Etched Secures $120m Funding for Transformer-Only ASIC Bet

The landscape of artificial intelligence is rapidly evolving, with companies like Sohu making bold claims about the capabilities of their cutting-edge technology. Sohu boasts that its architecture, specifically designed to run transformer models, outperforms Nvidia N100s by a staggering 20 times. This advancement positions Sohu at the forefront of Large Language Models (LLMs) and applications like ChatGPT, promising a new era of AI efficiency.

However, this impressive speed comes with limitations. Sohu is not compatible with traditional AI models such as Convolutional Neural Networks (CNNs), Recurrent Neural Networks (RNNs), or Long Short-Term Memory (LSTM) models. The company also points out that it cannot execute deep learning recommendation models like DLRMs, protein-folding models such as AlphaFold 2, or older image models like Stable Diffusion 2.

Despite these constraints, Etched, the company behind Sohu, remains undeterred. They argue that the majority of state-of-the-art AI models today rely on transformer software, including popular ones like ChatGPT, Sora, Google's Gemini, Stable Diffusion 3, and many others. Etched acknowledges the risk, stating that if transformers are replaced by new architectures, their chips could become obsolete.

CEO Gavin Uberti expressed the high stakes in an interview with CNBC, emphasizing the company's dependency on transformers for success. Uberti's sentiment reflects the critical role that transformer models play in shaping the future of AI technology. The potential demise of transformers could spell disaster for Etched, but their continued relevance could propel them to unparalleled success.

Etched argues that as AI models become increasingly complex and costly to train, specialized chips like Sohu's are essential for achieving optimal performance. The company asserts that even a marginal improvement in performance justifies the need for transformer ASICs, which currently outperform GPUs by orders of magnitude. Drawing parallels to the evolution of bitcoin mining chips, Etched predicts a similar transition in the AI industry, where GPUs will be supplanted by specialized chips.

Sohu's specialization in transformer inference sets it apart, focusing on algorithms like Llama and Stable Diffusion 3. By optimizing for a single algorithm, Sohu can eliminate much of the control flow logic, allowing for a higher density of mathematical operations. Etched claims that Sohu achieves over 90 percent FLOPS utilization, a significant leap compared to the approximately 30 percent utilization on GPUs.

Founded by CEO Gavin Uberti and Chris Zhu, a teaching fellow at Harvard University, Etched has secured $120 million in funding to advance its ASIC technology. The company plans to move forward with manufacturing through TSMC, although specific details regarding the manufacturing process node and chip availability remain undisclosed. As Sohu continues to make waves in the AI industry, its innovative approach to transformer models could reshape the future of artificial intelligence.