDR5 channels (50% more than the prior gen) and PCIe Gen 6 connectivity, supporting up to 9TB of DDR4 or 6TB of DDR5 per controller Addresses growing memory demand: Multi-host sharing and inline LZ4 compression deliver up to 2-3x more effective capacity, with up to 40-50% faster time to first token at equivalent context length Preview and availability: Demoing at the OCP Summit, the Structera X 3406 and 3506 are expected to sample in Q1 2027, extending the Structera X 2404 and 2504 architecture already shipping to hyperscalers Marvell Technology, Inc., a reference in data infrastructure semiconductor solutions, announced the second-gen Marvell Structera X 3406 and Marvell Structera X 3506 CXL memory expanders, designed to help cloud, enterprise and AI infrastructure providers build more efficient, scalable and higher-performance systems for the next era of AI and computing.Designed to address the evolving needs of leading hyperscale datacenters, Structera X addresses the most demanding data center workload requirements. Marvell Six-channel CXL 3.2 Memory Expander for AI Inference AI and ML workloads are rewriting the rules of data center design. Growing LLM sizes, expanding context windows and growing KV-cache requirements are driving unprecedented memory demands that traditional server architectures simply cannot meet. Memory is a critical constraint of AI performance and infrastructure efficiency. Modern AI inference workloads, from LLM serving to deep learning recommendation models (DLRM), are memory bound. KV-cache capacity directly determines time to first token (TTFT), throughput, and XPU utilization. Structera X 3406/3506 addresses these bottlenecks with: Industry-first six-channel CXL memory expansion: Six DDR4/DDR5 memory channels deliver 50% more channels than the previous gen, increasing memory bandwidth and capacity so AI servers can sustain larger models and longer context windows – up to 18 DDR4 RDIMM slots supporting 9TB for Structera X 3406 and 12 DDR5 RDIMM slots supporting 6TB for Structera X 3506 Industry-first four host connectivity: Enables memory capacity sharing with up to four hosts and up to 256GB/s PCIe6.0 x16 of host bandwidth, improving memory utilization and efficiency across AI clusters 5x higher LZ4 compression throughput from previous generation: Up to 250GB/s compression/decompression throughput unlocks up to 2 – 3x more effective memory capacity, depending on workloads StructGuard enterprise-grade RAS: Provides advanced error correction to enhance memory reliability and resilience vs. DRAM device failures Industry-first post-quantum security: Combines post-quantum cryptography, Caliptra 2.0, CXL IDE and AES-256 memory encryption to provide advanced data protection For DLRM workloads, Structera X delivers up to 2x higher queries per second and 2-3x more effective embedding-table capacity per controller. For LLM serving, Structera X delivers up to 40-50% faster TTFT at equivalent context length, or up to 2x longer supportable context windows at equivalent latency compared to the previous generation. Building on the success of the first-generation Structera X 2404/2504, the Structera X 3406/3506 leverages the proven software stack and ecosystem integrations enabling an accelerated ramp. The Marvell CXL ecosystem continues to expand across the full data center stack through close partnership with CPU and memory vendors, OEMs, ODMs and system integrators. “CXL is starting to gain widespread adoption in AI systems since they reap strong benefits from its very high bandwidth and support of large memory sizes,” said Jim Handy, general director, Objective Analysis. “In addition, today’s memory shortage and subsequent high prices are driving IT managers to use CXL to re-deploy used DDR4 DIMMs into DDR5 systems for immediate cost savings. In the near future, these same managers will use CXL’s memory pools to reduce stranded memory and drive greater savings, and they will find controllers that offer compression, as this one does, to be even more interesting.” By bringing together connectivity, memory and storage technologies, the expanded data center portfolio improves how compute and memory work across AI systems, increasing utilization, infrastructure efficiency and AI performance. These next-generation offerings increase access to memory resources, improving both memory and compute utilization, helping customers optimize AI infrastructure performance as workloads continue to grow in scale and complexity. “Marvell is the only provider with a complete end-to-end CXL portfolio spanning memory expansion (Structera X), near-memory acceleration (Structera A), and PFMA (Photonic Fabric Memory Appliance), pod-scale CXL memory-sharing appliance, giving customers a complete portfolio from direct-attach to disaggregated memory pooling,” said Will Chu, EVP and GM, custom cloud solutions, Marvell. “We are delivering industry-leading CXL controllers, co-developed with leading hyperscalers to address the most demanding AI infrastructure requirements.” Industry Ecosystem Support “AMD is advancing an open ecosystem that gives enterprises the freedom to build infrastructure around their evolving workloads,” said Raghu Nambiar, corporate VP, data center ecosystems and solutions, AMD. “CXL is an important part of that vision, enabling customers to expand memory capacity and use resources more efficiently as AI and data-intensive applications grow. Our collaboration with Marvell builds on the performance and scalability of AMD EPYC processors to help enterprises realize those benefits, bringing greater flexibility and long-term value to their data center investments.” “As AI becomes more agentic, memory and compute need to scale together to meet growing infrastructure demands. Marvell’s latest Structera X controllers bring CXL 3.2, memory expansion and compression to the Arm ecosystem, giving hyperscaler customers more options to efficiently support complex AI workloads,” said Eddie Ramirez, VP, GTM, cloud AI, Arm. “Marvell and Alibaba are collaborating to advance CXL memory solutions for the rapidly growing demands of AI infrastructure. As AI models and workloads continue to increase in scale, expanding memory capacity and bandwidth without compromising system performance and efficiency has become a critical requirement. CXL enables a more flexible approach to memory expansion and sharing, providing a path to scale memory resources across increasingly complex AI systems. Our collaboration with Marvell brings together its leadership in CXL technology with Alibaba’s experience in large-scale AI and cloud infrastructure to develop memory architectures designed for the next generation of AI data centers.” “Meta is pleased to work with Marvell to advance CXL solutions that address the growing memory demands of compute infrastructure,” said Yashar Bayani, director, hardware engineering, Meta. “As systems scale, increasing memory capacity while improving infrastructure efficiency and cost becomes critical. CXL provides a flexible approach to expanding memory resources and optimizing infrastructure economics, and we look forward to working with the ecosystem, including Marvell to help enable the next generation of scalable infrastructure.” “CXL is reshaping how memory is deployed in data centers, and our work with Marvell ensures customers can confidently deploy Structera with Samsung DDR memory for reliable, high-performance systems,” said Jangseok Choi, VP, memory product planning tea, Samsung Electronics. “The move to CXL 3.2 and six memory channels in Structera X 3406/3506 further strengthens the flexibility our joint customers need.” “Working with Marvell to validate Structera with SK hynix memory supports our shared goal of making memory expansion more accessible and flexible,” said Uksong Kang, head, next generation product planning and enabling, SK hynix. “Structera X 3406/3506 gives customers the tools they need to build future-ready, disaggregated architectures with less friction and more choice.” Availability Marvell Structera X 3406 and Structera X 3506 CXL memory expansion controllers are expected to sample in 1Q27.
