The first specialized ASIC is designed exclusively for transformer architectures, optimizing AI performance to unprecedented levels.
Etched has developed Sohu, the first specialized chip (ASIC) for transformers, which is essential for technologies like ChatGPT.
Sohu is designed for transformer architectures, limiting compatibility with other AI models such as those in Instagram ads, protein folding, older image models, and networks like CNNs and RNNs. It excels with transformers and high speeds. The Llama 70 B model processes over 500,000 tokens per second, surpassing even the latest NVIDIA Blackwell (B200) GPUs in speed and cost.
Every advanced AI model, including ChatGPT and Stable Diffusion 3, uses a transformer architecture. If future technologies move beyond transformers, Sohu’s relevance might decrease. Yet, if predictions hold, Sohu could transform the field.
The principle of scaling is crucial for AI superintelligence. In five years, AI has surpassed human performance in standardized tests through computational power, as shown by Meta’s Llama 400B model training.
The computational demands for AI are increasing rapidly, with significant companies ready to invest over $100 billion in the next few years. As we reach the limits of current technologies, specialized chips like Sohu are becoming necessary. Despite changes in AI architectures, transformers have kept a consistent design, making specialized chips a wise investment. This consistency and high computational needs make a strong case for custom ASICs.
Sohu is a significant development in AI hardware. Optimized for transformers, it achieves over 90% FLOPS utilization—a contrast to general GPUs’ broader functionality but lower efficiency.
This focus on transformers positions Sohu as a leader in AI acceleration and a vital part of managing future AI computational demands.
The company claims that if its strategic bet on Sohu is correct, it will alter not only the AI landscape but potentially the broader technological and computational paradigms.
