US AI startup Tensordyne claims 3nm Napier chip outperforms NVIDIA Blackwell by 13x in tokens per second - TweakTown

US AI startup Tensordyne has announced the successful tape-out of its 3nm Napier chip, claiming a 13x performance advantage over NVIDIA's Blackwell in tokens per second and a 17x improvement in tokens... Read original →

If Tensordyne’s Napier delivers on its claims, it will trigger a stack-wide redesign: its logarithmic math replaces multipliers with adders, reallocating silicon to on-die SRAM—five times Blackwell’s—undermining HBM3E’s dominance in inference and pressuring Micron and Samsung. Geopolitically, reliance on TSMC’s 3nm EUV node in Taiwan, China exposes it to U.S. export controls on advanced semiconductors, risking production delays. NVIDIA will likely accelerate Rubin Ultra and deepen FP8 software optimization, while Broadcom may pivot to UCIe-based chiplets to bypass monolithic scaling limits. Within 18 months, successful Q2 2027 volume ramp could shift AI infrastructure toward memory-compute co-design, favoring MoE and edge inference—but without a robust compiler stack and developer ecosystem, raw transistor advantage won’t guarantee market entry.