KIOXIA has announced the addition of direct liquid cooling (DLC) models to its KIOXIA CM9 Series enterprise SSD lineup to support cold-plate-enabled server designs. The new E3.S DLC models also support traditional air-cooled environments. As AI and other high-performance enterprise environments continue to advance, direct liquid cooling is becoming increasingly important to address growing computing power demands and associated cooling requirements. In order to fulfill the growing compute demands, increased power requirements are driving the need for innovative heat dissipation solutions. DLC infrastructure can offer significant energy savings compared with conventional air cooling, while enabling more servers to be cooled per rack. Ideal for SSD Context Caching, the new KIOXIA CM9-DLC models include read-intensive (1 DWPD) and mixed-use (3 DWPD) endurance options. “The rise of AI infrastructure is driving a fundamental shift in data center design toward higher power densities and advanced cooling technologies,” said Maulik Sompura, Sr. Staff Director, Product Management, SSD business unit, KIOXIA America, Inc. “The KIOXIA CM9-DLC Series addresses these evolving requirements by combining industry-leading enterprise SSD performance with direct liquid cooling support in the E3.S form factor. This enables customers to deploy denser, more energy-efficient AI systems while providing the performance, endurance, and capacity needed for data-intensive workloads, including emerging AI architectures that rely on large-scale memory and KV cache resources.” KIOXIA CM9-DLC Series SSD highlights include: • PCIe® 5.0, NVMe™ 2.0, NVMe-MI™ 1.2c specification compliant • Support for OCP Datacenter NVMe SSD v2.5 specification • Dual-port support in the E3.S SSD form factor • Capacities up to 30.72 TB read-intensive (1 DWPD) and 25.6 TB mixed-use (3 DWPD) endurances • Sequential performance (128 kibibytes (KiB)/QD32) of up to 14.8 GB/s read and 11 GB/s write • Random performance (4KiB) of up to 3,400 KIOPS (QD512) and 800 KIOPS (QD32)
Marvell Structera X 3506 CXL has seen a massive surge recently, and Marvell announced its second-generation Structera X CXL memory expanders, adding CXL 3.2 support, PCIe Gen6 connectivity, and increasing memory channels from four to six. For hyperscalers running DDR4 and lower-speed DDR5 reclamation projects, that means the DDR4 version can support up to 18 DDR4 RDIMMs. Marvell Structera X 3406 and 3506 CXL 3.2 Memory Expanders Patrick gave us the decoder for the models: CXL 3 support is indicated by the first digit 3, just as the 2 in last year’s 2404 and 2504 marked CXL 2.0. 5 or 4, is the DDR generation The trailing two digits are the memory channel count. Structera X 3406 is therefore a CXL 3 DDR4 part with six channels, while the 3506 is the CXL 3 DDR5 part, also at six channels. This announcement covers only the Structera X memory expansion line, not the Structera A near-memory compute controller that Marvell showed us last year. Still, CXL is focused on optimizing either bandwidth per core (since CXL bandwidth is in addition to native DDR bandwidth) and also capacity expansion. Marvell CXL Addresses Memory-intensive Applications Bandwidth / core High-bandwidth ML/AI inference DLRM Maximize memory bandwidth PCIe Gen6 infrastructure and CXL 3.2 means we get roughly twice the bandwidth over the link, so that is how Marvell is adding more bandwidth from more memory channels. Marvell New: CXL 3.2 Memory Expander CPU Structera X 3406/3506 CPU … Memory-expansion controllers 2x Both new controllers stay on Marvell’s 5nm process and have LZ4 compression. Compression means the installed memory is storing more than the physical DRAM capacity. In this generation, Marvell is ensuring that the LZ4 can keep up with the interface so it is not a bottleneck. Here is the comparison table with key specs: Marvell CXL Expander Product Overview Category DDR4 – Stuctera X 2404 DDR5 – Stuctera X Structera X 3506 pairs six DDR5 channels with LZ4 compression at 250GB/s, while the 3406 runs DDR4 with six channels at 1600MT/s and 3DPC. That lets the DDR4 part take 18 DIMMs for up to 9TB per controller, and the DDR5 3506 takes 12 DIMMs for up to 6TB. CXL 3.2 lets up to four hosts share pooled memory, with up to 256GB/s of host bandwidth, double the prior generation’s 128GB/s. Marvell Page Compressibility With Structera X Expander Frequency of obtaining given compression ratio by data’s We have known that memory pages have varying levels of compressibility, and Marvell gave us some data on that. Final Words We covered the first-generation Structera X and Structera A when we visited Marvell’s labs last year, including how line-rate LZ4 compression makes attached memory roughly half as costly per gigabyte. That hands-on look explains why CXL expanders are attractive as AI demand tightens DRAM supply and pushes DDR5 prices up. VIDEO For hyperscalers, this seems like a better version of the 2000 series. Unlike many of Marvell’s competitors who are moving to four channels in their CXL 3.x generation, six channels put 50% more DRAM behind a single controller.
