EEtched
Transformer ASIC, AI inference chip, Sohu
Etched builds Sohu, an ASIC designed to run transformer inference at extreme speed. Sohu replaces GPUs for inference-only workloads, processing tokens at 500,000 per second on Llama 70B and using low-voltage silicon to stay cool at full power.
- Sohu ASIC: 500,000 tokens/sec on Llama 70B; 20x throughput vs. 8x H100 GPUs.
- Built on TSMC N4P; proprietary low-voltage (LVI) and cluster-scale memory (CSM) techniques.
- August 2026: Shipped first production racks to Jane Street; $21B valuation after $700M Series D.
etched.com↗ (opens etched.com)Company profile →