Move over, Nvidia H100, there’s a new sheriff in town. American startup Etched has sent shockwaves through the AI hardware landscape with the release of their game-changing AI chip, the Sohu. This revolutionary chip boasts a staggering 20 times the performance of Nvidia’s current flagship, the H100, specifically designed for processing the ever-evolving world of transformer models.
Chip Sohu by Etched. Photo: Etched.
Transformer models are the powerhouse architecture behind some of the most impressive advancements in AI, powering tools like ChatGPT and Google’s LaMDA. However, their immense processing needs have pushed traditional GPUs to their limits. This is where Etched’s Sohu steps in.
Unlike traditional GPUs, the Sohu is an application-specific integrated circuit (ASIC). This means it’s custom-built for the specific task of running transformer models, allowing for a level of efficiency and power that surpasses even the most powerful GPUs. Benchmarks show the Sohu achieving a remarkable 500,000 tokens per second, compared to the H100’s 25,000 – a truly night and day difference.
This breakthrough has significant implications for the future of AI. Faster processing translates to faster development and deployment of cutting-edge AI applications. Researchers can train even more complex and powerful models, pushing the boundaries of what AI can achieve.
Etched’s focus on transformer-specific hardware represents a potential shift in the AI hardware landscape. Traditionally, GPUs have been the go-to option for AI development due to their versatility. However, the Sohu demonstrates the efficiency gains possible with specialized hardware.
While the cost and availability of the Sohu remain to be seen, its arrival marks a significant moment in AI development. With this level of processing power at their disposal, researchers and developers can usher in a new era of AI innovation.
Admin