Read also : IT Press Tour 69: HYCU Going beyond SaaS data protection and ransomware with advanced AI control By Philippe Nicolas | September 8, 2026 | News
IT Press Tour 70: Cohesity
7 d ago
11:11 Systems , a leading managed infrastructure solutions provider, announced it has completed the acquisition of select IBM VMware Cloud Service Provider enterprise customers.Broadcom has recently optimized its VMware Cloud Service Provider (“VCSP”) partners program to improve customer outcomes, and as the largest, private, global VCSP, 11:11 Systems has established itself as the leader in this space focused on helping VCSPs who desire to transition their customers to its platform. “Enterprises today are crossing the Rubicon into a new era of cloud, containers and private AI, and the path forward requires both fortitude and practical guidance,” said Brett Diamond, CEO, 11:11 Systems. “11:11 Systems was built for that role. With our resilient platform, deep VMware and cloud expertise and global reach, we help customers modernize on their terms, protect the applications and data that matter most and make the infrastructure decisions today that turn into a competitive advantage in the future.” The acquisition further expands 11:11 Systems’ global reach across North America, EMEA, APAC and South America and expands its ability to support complex enterprise migrations through an integrated portfolio of cloud infrastructure, connectivity, cyber resilience and recovery services. It also reflects the strong alignment between 11:11 Systems, IBM and Broadcom/VMware particularly around customer experience, innovation and long-term service continuity. “Broadcom’s Cloud Service Provider strategy is centered on partners that can deliver the operational discipline, technical expertise and enterprise scale our customers require,” said Ram Velaga, president, infrastructure software group, Broadcom. “11:11 Systems is a partner that reflects our model. With a global platform, extensive VMware experience, more than 6,000 customers and a portfolio spanning cloud, connectivity and security, they are well positioned to support enterprise customers as they modernize critical workloads utilizing VMware Cloud Foundation.” 11:11 Systems has acquired seven major VMware-based businesses – iland Cloud, Green Cloud Defense, Unitas Global, Sungard Availability Services, Faction, Ntirety and Digital Sense – establishing an unmatched foundation of operational expertise and commercial insight. The acquisition of select IBM VMware Cloud Service Provider enterprise customers will mark 11:11’s eighth purchase of a VMware CSP business and eleventh acquisition overall. Through its global cloud platform and partner-first approach, 11:11 Systems helps customers and providers run critical environments with greater resilience and scale. Moelis & Company LLC and Fifth Third Securities, Inc. served as financial advisors for 11:11 Systems, while Cooley LLP acted as legal advisor for the company in the transaction. The terms of the transaction were not disclosed.
Kioxia LD4 E1.L SSD in slot-in orientation Kioxia has introduced its first E1.L form factor SSD, built with BiCS FLASH generation 8 QLC flash memory. We did a piece a few days ago showing BiCS under a microscope. Kioxia is targeting read-intensive hyperscale workloads in high-density 1U scale-out servers with the LD4 Series, launching at 15.36TB and 30.72TB with a validated architecture the company says supports up to 122.88TB. For “ruler” SSD fans, this should be exciting, but it also shows some clear industry trends. Kioxia LD4 E1.L QLC SSD Announced The Kioxia LD4 is clearly a read-focused drive. It uses QLC NAND to boost capacity at the expense of raw speed. E1.L uses a 9.5mm thickness and a 318.75mm length, a shape designed to put more flash capacity into a 1U chassis by extending the depth for longer PCBs and therefore more NAND sites. We covered this in our EDSFF primer years ago. VIDEO Although PCIe Gen6 servers have started to arrive in the STH lab, this is still a PCIe Gen5 NVMe SSD. Our sense is that storage applications, especially those using QLC NAND will be slower to transition to Gen6 than the drives we see in GPU nodes. KIOXIA LD4 Series Overview Slide 3: KIOXIA LD4 Series data center NVMe SSD specifications At launch, Kioxia is offering 15.36TB and 30.72TB models, while the company says the underlying architecture has been validated for capacities up to 122.88TB. Kioxia LD4 E1.L SSD front view As a “D” series drive, we know that this is designed for hyperscale data centers. The LD4 is a single-port design and is apparently already sampling. Final Words While folks are looking at E3.L and E2 for capacity, this form factor matters for 1U servers, where you can either use less front panel space for a given capacity or cram more drives in. Two 30.72TB 9.5mm E1.L drives do not take up much front panel area, so if you are building front-cabled servers, that is important. Also, if you look at ORv3 racks that have power, liquid cooling, and in-rack interconnect in the rear, you can quickly see how having capacity in a 1U-compatible E1.L drive can make a lot of sense. Hopefully we get to see more at OCP Summit next week.
Netlist Files New ITC Action Against Micron Technology Involving HBM Patents for AI Computing
10 d ago
AMD announced that it has entered into a definitive agreement to acquire World Labs, an AI model and research lab led by AI pioneer Dr. Fei-Fei Li.The acquisition will bring a world-class team of researchers and model experts to AMD, strengthening its ability to develop AI hardware, software and systems around the needs of emerging models and applications. As AI expands into reasoning, robotics, simulation and physical AI, the demands on compute infrastructure become more diverse. World Labs‘ expertise in developing advanced models will give AMD deeper insight into how workloads are evolving and help shape its future technology roadmaps. The acquisition advances AMD’s strategy to deliver AI infrastructure for an open ecosystem. “Building the compute platforms for the next gen of AI requires a deep understanding of how models are evolving,” said Dr. Lisa Su, chair and CEO, AMD. “Fei-Fei and the World Labs team bring exceptional research leadership and model expertise. Together, we can use that insight to develop the hardware, software and systems that will power the next gen of AI and strengthen the open AI ecosystem.” “Advancing the next generation of AI technology requires close collaboration across model research, systems and compute,” said Dr. Fei-Fei Li, co-founder and CEO, World Labs. “Joining AMD will give our team the resources and engineering depth to accelerate our research and help define the infrastructure needed for the next era of AI.” Headquartered in San Francisco, World Labs develops spatial-intelligence models that generate, reconstruct and simulate interactive 3D environments from text, image and video inputs, as well as technology for robotic learning and simulation. Following the close of the transaction, the World Labs team will continue to focus on advancing AI model research, and Dr. Fei-Fei Li will join AMD as executive VP and chief scientist, reporting to Dr. Lisa Su. The all-stock transaction is valued at approximately $8.2 billion and is expected to close by the end of 2026, subject to regulatory approvals and other customary closing conditions.
