Tensordyne is making AI compute more efficient using logarithms

Tensordyne builds AI accelerators around a logarithmic number format that stores values as log2(x), turning multiplications into additions and enabling much smaller, more efficient multipliers. That advantage trades off by making addition and accumulation harder; Tensordyne compensates with approximations (e.g., using LOG2(1+X) ≈ X for small X) plus proprietary corrections and a software toolchain that converts PyTorch models to their numeric format. Those choices constrain the platform to inference workloads: retraining or numerical retooling was used initially, kernel-level porting is necessary to match GPU-optimized performance, and training is impractical because of numerical stability and the dominance of memory and networking in training cost.

Against the prevailing industry focus on memory hierarchy improvements, Tensordyne targets compute efficiency and smaller multiplier area, which lets them pack more on-chip SRAM and reduce off-chip traffic. Under the roofline model, their hardware shines when workloads are compute-bound (high arithmetic intensity from batching), while low-latency, single-query agentic workloads typically remain memory-bound and benefit more from all-SRAM designs. The net effect is a clear tradeoff: significant efficiency and die-area gains for compute-heavy inference and some memory benefits from extra SRAM, at the cost of software complexity and potential accuracy risk. The approach is novel and could be commercially valuable if customers accept the porting and compatibility work.