Gerard Williams III, the former lead CPU architect at Apple, discussed the role of CPUs in AI data centers during a recent Power On interview. Williams, who led the development of Apple's A-series and M-series processors, is currently the co-founder of NUVACORE, a company focused on data center chips.
Williams argues that while GPUs act as the primary compute engines for AI, CPUs remain essential as the central schedulers for data centers. They handle critical tasks such as memory management, security protocols, and operating system operations that GPUs cannot perform alone.
NUVACORE utilizes a 'Core First' architecture that remains flexible regarding instruction sets like x86, ARM, or RISC-V. This approach allows the company to optimize designs for specific ecosystems only after the core architecture is established.
The company has significantly accelerated its design process by integrating AI tools. While traditional CPU development typically requires 2.5 to 3 years, with 7 to 12 months spent on initial modeling, NUVACORE has reduced this design phase to just 2 to 3 months.
Williams emphasizes that the power-efficiency techniques developed for mobile chips are now vital for AI infrastructure. By applying mobile-grade power management to server CPUs, he believes data centers can achieve 2 to 3 times the computing power within existing energy constraints.