Xsight Labs announced the E1L, the newest member of its E-Series data processing unit (DPU) portfolio.The E1L brings fully programmable DPU capability, at up to 200G, to as little as 25 watts TDP (15 watts typical) in its eight-core, 100G configuration, powered by Arm Neoverse N2 cores and designed for edge, embedded and other power-constrained infrastructure. Like the E1, it is built on open standards and runs standard software without proprietary licensing, and with full cross platform compatibility, software written for one, runs on the other. For operators, bringing DPU offload into a power-constrained site has meant a proprietary software stack sized for one chip, or leaving offload out. The E1L changes that: full DPU offload from 25 watts, on the same architecture and software as the 800G E1, so an offload written once runs from a 25-watt DPU at the edge to an 800G DPU in the data center, on standard Linux tools and with no proprietary licensing. AI is shifting from models that answer questions to agents that take action. Every agent query, lookup and tool call becomes traffic to databases, storage and the network, and that infrastructure work lands on the same host CPUs meant for the AI workload. Agents touching critical systems also raise the bar for isolation. The E1L takes that infrastructure work off the host, fully or in part, and runs it in a separate, isolated domain, at a power envelope that fits sites outside the data center. The E1L is scheduled to begin sampling in the second quarter of 2027. Open, programmable silicon for the modern data center Unlike conventional network interfaces, DPUs offload networking, security and storage tasks from the main server CPU. A network interface card moves data; a DPU also carries its own processors and memory and runs the infrastructure software itself. By building on open standards, Xsight Labs provides a fully programmable, merchant-silicon alternative to proprietary DPUs, one that accelerates software-defined infrastructure without proprietary architectures or licensing. The E1L: low-power DPU for edge and embedded systems Engineered for space- and power-constrained environments and for embedded applications, where specialized data center cooling is limited or absent, the E1L brings fully programmable DPU capability, at up to 200 Gbps, with TDP from 25 W to 40 W and typical draw of 15 to 25 W. The power figure starts with the silicon: the E1L is built on TSMC’s N5 process, the same node as the E1, so it does the same work in a smaller form factor, with less energy and less heat. Built on Arm Neoverse CSS N2 and designed for Arm SystemReady compatibility, the E1L runs standard Linux and familiar software frameworks including XDP, DPDK and SPDK, so infrastructure teams can develop and deploy programmable networking, storage and security services without a proprietary software environment. The E1L is offered as a discrete device, a COM Express module and a 1RU server, all sharing the same software and feature set. Alongside the 800G E1, which draws under 75 W, the E1L completes an E-Series that runs one software stack from 100G at the edge to 800G in the core. E1L key specifications Compute: 8, 16 or 22 Arm Neoverse N2 cores Networking: up to 200 Gbps over 4 SerDes lanes Memory: up to 2 DDR5-5200 interfaces with inline encryption Security: line-rate AES-GCM for IPsec offload Host interface: 20 PCIe Gen5 lanes across 6 dual-mode controllers with SR-IOV support Power: configurable 25 W, 30 W and 40 W TDP (typical draw 15 W, 18 W and 25 W) Software: off-the-shelf Linux distributions; XDP, DPDK and SPDK integration “The E-Series now scales from 100 gigabits at the edge to 800 gigabits in the core on a unified architecture and software stack. Operators told us they needed DPU capability within the power envelopes they already rely on. The E1L delivers that,” said Yossi Meyouhas, CEO, Xsight Labs. “AI infrastructure is expanding across a broader range of environments, making performance per watt and software flexibility critical for deployment. By building on Arm Neoverse CSS N2, Xsight Labs is bringing programmable, efficient compute from the edge to the data center while giving developers the benefits of a common software ecosystem,” said Eddie Ramirez, VP, go-io-market, cloud AI, Arm. Availability The E1L is scheduled to begin sampling in the second quarter of 2027. Xsight Labs will present the E1L at the 2026 Open Compute Project (OCP) Global Summit, October 12 to 15, in San Jose, as an OCP Bronze member.
KIOXIA has introduced its first E1.L form factor SSD built with KIOXIA BiCS FLASH™ generation 8 QLC flash memory. Designed for high-density 1U servers and read-intensive storage applications, the new KIOXIA LD4 Series targets hyperscale data center-class workloads and scale-out environments. As data requirements continue to grow, hyperscale data centers continue to increase the storage capacity of their highest density systems. The KIOXIA LD4 Series brings high-capacity QLC flash to the E1.L form factor for read-intensive applications in these environments. “The continued growth of data-intensive workloads is driving demand for higher-capacity storage in hyperscale data centers,” said Neville Ichhaporia, senior vice president and general manager, SSD business unit, KIOXIA America, Inc. “QLC flash is particularly compelling for read- intensive applications, where high capacity and high read performance supports the vast data storage requirements of today’s AI workloads.” Key features of the KIOXIA LD4 Series include: • PCIe® 5.0 (Gen 4 x4) 1, NVMeTM 2.0e specification compliant • Support for OCP Datacenter NVMe SSD Specification 2.6 • Single-port design for hyperscale and data center-class servers • Available in 15.36 TB2 and 30.72 TB capacity points today, with a validated architecture enabling capacities up to 122.88 TB • Built with KIOXIA BiCS FLASH generation 8 QLC flash memory KIOXIA LD4 Series drives are currently sampling to select customers and will be showcased at the 2026 OCP Global Summit, taking place October 12 – 15 in San Jose, Calif. Company Contact: Mia Cool KIOXIA America, Inc. Tel: (408) 526-3087 mi******@****ia.com
