Tensordyne makes a big bet on log math to beat Nvidia

Tensordyne, an AI infrastructure startup, has taped out its first commercial accelerator using logarithmic math to reduce computational intensity in AI workloads, claiming higher throughput and lower power consumption than GPUs. The chip, developed with Juniper Networks and Broadcom, uses the Mitchell approximation for log/antilog estimation and aims to outperform Nvidia's Blackwell systems by 13x in throughput and 17x in tokens per watt.