Member-only story
Etched claims that “Sohu is >10x faster and cheaper than even NVIDIA’s next-generation Blackwell (B200) GPUs”
A Game Changer in AI Processing
Etched, a two-year-old AI startup founded by Harvard dropouts, has made headlines with its new achievement. The company has raised $120 million in Series A funding to develop Sohu, a specialized AI chip designed to revolutionize the processing of transformer models. Unlike general-purpose GPUs, Sohu is an application-specific integrated circuit (ASIC) optimized solely for transformers, promising unprecedented speed and efficiency.
Specialized Design for Unmatched Performance
Sohu’s key differentiator lies in its specialized design. Unlike Nvidia’s general-purpose GPUs, which allocate a significant portion of their transistors to general programmability, Sohu dedicates its architecture entirely to the critical tasks of transformer models. This focus allows Sohu to perform an order of magnitude faster and more efficiently. Etched’s is an ASIC approach which involves burning the transformer architecture directly into the silicon, resulting in a chip that can outperform Nvidia’s latest GPUs significantly.
