Tensordyne Tapes Out LNS-Based AI Chip, Claims Power Advantages (www.eetimes.com)

Tensordyne, an AI chip startup, has announced the successful taping out of its logarithmic number system (LNS)-based AI chip, targeting significant power efficiency improvements for data center inference. Promising up to 17 times the tokens per second per watt compared to Nvidia's GB300 systems, this chip addresses two critical challenges in AI—speed and cost—especially as models grow larger. The chip integrates proprietary number systems and hardware acceleration specifically designed for LNS, consuming 300 W per package and achieving up to 2.1 PFLOPS of dense FP8 compute. This innovative architecture features a patented cell-based network-on-chip, reducing latency and enhancing performance, particularly for complex workloads like mixture of experts (MoEs). Tensordyne's system can potentially deliver 3 million tokens per second per megawatt, vastly outperforming existing solutions in both speed and operational costs. By allowing various models to operate concurrently within its design, Tensordyne aims to revolutionize AI compute efficiency, with initial shipments expected by Q2 2027. This development suggests a pivotal shift in how data center AI efficiency is approached, matching the scalability of existing infrastructure while offering substantial advancements in speed and resource utilization.