Xsight Labs E1L DPU Package Xsight Labs has an interesting announcement today. The new Xsight Labs E1L DPU is designed to be a lower-power DPU, bringing 8-22 Arm Neoverse N2 cores into a power envelope that is low enough to put into a network switch or other low-power footprints. Xsight Labs E1L DPU for Lower-Power 200Gbps DPUs Xsight told us that the E1L DPU is designed to be a lower-power alternative to the 64-core E1 DPU that we have featured on STH several times now. Here are the key specs: 8, 16, or 22 Arm Neoverse N2 cores Up to 200Gbps over 4x SerDes lanes Up to 2x DDR5-5200 interfaces with inline encryption Line-rate AES-GCM for IPsec offload 20x PCIe Gen5 lanes across 6x dual-mode controllers with SR-IOV support Configurable 25 W, 30 W and 40 W TDP (typical draw 15 W, 18 W and 25 W) Xsight’s idea is basically the anti-AMD Pensando. While AMD says running network offloads via Linux and DPDK on a sea of Arm cores is the wrong approach, and instead focuses on P4 programmable engines, Xsight focuses on running on Arm cores. The Neoverse N2 cores on the E1 DPU are so vanilla in their implementation that we featured Installing Out-of-the-Box Ubuntu LTS on Xsight Labs E1 64-Core Arm 800G DPU. Xsight Labs E1 DPU Ubuntu Installation Step 11 Installation Complete Time To Reboot The new DPU is designed to run on a COM Express Type 7 module. When I heard this, the first thought I had was, “That works in network switches.” The E1L is also an interesting design for storage servers, both NVMe and disk. The storage path is perhaps exciting because, typing that, I realize you could theoretically attach E1.L SSDs to the PCIe lanes of the E1L DPU and have the entire system be built around E1L/E1.L. Xsight Labs E1L DPU COM Express Type 7 The Arm Neoverse N2 cores are actually pretty slick. Just to give you some idea of how they fare compared to some other options on the STH AgentSTH V7 workloads, we charted a few interesting data points: Xsight Labs E1 DPU Single Arm Neoverse N2 Core AgentSTH V7 Performance We are using a single-core score here just because the E1 DPU has 64 cores and we wanted to give some sense of relative performance. Just given memory constraints, we have not been able to get a BlueField-2 DPU running AgentSTH V7, but a BlueField-4 is based on the Neoverse V2 core that we found on the Grace CPU in the above (likely higher clocks in the GB300 CPU above), and here is the BlueField-2 DPU on Geekbench 5: XSight Labs E1 DPU Versus NVIDIA BlueField 2 DPU Geekbench 5 The big multi-core delta is also because the E1 has 64 cores and a newer architecture. Still, an 8-22 core part is very capable for applications like storage, switch control planes, and virtualization hosts. Final Words Xsight Labs is just different. While a lot of the industry is chasing 800Gbps and higher speeds, the company is taking its 800Gbps design and scaling it to 200Gbps. More importantly, it is scaling the design to a lower power level, which will allow it to go into many more applications where power is a bigger concern than hitting 800Gbps. Also, from an architecture scaling standpoint, since the company uses Neoverse N2 Arm cores with DPDK and other standard Linux features, it means an acceleration path should scale easily down from something running at 800Gbps on the 64-core E1 to the lower-core-count E1L.
Surface-Laptop-Ultra-and-Surface-RTX-Spark-Dev-Box Microsoft today has formally launched its Surface Laptop Ultra, the first in a series of developer-focused machines from the company that are based around NVIDIA’s RTX Spark processor. The consumer iteration of NVIDIA’s GB10 chip, which has been gracing Linux-based AI mini-PCs for the past year as part of NVIDIA’s DGX Spark ecosystem, RTX Spark marks the chip’s formal entrance into the consumer market. And, for partner Microsoft, the company gets the honor of being one of the first vendors to ship an RTX Spark laptop with its Surface Laptop Ultra. Itself the new face of Microsoft’s high-end laptop lineup, the Surface Laptop Ultra is a fresh effort to jump-start the Windows-on-Arm ecosystem with some powerful mobile hardware, all the while serving as Microsoft’s latest effort to woo developers and impress upon them the importance of developing for Arm, and not just for x86. The new Surface laptop is also a fresh broadside at perennial rival Apple and their MacBook Pro laptops, which have for the last several years successfully captured a surprisingly large chunk of the software developer market thanks in part to the strength of their Arm-based Apple-designed SoCs. Consequently, the Surface Laptop Ultra is set to become an important product for Microsoft on multiple levels. Even though it is not a laptop for the broadest parts of the consumer market – with pricing starting at $2600, well above the standard Surface Laptops or Surface Pros – it has the potential to be the most important Surface laptop release from Microsoft in the last several years. But it also means that Microsoft has more riding on the Windows-on-Arm ecosystem than ever before, so the Surface Laptop Ultra needs to be a clear flagship showcase for the ecosystem, and for what Microsoft’s hardware design teams can do in conjunction with NVIDIA’s. Meanwhile, launching alongside the Surface Laptop Ultra on the desktop is the Surface RTX Spark Dev box, Microsoft’s even more developer-focused RTX Spark mini-PC. Microsoft Surface Laptop Ultra Specifications First off, let us take look at the Surface Laptop Ultra’s specifications. Microsoft Surface Laptop Ultra Key Specs Processor Options NVIDIA RTX Spark N1X 18-core, NVIDIA RTX Spark N1X 20-core Operating System Windows 11 Home (Entry-Level SKU) Windows 11 Pro (All Others) Memory Options 24GB – 128GB LPDDR5X Storage Options 512GB SSD (M.2 2280, PCIe Gen4), 1TB SSD (M.2 2280, PCIe Gen5), 2TB SSD (M.2 2280, PCIe Gen5) GPU Options NVIDIA Integrated Blackwell RTX 5120 GPU core, NVIDIA Integrated Blackwell RTX 6144 GPU core Display 15-Inch PixelSense, 3270×2180 120Hz, 2000nits Max Brightness, 100% DCI, Touchscreen Battery 92Wh Form Factor Laptop Dimensions 328.8 x 238.7 x 19.16 mm (12.95 x 9.40 x 0.75 in) Weight 2.0 kg (4.41 lbs) Wireless Wi-Fi 7 + Bluetooth 5.4 Colors Platinum (Gray) Nightfall (Black) Ports Left: 1x HDMI 2.1b, 2x USB-C 40Gbps, 3.5mm Combo Audio Jack Right: 1x USB-C 40Gbps (Magnetic), 1x USB-A 10Gbps, SD Card Reader Microsoft is positioning the Surface Laptop Ultra as a premium laptop, and the specifications largely back that up. The 15-inch laptop weighs 2kg, in line with other productivity laptops and underscoring that although this is not an ultraportable laptop, it is not on the DTR-side of the spectrum, either. Surface Laptop Ultra Nightfall-Front There are several configurations of the Ultra available, with the SoC, memory, and storage all configurable to an extent. The entry-level configuration features a cut-down version of the RTX Spark N1X chip, with 18 CPU cores and a 5120 core GPU configuration. On the CPU side of matters, Microsoft has