Etched.ai, a start-up based in Cupertino, California, has introduced its Sohu ASIC, a specialized chip designed specifically for transformer architecture-based artificial intelligence models. According to the company, Sohu is capable of processing over 500,000 tokens per second in Llama 70B throughput, surpassing the performance of leading AI chips from Nvidia and other major players.
Key Points:
- Sohu ASIC claims to be the fastest AI chip of all time
- Processes over 500,000 tokens per second in Llama 70B throughput Designed specifically for transformer architecture-based AI models, such as ChatGPT, SD3, and Sora
- Partners with TSMC on 4nm process and secured HBM and server supply from top vendors
- Raised $120M in funding from multiple investors, including Primary Venture Partners and Positive Sum