June 26, 2024, 4:34 a.m.
3923
None
Brief news summary
Etched, a startup specializing in transformer-focused chips, has introduced Sohu, an application-specific integrated circuit (ASIC) that claims to surpass Nvidia's H100 in AI LLM inference. A single 8xSohu server supposedly matches the performance of 160 H100 GPUs, potentially reducing costs for data processing centers. Current AI accelerators are designed to support various AI architectures, resulting in a large portion of computing power being allocated to programmability. However, the popularity of the transformer AI architecture has prompted Etched to develop Sohu specifically for transformer models. This move could pose a threat to Nvidia's dominance in the AI space. Additionally, Sohu's specialized design may help address the power consumption concerns associated with AI data centers.Startup Etched has unveiled a new application-specific integrated circuit (ASIC) called Sohu, which claims to outperform Nvidia's H100 in AI language model inference. This chip is specifically designed for transformer models, allowing it to allocate more transistors for AI computation.
Sohu's launch poses a potential threat to Nvidia's dominance in the AI space, as companies that exclusively use transformer models may migrate to Sohu for its efficiency and affordability. By reducing power consumption, Etched's approach could also address concerns about the environmental impact of power-hungry AI infrastructure.
Watch video about
None
Try our premium solution and start getting clients — at no cost to you
