Etched is a startup developing a specialized chip optimized for transformer workloads, using ultra low voltage compute and cluster scale memory to enhance efficiency and scalability in AI processing. Founded by Harvard dropouts and backed by over $1 billion in funding, the company aims to challenge industry leaders like Nvidia with innovative hardware tailored for modern language models.
Etched is a startup emerging from stealth with ambitions to challenge industry giants like Nvidia by designing a chip specifically optimized for transformer workloads, which are central to modern language models. Founded by two young Harvard computer science dropouts, Gavin Uberti and Robert Wacken, along with CTO Mark Ross, who has a strong background in networking hardware, the company has raised significant funding—over $1 billion in recent rounds. Their focus is on creating a highly efficient, transformer-dedicated chip that leverages unique architectural innovations to deliver superior performance.
The core innovation behind Etched’s chip lies in two main technologies: ultra low voltage compute and cluster scale memory. The low voltage compute approach involves running transistors at extremely low voltages (around 0.45 volts), which drastically reduces power consumption while maintaining efficiency. This technique is similar to what is used in Bitcoin mining hardware but adapted here for transformer inference workloads. However, low voltage operation typically limits frequency, so the design must balance power savings with performance needs.
To address the data communication challenges inherent in machine learning workloads, Etched employs cluster scale memory. This involves a large SRAM scratchpad memory accessible not only within a chip but also across multiple chips via high-speed interfaces. This architecture allows for direct memory access between chips, potentially solving bandwidth bottlenecks that often limit scaling in AI accelerators. The trade-off is increased complexity and cabling, as seen in their server setups, but it could enable efficient scaling of transformer computations across many chips.
Despite these promising technologies, many details remain undisclosed, such as exact bandwidth, power consumption, and memory management strategies. The chip is reportedly taped out, and Etched has demonstrated server prototypes, but questions about compiler support and memory coherence remain open. The company is actively engaging with hyperscalers and other potential customers, aiming to scale production and deployment rapidly, with an aggressive roadmap to release new chip generations frequently.
Overall, Etched represents a bold attempt to carve out a niche in the AI hardware market by focusing exclusively on transformer workloads with specialized hardware. Their approach combines novel low voltage transistor design with innovative memory architectures to tackle both compute and communication challenges. While still early in development and with many unknowns, the company’s strong backing, experienced leadership, and unique technology have generated significant interest and anticipation in the industry.