All 2 min read

Marvell Debuts Industry-First 2nm SRAM to Power Next-Gen AI Chips

Marvell introduced the industry’s first 2nm custom Static Random Access Memory (SRAM), targeting next-generation AI accelerators and data center silicon. This new memory block achieves up to 6 gigabits of high-speed SRAM and claims the industry’s highest bandwidth per square millimeter, while reducing memory standby power by up to 66% and reclaiming as much as 15% of the die area—key for dense, high-performance chip designs in hyperscale AI infrastructure.

The 2nm SRAM is part of Marvell’s broader custom silicon platform strategy, which also includes co-packaged optics, HBM integration, and CXL memory expansion. The new SRAM complements this portfolio by improving on-die memory efficiency and enabling chip designers to use reclaimed silicon for more compute cores, higher memory density, or power-optimized custom solutions. Designed to operate at frequencies up to 3.75 GHz, the SRAM uses custom circuitry and architectural enhancements co-developed with leading foundry partners.

This latest innovation reinforces Marvell’s bet on custom silicon as transistor scaling slows in the post-Moore’s Law era. The company is building out a portfolio of advanced IP—including PCIe Gen 7 interfaces, die-to-die interconnects, optical I/O, and advanced SerDes—to support customer-driven chip development in AI, cloud, and networking infrastructure.

  • Industry’s first 2nm custom SRAM delivers 6 Gbit capacity and industry-leading bandwidth density
  • Reclaims up to 15% of die area and reduces standby power by up to 66%
  • Complements Marvell’s custom silicon platform including HBM, CXL, and optical interconnects
  • Operates at up to 3.75 GHz and built for AI, cloud, and custom XPU designs
  • Aimed at enhancing compute/memory balance in data center and AI infrastructure

“Custom is the future of AI infrastructure. The methodologies and technologies used by hyperscalers today to develop cutting-edge custom XPUs will percolate to more customers, more classes of devices, and more applications,” said Will Chu, senior vice president of Custom Cloud Solutions at Marvell.


Share this article

Help others discover this reporting.

From the Technology Graveyard

You just read about All. A technology from the archive that helped get us here:

Google/Alphabet Project Loon served 2011–2021. What ultimately replaced it?

On this day

June 17

From 25 years of the Converge Digest archive.