not clarified whether the cut CPU cores come from the performance (Cortex-X925) or efficiency (Cortex-A725) core pools, so it is not entirely clear right now what the performance ramifications are for the lower-tier chip. Though at 2 CPU cores it is only a small fraction of the overall core collection. The most configurable option is system RAM, which is unsurprising given the current extreme cost of RAM. Underscoring that this is not a standard consumer laptop, the minimum memory configuration here is 24GB, with further configurations for 32GB, 48GB, 64GB, and finally 128GB of LPDDR5X memory. Microsoft is taking a page from Apple’s book here and is tying memory and SoC options together: the 18 core RTX Spark can only be had with 24GB or 32GB of RAM, while the 20 core part cannot be had with anything less than 32GB. Surface Laptop Ultra Serviceability Meanwhile storage options range from a 512GB PCIe Gen4 SSD to a 2TB PCIe Gen5 SSD. Microsoft has confirmed that this storage is not soldered down and is user upgradable, and photos of the laptop posted by the company show that it accepts M.2 2280 form factor SSDs, allowing for a significant variety of upgrades if Microsoft’s offerings do not fit the bill. All models of the Surface Laptop Ultra ship with the same 15-inch “PixelSense Ultra” touchscreen display. This is a 3270×2160 pixel panel in Microsoft’s Surface-trademark 3:2 aspect ratio. The LCD panel can run at refresh rates up to 120Hz and is capable of covering 100% of the DCI-P3 color space (with a D65 white point), thanks in part to its mini-LED backlighting system. This also means that the panel can provide HDR as well, with Microsoft touting a peak brightness of a staggering 2000nits. Though it is notable that Microsoft has not received DisplayHDR certification for the display, so we will have to see how accurate (and under what conditions) those brightness figures are once the system reaches consumer hands. Surface-Laptop-Ultra-Ports-Side-by-side As for connectivity, Microsoft is making heavy use of USB-C here. All told, the laptop features three USB-C ports, all of which offer 40Gbps data rates. Meanwhile a lone USB-A port sits on the right side of the device for older devices. Alongside a dedicated HDMI 2.1b port, these USB-C ports also double as display outputs via DisplayPort Alt Mode. And being that this is a professional laptop, there is also a SD card reader on the right side of the laptop (though Microsoft has not specified what generation of cards it can accept). Surface Laptop Ultra Magnetic Connect The most novel aspect here is arguably Microsoft’s solution to offering a magnetic charging port, which has traditionally been accomplished via their proprietary Surface Connect port. But for the Surface Laptop Ultra, the company has instead developed a magnetic USB-C port, which they have dubbed “Magnetic Connect.” The magnetic USB-C port places magnets around both sides of the port to secure the charging cable, allowing the area around the port to be magnetized while keeping the port itself to USB specifications. The laptop is capable of charging from any of its USB-C ports, and Microsoft’ magnetic charging cable is not required, but only the USB-C port on the right side of the system is magnetized. Powering the system when it is not plugged in is a 92Wh battery. Microsoft has not said too much about its chemistry or cell structure, though it is otherwise just a bit under the maximum ceiling for a laptop battery. In terms of real-world performance, the company is making high-level claims of 15 hours of video playback, or 12 hours of “active web usage.” As for operating systems, in a curious move, the version of Windows that ships with the Ultra will depend on the memory configuration. 32GB and higher configurations all come with Windows 11 Pro. However, the entry-level 24GB configuration gets downgraded to Windows 11 Home, despite the laptop’s overall price tag and professional audience target. Finally, taking a look at pricing, with multiple system configurations the Surface Laptop Ultra covers a wide swath of prices. The entry-level tier laptop with an 18 core CPU, 24GB of RAM, and a 512GB kicks things off at $2600. Meanwhile the cheapest model with a full configuration RTX Spark chip, which also brings the memory up to 32GB and 1TB of storage, carriers a $3700. Finally, at the very high end, the most expensive configuration with 128GB of RAM and a 1TB SSD runs for $5900. While a hefty price tag for a laptop, this high-end configuration is surprisingly competitive; the cheapest 14-inch MacBook Pro with the same amount of memory is $800 more, so Microsoft is noticeably undercutting Apple on this front. Surface Laptop Ultra: For Creators, Developers, and AI Builders The Surface Laptop Ultra’s hardware specifications paint a picture of a laptop chiefly aimed at the professional market, and this is primarily how Microsoft is promoting the system as well. Rather than being an all-around consumer and professional laptop ala the Surface Pro or Surface Laptop lineups, the company has their eye especially focused on professional users and their workstation-light workloads for this latest laptop. The biggest factor here is that this is the most powerful (and highest memory capacity) laptop with a unified memory pool that they have been able to offer to date. We have spilled plenty of ink about the GB10 SoC in this regard over the last year, so we will not repeat ourselves in depth here, but the integrated Blackwell RTX GPU is leaps and bounds ahead of the integrated GPU in any previous Surface portable device. And because it uses the same LPDDR5X memory pool as the rest of the system, it can gain access to far more memory than any discrete NVIDIA laptop chip (which top out at 24GB). All of which allows the system to run far larger AI models than any earlier Surface laptop. Surface Laptop Ultra Development The launch of the laptop comes alongside a broader push for AI at Microsoft, which is putting together a fresh take on the subject after 2024’s Copilot+ efforts fizzled. Microsoft’s new tack is “hybrid intelligence,” using local models where appropriate, while making it easy for tasks to switch over to cloud services when additional performance is required. Though it is certainly a more
