Etched Sohu vs. Nvidia: Transformer ASIC vs. GPU (2026) – Spheron Blog
Summary
Etched AI's Sohu transformer-only ASIC claims very high per-chip throughput for autoregressive transformer inference, contrasting with NVIDIA GPUs that are programmable and broadly supported. The article notes Sohu's fixed-function design cannot run vision, diffusion, MoE with dynamic routing, SSM/Mamba, or training, and requires a proprietary toolchain, with first racks due in 2026 and no independent benchmarks yet. It framingly compares Sohu to Groq's LPU and GPUs, emphasizing architectural rigidity, supply risk, and migration costs as key decision factors. A practical framework is offered for when ASIC bets pay off, including baseline GPU benchmarks and cautious adoption timelines.