Tech news in 3 minutes

Chips Media Signs Next-Gen AV2 Video IP Licensing Deal with North American Big Tech, Strengthening Global Standards Leadership

27 d ago

AI inference is rapidly becoming the key battleground for next-generation computing, moving from training headlines to real-world business, service, and infrastructure challenges. Every chatbot, agentic workflow, code assistant, scientific model, and enterprise copilot depends on running large models efficiently at scale. This context underscores Semidynamics’ appearance at ISC High Performance 2026 in Hamburg this week, where the company is presenting a full silicon-to-rack inference stack, positioning itself as a systems-level AI infrastructure company rather than just a RISC-V IP supplier. The core message: inference performance is no longer defined by peak TOPS alone. Instead, it depends on usable compute after considering memory, data movement, latency, model size, and rack-level integration. Modern AI workloads are increasingly memory-bound; large language models require massive movement of weights, activations, and KV-cache data, with growing context windows and agentic AI sessions dramatically increasing memory footprints. Semidynamics attacks this problem at the architectural level with an “all-in-one” RISC-V design integrating scalar, vector, and tensor processing more tightly than traditional accelerators. Its Gazzillion Misses technology hides memory latency, while vector and tensor capabilities handle mixed inference workloads. At ISC HPC 2026, Semidynamics is showcasing 3nm silicon, boards, and liquid-cooled OCP-compliant racks—a major shift from selling IP cores to offering a full-stack data center inference platform. This matters for three reasons: First, AI infrastructure buyers need to reduce total cost of ownership; memory-centric architectures that improve usable compute can yield large economic benefits at scale. Second, the market needs credible alternatives to dominant incumbents, and a European RISC-V company showing a complete inference stack is strategically significant for AI compute sovereignty. Third, inference workloads are evolving with agentic AI—involving planning, tool use, memory, and multi-step reasoning—shifting bottlenecks from raw matrix math to memory capacity, bandwidth, latency, and system orchestration. While Semidynamics must still prove itself in silicon, software maturity, and customer deployments, its direction is notable. The company argues that the next phase of AI inference will be won by architectures designed around memory realities from the start—a conversation HPC needs to have.