Marvell’s 2nm Custom SRAM is the industry’s first custom-built on-chip memory at the 2nm node, designed specifically to meet the evolving needs of AI infrastructure. This groundbreaking memory solution delivers up to 6 gigabits of high-speed SRAM, while saving up to 15% of total die area and reducing standby power by up to 66%—a transformative improvement in performance and power efficiency for next-generation XPUs.

Unlike conventional SRAM, Marvell’s custom design leverages proprietary data path architecture, advanced process techniques, and co-optimized hardware/software interfaces to maximize bandwidth and silicon utilization. Capable of operating at up to 3.75 GHz, the solution provides the highest bandwidth per mm² in the industry, enabling chip architects to either increase compute, shrink chip size, or reduce cost—all essential to AI cluster scalability.

This custom SRAM builds on Marvell’s holistic memory strategy alongside innovations in HBM, CXL, and die-to-die interconnects. Together, they offer a fully integrated memory hierarchy for custom AI silicon, empowering hyperscalers and data center providers to scale up with performance and efficiency. The solution is already being adopted by industry leaders focused on extending the frontier of AI compute.