Etched's Sohu Chip: A New Player in the AI Hardware Market

Etched's Sohu Chip: A New Player in the AI Hardware Market
Image Source: EtchedBy Wed, 26 Jun 2024 04:02:07 GMT

As generative AI makes its mark across various sectors, the companies manufacturing the necessary chips are reaping substantial rewards. Nvidia, in particular, has a commanding presence in the AI chip market, controlling an estimated 70% to 95% of the market share. Giants like Meta and Microsoft are investing heavily in Nvidia GPUs to maintain their competitive edge in the AI arena. However, this dominance isn't without its challenges. Many generative AI vendors, reliant on these chips, are seeking alternatives. This has led to a surge in interest from venture capitalists and AI vendors in finding new players to rival Nvidia's dominance.

One of the most promising new entrants is Etched, a startup founded just two years ago by Harvard dropouts Gavin Uberti and Chris Zhu, along with Robert Wachen and former Cypress Semiconductor CTO Mark Ross. Based in Cupertino, California, Etched focuses on developing specialized chips for AI models, specifically transformers. Transformers have become the foundation of many generative AI applications, including OpenAI's ChatGPT, Google's Gemini, and Anthropic's Claude.

Etched's flagship chip, Sohu, is an application-specific integrated circuit (ASIC) optimized for transformers. Using TSMC's 4nm process, Sohu is designed to outperform general-purpose GPUs in inferencing tasks while consuming less power. Uberti claims that Sohu is significantly faster and more cost-effective than Nvidia's latest GPUs, with a single Sohu server capable of replacing up to 160 H100 GPUs.

The efficiency of the Sohu chip stems from its streamlined design, which allocates more resources to computational power and less to memory. This is possible due to the predictable data flow in transformer models, allowing Etched to minimize unnecessary hardware components. Uberti likens the potential impact of specialized AI chips to the rise of ASICs in cryptocurrency mining, which dramatically reduces the demand for Nvidia GPUs in that field.

Despite Nvidia's substantial market lead and the challenges inherent in the semiconductor industry, Etched has attracted significant investor interest. The company recently secured $120 million in a Series A funding round, led by Primary Venture Partners and Positive Sum Ventures, with contributions from notable investors like Peter Thiel and GitHub CEO Thomas Dohmke. This brings Etched's total funding to $125.36 million.

Etched is not alone in this endeavour. Other startups, such as Cerebras Systems and Tenstorrent, are also developing innovative AI chips, aiming to capture a share of the expanding market. However, Etched's focus on transformer models sets it apart, and the founders believe their chip will be crucial for future AI applications, including video generation, robotics, and real-time voice interactions.

Etched plans to bring Sohu to market soon, with preliminary commitments already in place from several customers. The company is also launching the Sohu Developer Cloud, an interactive platform that will allow potential clients to test the chip's capabilities.

Although the path ahead is fraught with challenges, Etched's founders are optimistic. They acknowledge the high stakes but believe their specialized approach and rapid development will give them an advantage. As Uberti aptly notes, the future of AI technology depends on scalable infrastructure, and Etched aims to be at the forefront of this transformation.