Etched Raises $120M to Build Custom LLM Chips
Built for high efficiency, Etched looks to scale transformers without the cost!
Created on June 26|Last edited on June 26
Comment
A group of Harvard dropouts has raised $120 million to develop a new AI chip, aiming to compete with NVIDIA’s dominance in the industry. The startup, Etched, is based in Cupertino, California, and is focused on creating the Sohu chip, designed specifically for handling transformers, the core technology behind advanced AI models like OpenAI’s ChatGPT.
Specialized Hardware
Etched's co-founders, Gavin Uberti, Robert Wachen, and Chris Zhu, believe that the future of AI lies in specialized, hard-wired chips known as ASICs (Application-Specific Integrated Circuits). These chips are tailored to execute specific AI models more efficiently than general-purpose GPUs (Graphics Processing Units), which, while versatile, are also more costly. The Sohu chip aims to provide significant performance improvements and cost savings by optimizing for the predictable data movement of transformers.
Betting on Attention
The timing of Etched’s development is significant as transformers are increasingly being adopted in various AI applications, including image generation. This trend suggests a growing demand for chips that can efficiently handle transformer-based models. However, this also means that Etched is making a high-stakes bet: if transformers remain central to AI advancements, the company could see immense success. Conversely, if the technology shifts away from transformers, Etched could face substantial challenges.
Despite Nvidia’s strong market position, Etched is confident in its approach. The Sohu chip is designed to reduce latency and boost efficiency by embedding AI architectures directly into the hardware. This specialization is expected to outperform general-purpose GPUs in specific AI applications.
Investors
Etched has attracted investment from notable figures like Peter Thiel and Stanley Druckenmiller, as well as venture capital firms Primary Venture Partners and Positive Sum Ventures. The startup plans to launch the Sohu chip later this year and is currently working to secure customers interested in its innovative approach.
High Risk, High Reward
By focusing on the specific needs of AI models and leveraging hard-wired chips for greater efficiency, Etched hopes to make a significant impact in the AI hardware space, offering an alternative to Nvidia’s powerful but costly GPUs. If transformers continue to dominate AI development, Etched's specialized approach could position them as a major player in the industry.
Add a comment
Tags: ML News
Iterate on AI agents and models faster. Try Weights & Biases today.