“Etched Secures $120M to Compete with Nvidia in AI with Advanced Transformer Chips”
GamesBeat has teamed up with Lil Snack to offer custom games tailored just for their audience. As gamers ourselves, we see this as a fantastic way to engage with the GamesBeat content you already enjoy. You can start playing these games here.
Etched has announced raising $120 million to challenge Nvidia in the AI chips space. The company is working on a new chip called Sohu, which focuses on a crucial part of AI processing: transformers. By integrating transformer architecture into their chips, Etched aims to produce the world’s most powerful servers for transformer inference, claiming Sohu is the fastest transformer chip ever created.
This is noteworthy given Nvidia’s significant market presence, having surpassed Microsoft as the world’s most valuable company with a $3.3 trillion valuation. Etched, despite having only 35 employees, believes it can outperform Nvidia, attracting backers like Peter Thiel.
The funding round was led by Primary Venture Partners and Positive Sum Ventures, with contributions from institutional investors such as Hummingbird, Fundomo, Fontinalis, Lightscape, Earthshot, Two Sigma Ventures, and Skybox Data Centers. Angel investors included notable figures like Peter Thiel, Stanley Druckenmiller, David Siegel, and others.
Alex Handy, director of the Thiel Fellowship, stated that investing in Etched is a strategic move given the value of AI. He highlighted that Etched’s chips address scalability issues that competitors avoid, making it a promising venture in the semiconductor industry.
Etched was founded by three Harvard dropouts: Robert Wachen, Gavin Uberti, and Chris Zhu. In a blog post, the founders mentioned that they made a significant bet on AI in June 2022, focusing on the potential of transformer models, which have since become dominant in the AI landscape. They spent the last two years developing Sohu, an ASIC chip specialized for transformers, which is significantly faster and cheaper than Nvidia’s upcoming Blackwell GPUs.
Uberti explained that despite not supporting other AI models like CNNs or RNNs, Sohu excels at transformer inference, which powers most major AI models today. If other architectures replace transformers, Etched’s chips would become obsolete, but Uberti is confident that this won’t happen anytime soon.
Etched is partnering with TSMC to start production on their 4nm node, aiming to revolutionize the AI chip market. Uberti emphasized that specialization in AI chips is inevitable as the demand for transformer models grows. Despite the challenges, Etched aims to be the first company to succeed in building an architecture-specific AI chip and is moving quickly to make Sohu a reality.
Uberti mentioned that if transformers maintain their dominance, Sohu will significantly impact the world, reducing the costs and improving the speed of AI models. Etched believes that specialized chips will eventually dominate the market due to their efficiency and performance advantages over general-purpose chips like GPUs.
The company is confident that it will dominate the AI chip market for transformer inference, potentially becoming one of the largest companies in the world due to the transformative power of Sohu.