The world's first Transformer-specific AI chip Sohu is launched: it can process 500,000 tokens per second, 20 times faster than NVIDIA H100

Etched The company announced the completion of a US$120 million (currently approximately RMB 873 million) Series A financing.Will be used to develop and sell the world's first Transformer application-specific integrated circuit (ASIC) chip Sohu.

The world's first Transformer-specific AI chip Sohu is launched: it can process 500,000 tokens per second, 20 times faster than NVIDIA H100

According to public information, Etched was founded by two Harvard dropouts, Gavin Uberti and Chris Zhu, and was established less than two years ago.

The biggest highlight of the Sohu chip is that the Transformer architecture is directly etched into the chip. Uberti said that Sohu is manufactured using TSMC's 4-nanometer process, and its reasoning performance is much better than GPU and other general-purpose artificial intelligence chips, while also consuming less energy.

At Llama 70B throughput, Sohu can process over 500,000 tokens per second, allowing users to build products that are impossible with GPUs.

Sohu is capable of real-time voice agents, millisecond-level processing of thousands of words of text, more powerful code tree searches, parallel comparison of hundreds of responses, multicast speculative decoding, and real-time generation of new content, making it possible to run trillion-level parameter models in the future.

statement:The content is collected from various media platforms such as public websites. If the included content infringes on your rights, please contact us by email and we will deal with it as soon as possible.
Information

DingTalk will be open to all AI large model manufacturers, with the first batch of 7 companies joining

2024-6-27 8:44:24

Information

OpenAI and US pharmaceutical giant Eli Lilly have reached a cooperation to use generative AI to develop new antibacterial drugs

2024-6-27 8:46:25

Search