NVIDIA is a leader in the industry with its powerful AI GPUs, and it is set to strengthen this position with the new generation B200 Blackwell chips. However, a startup named Etched claims to have developed a faster and more economical solution: the Sohu ASIC (Application Specific Integrated Circuit) chip.
Etched’s Sohu ASIC chip is 20 times faster than NVIDIA GPUs
The new Sohu chip from the startup Etched offers superior performance compared to traditional GPUs. Thanks to its ASIC structure, this chip is optimized for specific workloads like machine learning, providing performance 10 times faster than NVIDIA B200 GPUs and 20 times faster than the most common accelerator, the NVIDIA H100 (Hopper) GPUs.
For example, in tests conducted on Meta’s Llama 70B architecture, a server using 8 Sohu chips can process 500,000 tokens per second. To compare, this processing power is equivalent to 160 NVIDIA H100 GPUs. Sohu achieves its high processing power by using a greater number of specialized transistors, providing the “raw power” needed for AI systems.
ASIC can reach a usage rate of 90% at full load, which is twice that of traditional GPUs. While GPUs are programmable and flexible, they cannot achieve the high efficiency of ASICs. According to Etched, GPUs do not make significant advances from generation to generation. Performance increases are typically achieved through larger chip sizes and higher energy consumption.
On the other hand, ASICs, optimized for specific tasks, can be more efficient and powerful. If the promised performance is realized, major data center providers will closely follow this new technology. There is no information yet on the release date of Sohu, but production will be done using TSMC’s 4 nm process.
This new technology could be a serious competitor to NVIDIA GPUs. What do you think about these new ASIC chips? Don’t forget to share your thoughts with us in the comments below!
{{user}} {{datetime}}
{{text}}