New AI Chip “Sohu” Promises to Surpass Current Technology, Claims Startup Founded by Harvard Dropouts
In an impressive move, the fast-growing startup Etched has announced its groundbreaking AI chipset called Sohu, claiming it to be 20 times faster than Nvidia’s leading H100 graphics processing unit (GPU) for running transformer models like ChatGPT. This bold declaration, made by a company founded by two Harvard dropouts, has already garnered widespread attention in the tech world.
The Dawn of a New Era in AI Hardware
Sohu is an Application Specific Integrated Circuit (ASIC) designed exclusively to handle transformer AI models. Transformers are the backbone of cutting-edge AI applications, powering high-profile systems from ChatGPT to Google’s Gemini, and even OpenAI’s text-to-video model, Sora. The specialized architecture of Sohu aims to offer unparalleled computational power to process the massive parameters that these models depend on.
While GPUs like Nvidia’s H100 and B200 can manage a variety of computational tasks, Sohu is optimized to execute only transformer models. This specificity allows it to outperform the traditional GPU-heavy approach, making it not just a competitor but a potential game-changer in the AI hardware market.
Understanding Transformer Models
Transformer models excel at a wide range of tasks due to their advanced architecture of self-attention mechanisms. These models can understand and generate human-like text, making them highly specialized tools for natural language processing. However, they require immense computational power to operate efficiently. This is where Sohu steps in, providing the needed horsepower to run these models more quickly and efficiently than traditional GPUs.
Innovating Beyond the Conventional GPU
Unlike GPUs that can juggle a variety of tasks, Sohu’s focus is strictly on transformer models. This intentional limitation means it can’t run models like Convolutional Neural Networks (often used for image recognition). However, this focused approach opens up exciting new possibilities for AI development. For instance:
- A real-time multi-lingual translator that can listen and translate languages like Hindi or Gujarati into French or English almost instantaneously.
- Integration of visual and language processing capabilities, allowing for applications where the model can understand both text and images simultaneously. Imagine AI systems capable of conducting an interview by interpreting visual cues as well as verbal responses.
These examples illustrate the transformative potential of Sohu’s design, though it’s worth noting that these applications are still theoretical. The true capabilities of Sohu will become clearer only with real-world testing and deployment.
Investment and Production Plans
On June 25, Etched announced it had raised $120 million to make the vision of Sohu a reality. Although a specific release date remains unclear, the momentum is encouraging. The company has already secured substantial preorders amounting to “tens of million dollars” worth of hardware. They have also confirmed a manufacturing partnership with TSMC (Taiwan Semiconductor Manufacturing Company) to produce the chip using 4-nanometer technology, suggesting a substantial head start in terms of production capabilities.
Today we announced our $120M fundraise to bring Sohu to the world. Thanks @edludlow for hosting us! https://t.co/khzMPAofng pic.twitter.com/Q4SeuUjP2h
— Etched (@Etched) June 26, 2024
What This Means for the Future of AI
With Sohu, Etched is positioning itself as a modern-day David taking on Goliaths like Nvidia. Their specialized chip could potentially redefine how future AI applications are developed, moving away from multi-purpose GPUs towards more narrowly focused, highly efficient hardware solutions.
In essence, the success of Sohu could signal a shift in the AI hardware landscape. If this new ASIC chip can deliver on its promises, we could very well witness a new era where computational speed and efficiency in AI models drive unprecedented advancements in technology. As developers and tech enthusiasts eagerly await more concrete performance data, the possibilities seem limitless.
Mehek 
Divya