Tensordyne Bets on Logarithmic Number Systems to Slash AI Compute Power

Tensordyne has taped out its Napier AI accelerator on TSMC 3nm, utilizing a logarithmic number system with hardware-based error correction to replace traditional floating-point multipliers and improve power efficiency for datacenter workloads.

Chips & Climate Desk
Updated Nov 7, 2026 2 min read
Tensordyne Bets on Logarithmic Number Systems to Slash AI Compute Power

Tensordyne has taped out its first commercial AI accelerator, Napier, fabricated on TSMC’s 3nm process. The architecture distinguishes itself by eschewing conventional floating-point multipliers in favor of a Logarithmic Number System (LNS). According to The Register, the design aims to reduce the computational intensity of matrix multiplication by converting operations into additions. To mitigate the precision errors inherent in LNS, Tensordyne utilizes the Mitchell approximation for log and antilog values, augmented by a proprietary section-wise correction mechanism in hardware. The company claims this approach achieves accuracy equivalent to FP16, while also supporting FP8 and 4-bit block floating data types.

The Napier accelerator is designed for high-density inference, featuring a 300-watt nominal TDP. Each chip is equipped with 144 GB of HBM3e memory across four stacks, delivering 4.7 TB/s of memory bandwidth. In terms of raw performance, the silicon is rated for up to 2.1 petaFLOPS of dense FP8 performance. According to The Register, these specifications position the chip as roughly comparable to Nvidia’s H200 accelerators, albeit with a 60 percent reduction in power consumption. Tensordyne developed the accelerator in collaboration with Broadcom and Juniper Networks, focusing on a system-level approach to scalability rather than relying solely on individual chip performance.

Tensordyne’s rack-scale deployment, the TDN72, consists of eight air-cooled compute blades. Each blade houses a 10-core Intel Xeon-D host CPU alongside nine Napier accelerators. The system utilizes a high-speed fabric topology developed with Juniper Networks, which connects each chip to six proprietary fabric switch blades in an all-to-all configuration. Unlike Nvidia's liquid-cooled GB200 NVL72, the TDN72 is air-cooled, which Tensordyne suggests will facilitate deployment in brownfield datacenters. The company reports that four 30 kW TDN72 systems can fit into a 52U rack, resulting in a 120 kW footprint capable of 608 petaFLOPS.

The company makes aggressive performance claims, suggesting the TDN72 will deliver 17x more tokens per watt and 13x higher throughput than Nvidia’s Blackwell systems. According to The Register, this density translates to 1.68x more dense FP8 compute per rack than the GB200 NVL72. However, these figures are based on peak performance metrics, which often diverge from real-world execution. Tensordyne cofounder Gilles Backhus acknowledged that max achieved FLOPS frequently fall short of peak ratings, advising that comparisons should be viewed with caution until the hardware is available for independent benchmarking.

Software maturity remains a critical hurdle for the startup. Earlier prototypes required quantization-aware training to adapt models, a process Tensordyne deemed unfeasible for trillion-parameter models. The current compiler stack is designed to convert existing models directly for the Napier architecture, similar to approaches seen at other startups like Tenstorrent. The company is also developing a proprietary serving platform and runtime environment to support existing tools like vLLM, with PyTorch support currently in development. With the Napier hardware slated for release in Q2 or Q3 of 2027, Tensordyne faces the challenge of ensuring software compatibility against upcoming competitors like Nvidia’s Vera Rubin platforms.

Sources

  1. 01 Tensordyne makes a big bet on log math to beat Nvidia — The Register — HPC