HPE announced an expanded collaboration with Oracle to help scale Oracle’s global AI infrastructure by deploying HPE Juniper Networking across Oracle’s AI data centers. The expanded collaboration builds on more than a decade of engineering work between Oracle and Juniper Networks and includes networking support services and financing capabilities.Today, HPE Juniper Networking routing and switching platforms support key elements of Oracle Cloud Infrastructure’s (OCI) data center and edge networks, providing the scale, performance, and availability needed for Oracle’s global cloud operations. That foundation is becoming increasingly important as OCI expands its AI superclusters. Large GPU clusters place significant demands on bandwidth, latency, congestion management, and fault recovery, while Oracle’s regional and edge networks must scale to connect clusters with services, customers, and external networks worldwide. “Oracle is building AI infrastructure at extraordinary scale, and HPE is uniquely positioned to support OCI with one of the industry’s most comprehensive networking portfolios for AI data centers,” said Rami Rahim, EVP, president and GM, Networking, HPE. “Together, we are advancing a high-performance networking architecture that helps Oracle address the traffic management, power, space, and operational requirements of increasingly large AI environments.” The expanded collaboration provides for the potential multi-year deployment of HPE Juniper Networking PTX and MX routing platforms and QFX and EX switching platforms across Oracle data centers globally, along with multi-year networking support services. Across OCI today, the MX and PTX routers support portions of the edge network, while QFX switches power multiple layers of the regional data center fabric. The latest QFX switches will provide the high-density connectivity, dynamic load balancing, and advanced congestion management required for large RoCEv2 (RDMA over Converged Ethernet version 2) AI backend networks. A common HPE Juniper Networking foundation across these domains helps simplify engineering and operations while enabling each environment to be optimized for its distinct requirements. “HPE Juniper Networking has been an important part of our cloud architecture and now plays a key role in our AI clusters, regional data centers, and edge networks with the scale, performance, and reliability our customers expect,” said Mahesh Thiagarajan, EVP, Oracle Cloud Infrastructure. “Building AI infrastructure at Oracle’s scale requires an integrated ecosystem where every layer of the stack works together,” said Mahesh Thiagarajan, executive vice president, Oracle Cloud Infrastructure. To help OCI sustain high performance and resiliency as its AI infrastructure scales, Oracle and HPE are also collaborating on intelligent telemetry that delivers deeper visibility into device and fabric behavior. Earlier detection of queue buildup, packet loss, traffic imbalances, and component degradation can help OCI resolve issues faster and minimize their impact on GPU utilization and network availability. “Every microsecond matters to AI models and agents,” said Praveen Jain, SVP and GM, data center and AI networking, HPE. “Through high-performance networking and our work with Oracle on deeper fabric visibility, we are helping OCI detect bottlenecks faster, maintain network availability, and improve the efficiency of AI infrastructure.” As part of the agreement, HPE has issued Oracle warrants to purchase shares of HPE common stock.
GIGABYTE W775-V10-L01-Front Angled 3 The Gigabyte W775-V10-L01 is an NVIDIA DGX Station GB300-based system that you can run at your desk. That may sound simple in concept, but in practice, the system is far beyond what you might expect. It is liquid-cooled, and with a 1.6kW maximum power draw, it can run on even most 120V circuits. We managed to get crazy performance from this in unexpected places, like Qwen3.6-35B-A3B Flash NVFP4, a very useful model, running at well over 26,000 tokens/second or over 2.25B tokens/ day. We tested the ConnectX-8 400Gbps RDMA networking and found something we did not expect. We even tested the high-speed C2C link between the CPU and GPU. VIDEO The first NVIDIA DGX Station-class system to cross our collective desk is Gigabyte’s W775-V10-L01, with Gigabyte taking the task of building a desk-side GB300-based system quite seriously. That may sound simple in concept, but in practice, the system is far beyond what you might expect. It is liquid-cooled, and with a 1.6kW maximum power draw, it can run on even most 120V circuits. We managed to get crazy performance from this in unexpected places, like Qwen3.6-35B-A3B Flash NVFP4, a very useful model, running at well over 26,000 tokens/ second or over 2.2B tokens/ day. We tested the ConnectX-8 400Gbps RDMA networking and found something we did not expect. We even tested the high-speed C2C link between the NVIDIA Grace 72-core CPU and the NVIDIA “Blackwell Ultra” (B300) GPU. We have been working on this article for some time, and it is just a wild machine. Gigabyte W775-V10-L01 Key Specs CPU NVIDIA Grace 72-Core Neoverse V2 (72C/72T) GPU NVIDIA Blackwell Ultra Operating System Ubuntu with NVIDIA AI Developer Tools Memory CPU: 496GB LPDDR5X (396GB/s) GPU: 252GB HBM3E (7.1TB/s) Storage 2x M.2 2280 (PCIe Gen6 x4) 2x M.2 2280 (PCIe Gen5 x4) Video Card Optional: NVIDIA RTX PRO Blackwell 2000/4000/6000 PSU 1600W (80 PLUS Platinum) Form Factor Tower Workstation 245 x 500 x 531mm Networking 2x 400Gbps (QSFP112, ConnectX-8) 10Gb Ethernet (RJ45, AQC-113C) 1Gb Ethernet (BMC) Ports Top: 1x USB-C 10Gbps, 2x USB-A 5Gbps, 1x 3.5mm Line-Out, 1x 3.5mm Mic-In Rear: 4x USB-A 10Gbps, 1x USB Micro-B UART (BMC), 2x 400GbE (QSFP112), 1x 10GbE LAN (RJ45, AQC-113C), 1x 1GbE LAN (RJ45, BMC), 3x 3.5mm Audio, 1x Mini DisplayPort (BMC) The W775 throws a lot of hardware at the user, and this is the crux of what makes DGX Station systems compelling. It is not just a Grace CPU or just a server-grade, HBM3E-equipped Blackwell Ultra GPU, or even a high-end dual 400Gbps ConnectX-8 network interface card. Instead, it is a combination of all of this hardware and more, all of the critical components of a GB300 server, brought down to a desktop workstation. Even calling it a GB300 server in a tower undersells the idea a bit, but it gets the concept across well enough. This is not a slice of the Grace Blackwell ecosystem like GB10-based Spark boxes, but it is the complete and unrestrained Grace Blackwell experience. Gigabyte loaned us this unit so we could make this content. We have to say this is sponsored. Gigabyte W775-V10-L01 External Hardware Overview Thanks to the combination of some very large chips inside as well as an extensive liquid cooling system, Gigabyte’s W775-V10-L01 resides in a full-size PC tower case measuring 50cm tall and 53cm deep. But more than the size, if you try