AI Chip Startup Tensordyne Bets on Logarithm Math to Outpace Nvidia

AI infrastructure startup Tensordyne has taped out its first commercial accelerator, the Napier chip, with fabrication on TSMC's 3nm process already underway. The chip uses an unorthodox approach to mathematics that relies on logarithms to make matrix multiplication heavy AI workloads less computationally intensive.

In conventional computing, addition is cheap and multiplication is expensive. Logarithms flip this on its head: multiplication becomes an addition problem, as a*b becomes log(a) + log(b). Tensordyne cofounder Gilles Backhus told The Register that the company uses a heuristic called the Mitchell approximation to estimate log and antilog for each value, then implements a section-wise correction mechanism in hardware to deliver accuracy equivalent to FP16.

Napier Chip Specs and Performance Claims

Tensordyne's Napier chip boasts a 300-watt nominal TDP, 144 GB of HBM3e memory across four stacks, 4.7 TB/s of memory bandwidth, and up to 2.1 petaFLOPS of dense FP8 performance. This makes it roughly comparable to Nvidia's H200 accelerators announced in 2023, while using nearly 60 percent less power.

• Each chip features roughly a terabyte of interconnect bandwidth, allowing for rack-scale deployments of up to 72 accelerators per pod.
• Tensordyne says its rack systems will achieve up to 17x more tokens per watt and 13x higher throughput than Nvidia's Blackwell systems.
• The TDN72 system consists of eight air-cooled compute blades, each with a single 10-core Intel Xeon-D host CPU and nine Napier accelerators.
• Each chip connects to six proprietary fabric switch blades developed by Tensordyne's networking partner Juniper in an all-to-all fabric.
• Up to four 30 kW TDN72 systems can be packed into a 52U rack, delivering 608 petaFLOPS in a 120 kW footprint.

Market Context and Competition

Tensordyne's TDN72 system is set to launch next year, competing against Nvidia's next-generation Vera Rubin and Vera Rubin Ultra systems. The startup is leaning heavily on the scalability of its accelerators rather than individual performance, with each chip connecting to a high-speed interconnect fabric topology reminiscent of Nvidia's GB200 NVL72 rack systems.

However, the company faces stiff competition from other AI chip makers. Amazon is reportedly considering selling its Trainium AI chips to data centers, a move that would challenge Nvidia's dominance. Meanwhile, Google is using Nvidia's own playbook of financial guarantees and circular financing to win data center customers for its own silicon, according to a Wall Street Journal investigation.

Backhus told The Register that Tensordyne's TDN72 will be much smaller than Nvidia's NVL72 racks and won't require liquid cooling, making it easier to deploy in older brownfield data centers. The company claims its rack systems will achieve 1.68x more dense FP8 compute per rack than Nvidia's GB200 NVL72, though peak FLOPS are not representative of real-world performance.