Tech news in 3 minutes
Anthropic Halts Access to Fable 5 and Mythos 5
Tensordyne announced Napier, a 3nm AI processor and rack-scale inference platform built on proprietary logarithmic mathematics that converts multiplication into addition, reducing multiplier area and increasing on-chip SRAM. The 138-billion-transistor chip (TSMC 3nm) delivers 2.1 petaflops per die, with a 1.33GHz accelerator core, 1.5GHz CPU, 256MB SRAM (5x NVIDIA Blackwell), and 144GB HBM3E. The TDN72 rack system integrates 72 nodes, 68 petaflops total compute, and 42TB HBM, targeting models up to 10–20 trillion parameters. Tensordyne claims one 120kW air-cooled rack achieves 1,300 tokens/s/user for a two-trillion-parameter GPT MoE model, matching nine racks (1.5MW) of NVIDIA/Groq or fourteen racks (800kW) of AWS/Cerebras. Interconnect uses TDN Link with sub-microsecond latency and 1TB/s bandwidth across the 72-chip system, supporting flexible chip grouping. Software includes a Hugging Face model hub, direct PyTorch/Triton compilation, and a custom Python eDSL. Partners include HPE and Juniper for chassis and infrastructure. Beta programs are planned for Q1 2027 with system shipments by end of Q2 2027. The company aims to improve inference speed and cost by balancing compute, memory, and interconnect, but faces challenges from established ecosystems and the need for real-world validation.