to lift it, you will quickly feel the weight of the system and the liquid cooling system inside, which brings the system to a hefty 29kg (about 64 lbs). It is heavy enough that Gigabyte strongly advises using two people to move the system (team lifting), and even specifies where each person should place their hands in order to avoid lifting the system by its plastic front panel (which cannot support the weight of the system). With that said, the W775 is otherwise a fairly unassuming system given that this is a hundred-thousand-dollar workstation. The case itself is a pretty typical PC chassis with only minimal customization for the hardware inside. So this is not a heavily customized design ala DGX Spark systems, and the ability to use standard PC cases is one of the few aspects of DGX Station systems that OEMs are given any liberty to customize. With that said, Gigabyte does seem to be using a unique design here for the W775, as we have not been able to find any other Gigabyte systems with a matching case. GIGABYTE W775-V10-L01-Front Angled 1 Surprisingly, the front of the W775’s case is merely decorative, rather than serving any cooling function. Gigabyte is not using the front of the system as an airflow intake (as we see on most full tower systems), and instead they are making extensive use of every other side of the case as either an intake or exhaust. This does help with acoustics some, as it means there are not any fans directly facing forward towards the user. GIGABYTE W775-V10-L01-Side 2 On the left side of the case we find a vent cutout covering a couple of fans to serve as an intake. GIGABYTE W775-V10-L01-Side 1 Meanwhile, on the right side of the case, we catch the back side of a radiator, with three fans pushing air through it as part of the system’s extensive liquid cooling system. This also means that there are no I/O ports or buttons on the front of the case. Given the size, weight, and the amount of heat this system puts out, Gigabyte intends for the W775 to be a floor computer, so its ports and controls have been configured accordingly, placing them along the top side of the case. GIGABYTE W775-V10-L01-Top IO 1 For all of the W775’s I/O capabilities, the top panel I/O is actually a bit spartan. Here we find a pair of 5Gbps USB-A ports and a single 10Gbps USB-C port towards the far end. You will not find any USB4-speed ports on the system. Any transfers faster than 10Gbps will need to go through the system’s massive throughput 400Gb QSFP112s via the integrated ConnectX-8 SuperNIC. Actually, the 10GbE NIC also goes through the ConnectX-8 as well. While this is absolutely nitpicking on our part, for the price of the system it is surprising to see that Gigabyte has not bothered to use any kind of official (and suggested) USB-IF logos to identify the USB ports. Especially as they have opted not to use blue plastic here for the USB-A ports, there is nothing to indicate that these are USB 3.x ports. GIGABYTE W775-V10-L01-Audio + Mic Jack 1 In any case, along with the trio of USB ports, the top of the system also includes discrete headphones-out and mic-in jacks for audio, saving a trip to the back of the heavy system. GIGABYTE W775-V10-L01-Rear Angled 2 Speaking of the rear of the system, here is a look at that. Since this is a (largely) standard PC chassis, we see the familiar fan, rear I/O, expansion slot, and PSU layout of the ATX form factor. The chassis can accept either a 120mm or 140mm fan for the rear exhaust, with Gigabyte opting for a 120mm Sunon fan. GIGABYTE-W775-V10-L01-Rear-IO-2 For the rear I/O, the system is dominated by networking. Including the 1GbE port for the BMC, there are four different Ethernet ports on the W775, which are routed to three different controllers. GIGABYTE-W775-V10-L01-QSFP-Ports-3 Chief among these are the two QSFP112 ports, which are fed by NVIDIA’s powerful ConnectX-8 “SuperNIC”. Each port offers 400Gbps networking, and like the QSFP ports on the DGX Spark, these are the primary means by which the system is meant to network with high-speed storage and/or other DGX Station-based systems. The ConnectX-8 SuperNIC driving these ports does a lot of heavy lifting beyond networking, and the capabilities and limitations of these QSFP ports are somewhat complex as a result. We will get into that more when we talk about the system topology. Gigabyte W775 NVIDIA ConnectX-8 SuperNIC Ports with two 400G DR4 Optics Large We tried both QSFP112 DACs and optics in here. Something I will point out is that there is not a ton of airflow around the optical cages, so the NVIDIA MMS1X00-N5400 QSFP112 optics at around 10W ended up working well for us. These cost a bit more, but in the data center tours we do, the data center tech tend to tell us that NVIDIA’s optics work better than some low-cost brands when installed in hot aisles. GIGABYTE-W775-V10-L01-Rear-IO-RJ45-USB-1 For more mundane networking use cases, the system also has a pair of RJ45 Ethernet jacks. The left port, labeled ‘M’, connects to the BMC and is for system management. Management is handled by OpenBMC with a HTML5 iKVM and everything, but since this is an enormous review, we are not going too deep into that. The RJ45 jack on the right side is a 10Gb Ethernet port powered by Marvell’s tried-and-true AQC113C controller. The topology of this system is fascinating, so the AQC113C actually is attached to the ConnectX-8 SuperNIC, not the CPU directly. Below are the rest of the W775’s USB ports. These four ports all run at 10Gbps, matching the speed of the USB-C port on the top of the system. Unlike the top USB-A ports, Gigabyte (or rather, NVIDIA) used blue plastic here to indicate these are high-speed ports. That said, Gigabyte’s block diagram indicates these ports route through a hub, so you shouldn’t expect to get 10Gbps out of each one simultaneously. GIGABYTE-W775-V10-L01-Audio-Jack-4 To the right of that, we have a trio of audio jacks, all powered by a Realtek ALC4080 audio codec. GIGABYTE-W775-V10-L01-Rear-IO-BMC-Display At the end of the rear I/O panel, we have two more ports that are tied to the BMC. On the left is a serial UART in a micro-USB form factor. To the right is a Mini DisplayPort output from the BMC for system management and configuration. DGX Station systems do not have a true integrated display GPU within them (Blackwell Ultra is more of a compute accelerator), so the BMC is the only graphics output out of the box. For a full-featured graphics solution, DGX Station systems can be paired with an NVIDIA RTX PRO Blackwell video card. Briefly, we also find a pair of plugs here. While th