AI startup, Etched, has revealed its Sohu ASIC for AI, claiming it to be much faster and cheaper than NVIDIA's next-gen Blackwell B200 GPUs.
Etched Makes The "Biggest Bet in AI" With Sohu ASIC With Transformer Engine, Claims Big Numbers
It's not every day that firms pop up and challenge the likes of NVIDIA in AI computing performance, but it looks like Etched might be here to do just that.
The firm currently relies on the "Transformer Engine," a neural network architecture utilized in OpenAI's ChatGPT and Meta's Llama LLM AI models. However, they have decided to take things a step further by unveiling a dedicated ASIC machinery, called the Sohu, with the sole aim of providing customers with an option that is claimed to be 10x faster than NVIDIA's Blackwell and 20x faster than NVIDIA's Hopper AI GPUs.
.png)
These are certainly some bold claims made by Etched; they also have some benchmarks to back them. The firm says that one Sohu ASIC can manage to serve over 500,000 Llama 70B tokens per second, equivalent to a whopping 160 units of NVIDA's H100s.
Apart from that, Sohu can serve every modern AI model out there, whether it is Meta's Llama or Google's Anthropic, and the team has said that they can make adjustments for further integrations as well, ensuring the ASIC's widespread adoption.
-p-800.png)
"Squeezing in FLOPS" is what Sohu has done best, as the firm has revealed that their ASIC has managed to attain an impressive 90% utilization rate, double what the market offers. Etched has done this by dedicating a larger portion of transistors specializing in transformers, whereas other firms like NVIDIA focus more on programmability, which is why their ecosystem doesn't manage to squeeze out FLOPS comparable to Sohu.
Discussing the future, Etched says that they are the first ones in the ASIC race and probably one of the dominant firms in it, considering how they have managed to capitalize on the hype. They have partnered up with TSMC, employing their 4nm process in Sohu, and have reportedly raised $120 million in funding, showing that the firm has seen massive interest from the industry.
With Sohu, Etched aims to take AI real-time processing to new heights, eventually making it a commonality rather than a barrier.
News Source: Etched
About the author: Muhammad Zuhair is a hardware and technology reporter for Wccftech, specializing in the semiconductor industry and the complex interplay between technology, manufacturing, and geopolitics. His coverage focuses on the corporate strategies and technological roadmaps of industry giants like TSMC, NVIDIA, Samsung, and Intel. Zuhair's expertise lies in deconstructing complex topics such as fabrication nodes (e.g., 2nm process), the economic impact of policies like the CHIPS Act, and the strategic development of AI infrastructure from NVIDIA, AMD and Intel.
Follow Wccftech on Google to get more of our news coverage in your feeds.
