Sohu AI chip claimed to run models 20x faster and cheaper than Nvidia H100 GPUs

Etched, a startup, has developed the Sohu AI chip, an ASIC designed for AI LLM inference that claims to outperform Nvidia's H100 GPUs. The Sohu chip is specialized for transformer models, allocating more transistors to AI compute tasks compared to general-purpose AI chips like Nvidia's H100. This specialization allows Sohu to run models 20 times faster and cheaper than the H100, potentially disrupting Nvidia's dominance in the AI space. By focusing on transformer architectures popularized by models like ChatGPT, Sohu aims to provide more efficient AI processing, reducing power consumption in AI data centers. While concerns over power consumption in AI infrastructure persist, Sohu's approach could offer a more sustainable solution by optimizing hardware for specific AI tasks. The success of Sohu could lead to a shift in the industry towards more specialized AI chips tailored for specific models, potentially reshaping the landscape of AI computing.