Tensordyne's 17x more efficient log-math chips beat NVIDIA Blackwell
New hardware claims 17x tokens per watt using logarithmic number system.
Deep Dive
Tensordyne announced a breakthrough inference system that uses Logarithmic Number System (LNS) hardware to replace multiplications with additions. By solving the long-standing "addition problem" in LNS, they free up die space for more SRAM cache, reducing power and cost while boosting performance for AI workloads.
Key Points
- 17x more tokens per watt and 13x higher throughput than NVIDIA Blackwell using logarithmic math.
- Replaces floating-point multiplication with simple log-domain addition, reducing transistor count and power.
- Solves the decades-old 'addition problem' in LNS hardware, enabling practical AI inference chips.
Why It Matters
If real, this could slash AI inference costs and energy use while boosting performance beyond current GPU limits.