AI designs an AI accelerator in two weeks, not a year – all that’s left is to test in wafers

Architect Labs, a startup founded last year, is betting that artificial intelligence can radically speed up chip design. The company claims to have developed and validated its experimental Redwood AI accelerator in just two weeks: two experts defined the high-level architecture and refined the master plan, while the main engineering work was done by AI.

Image credit: Igor Omilaev / unsplash.com

Under normal circumstances, developing a complex processor can take years, cost tens or hundreds of millions of dollars, and require a large team of experts. In the field of artificial intelligence, such long lead times are particularly problematic because models are developed faster than the wafers optimized for them can be created.

The Redwood Accelerator showcases the capabilities of an AI-accelerated design system created by startup Architect Labs. After receiving specifications prepared by two experts, she independently created the accelerator’s RTL description, UVM verification and formal verification tools, as well as firmware, drivers and dedicated computing cores.

However, mature rosewood wafers do not yet exist. This artificial intelligence-developed project is modeled on the Programmable Matrix AMD Versal FPGA, which is already capable of executing inference models with billions of parameters, including Meta✴ Please call Alibaba Qwen and Kimi.

Architect Labs also evaluated Redwood’s potential performance after moving the project to silicon. According to the company’s calculations, when implemented on Samsung’s 8nm process technology, the accelerator will be able to provide a 1.75x performance improvement and approximately 1.9x power consumption reduction compared to the Nvidia Jetson Orin Nano. Therefore, Redwood’s performance advantage per watt can reach 3.4 times. However, for now these are calculated metrics: only the release and testing of real chips can confirm them.

Architect Labs isn’t in a rush to put the project into production just yet. As artificial intelligence continues to optimize architecture, developers need to ensure that no errors are left in the design: defects discovered after the wafer is manufactured can cost millions of dollars. In the future, the company expects to transfer the project to TSMC for production.

Developers see this approach as an opportunity to spark a new transformation in the semiconductor industry. Decades ago, contract manufacturers such as TSMC enabled the emergence of fabless companies that no longer needed their own factories. Architect Labs hopes to do something similar with design, and talks about the advent of a “design-free” era, where companies will be able to order specialized chips for their tasks without having to maintain huge engineering departments.

Additionally, the startup’s founders anticipate a positive feedback: more advanced AI models will enable the design of more efficient specialized accelerators, which in turn will allow them to launch more advanced models. Redwood should be the first to demonstrate the feasibility of this approach.