Etched AI is a California-based startup founded in 2022 that specializes in developing hardware specifically designed to accelerate transformer-based AI models. Its flagship product, the Sohu chip, is an application-specific integrated circuit (ASIC) optimized exclusively for running transformer architectures, which are foundational to many modern AI systems. By focusing solely on transformers, the chip offers significant improvements in performance and energy efficiency compared to general-purpose GPUs. Backed by $120 million in Series A funding, including investments from notable figures like Peter Thiel, Etched AI aims to reshape the AI hardware landscape with highly specialized, high-performance solutions.

The Birth of Etched and Sohu

Etched is a company founded by Gavin Uberti and Chris Zhu, who left Harvard. It is backed by an investment round of approximately $120 million. Etched's vision is to design efficient ASIC chips optimized for transformer-based AI models that channel their power into a very specific task. This initiative seeks to challenge NVIDIA’s dominant position in the AI hardware market.

Technical Specifications of Sohu

Sohu is an ASIC chip manufactured with TSMC’s 4nm technology, and it is optimized exclusively to run transformer architectures. Unlike conventional GPUs, it does not contain hardware designed for general-purpose computing, but instead focuses solely on a specific function. This results in significantly higher efficiency and energy savings.

^{Sohu: The World's First Transformer-Specific ASIC Chip}

Performance and Efficiency

The key advantage of Sohu lies in its speed and cost-effectiveness, claimed to be 10 times faster and more efficient than NVIDIA’s latest GPU models. In benchmark tests, 8 Sohu chips can replace 160 NVIDIA H100 GPUs.

• With 144 GB of HBM3E memory, it is capable of running trillion-parameter AI models.
• Its massive single-core architecture processes data more directly and efficiently.
• Lower power consumption enhances the energy efficiency of AI infrastructures.

^{Performance Comparison: Sohu vs NVIDIA Products}

Specialized Software Support

Sohu provides fully open-source software support, giving developers more flexibility. With a software stack specifically optimized for the transformer architecture, it eliminates unnecessary software layers commonly found in traditional AI hardware, allowing for more direct and efficient computation.

Use Cases

Sohu’s low latency and high efficiency make it particularly suitable for:

• Real-time voice assistants: It can process thousands of words in milliseconds, enabling more fluid and meaningful human-machine interactions.
• Code generation and deep search algorithms: It can compare thousands of responses in parallel, making the coding process more efficient.
• Multicast speculative coding: It can produce real-time outputs for more complex AI models.

Market Potential and Risks of Sohu

Sohu is positioned as a more energy-efficient and cost-effective alternative for companies offering AI solutions. While NVIDIA’s dominant position in the AI market makes it difficult for competitors to gain market share, solutions that focus on specific areas may become more widely adopted over time.

However, the long-term success of Sohu hinges on whether the transformer architecture remains dominant in the AI field. CEO Uberti is aware of this risk and makes the following striking comment:

“If transformers disappear, we disappear. But if the architecture endures, we become one of the world’s biggest companies.”

Etched and Sohu could redefine the future of AI hardware. This project shows that specially optimized AI chips may offer a competitive edge against general-purpose but powerful GPUs. Whether this venture succeeds or not will be determined in the coming years. If Etched can bring Sohu to market in time and meet the growing demands of the AI industry, it may mark a new threshold in AI hardware development.