Etched, a startup focusing on transformer-based chips, has unveiled Sohu, an ASIC aimed at outperforming Nvidia’s H100 in AI LLM inference. The company claims that a single 8xSohu server can offer the performance of 160 H100 GPUs, potentially saving data centers on upfront and operational costs.
Current AI accelerators, be it CPUs or GPUs, are designed to support various AI architectures, leading to a significant portion of computing power being devoted to programmability. Etched’s Sohu chip is specifically designed for transformer-based models, which are gaining popularity in the AI space.
Transformative AI architectures like ChatGPT, Sora, Gemini, Stable Diffusion, and DALL-E are all based on transformer models. Etched made a strategic decision to focus on transformers with the Sohu project, allocating more transistors for AI computing tasks like matrix multiplication.
By specializing in transformer models, Sohu aims to optimize AI computing efficiency and speed, challenging Nvidia’s dominance in the AI field. As demand for AI GPUs grows, the energy consumption of AI data centers has raised concerns among experts and government officials.
Etched’s Sohu chip could potentially reduce energy demands in AI computing, making it more sustainable and environmentally friendly. The efficiency of running transformer models on specialized hardware like Sohu could revolutionize the AI industry, paving the way for a more energy-efficient future as computing needs continue to grow.
Article Source
https://www.tomshardware.com/tech-industry/artificial-intelligence/sohu-ai-chip-claimed-to-run-models-20x-faster-and-cheaper-than-nvidia-h100-gpus
