GF Pairs 22nm/40nm Edge AI Platforms With Gain-Cell GCRAM Memory

GlobalFoundries opened two fronts in the edge-AI battle on September 2, 2026, at its annual Technology Summit. The company announced customer availability of a new UX platform family — 40UX on 40nm and 22UX on 22nm — targeting intelligent edge, connectivity and so-called "Physical AI." On the same day, it revealed a strategic collaboration with fabless memory designer RAAAM Memory Technologies to develop and qualify GCRAM, a gain-cell memory IP, on GF's FDX platform, with a test chip already taped out.

The two announcements are related but distinct, and keeping them separate matters for understanding GF's strategy. The UX family is a process-platform play: PDK availability for customers building cost-sensitive edge devices. GCRAM, by contrast, is a memory-IP play — a partner's patented gain-cell RAM that GF will manufacture and help qualify on its 22nm FD-SOI process, aimed at the growing on-chip memory bottleneck in edge AI and automotive chips.

The UX family: two nodes, one edge-AI bet

The UX platform family is designed around where GF already has manufacturing scale. The 40UX platform builds on GF's high-volume 40nm process with embedded flash — a node that has shipped more than a million wafers — and is positioned for intelligent edge and connectivity products where cost and maturity matter more than bleeding-edge density. Volume production is slated for Singapore in 2027, with Malta, New York to follow. The 22UX platform is an analog-optimized 22nm process built on GF's proven 28nm base, also with more than a million wafers of heritage, and is manufactured in Dresden, Germany, with a roadmap toward multi-site production. Together the two platforms give customers a ladder: 40nm for low-cost high-volume, 22nm for higher analog/mixed-signal density.

"Our new UX platform family gives customers a scalable path to develop differentiated products for the intelligent edge," said Ed Kaste, senior vice president of CMOS at GF. The timing is deliberate: edge inference and Physical AI — machines that perceive and act in the physical world — are widely expected to be the next volume driver for mature-node fabs, and GF is positioning its differentiated, non-planar technologies against the scale economics of the leading-edge foundries.

GCRAM: attacking the on-chip memory bottleneck

The GCRAM collaboration addresses a different constraint. On-chip memory increasingly dominates die area and power in edge-AI chips, and conventional SRAM is the default but the most area-hungry option. GCRAM is a gain-cell RAM — a DRAM-like bitcell architecture that replaces the six-transistor SRAM cell with fewer transistors, trading a small amount of access complexity for significantly smaller area and lower leakage. RAAAM, an Israel-based fabless startup, has spent years commercializing the technology; its claims for GCRAM include up to 50 percent area reduction and up to three times lower power versus SRAM.

On GF's FDX platform, the two companies say the co-designed GCRAM bitcell achieves a 40 percent memory area reduction and up to 60 percent memory power reduction compared with commodity SRAM. "The successful tape-out of our GCRAM test vehicle on GF's FDX platform is a testament to the powerful synergy between RAAAM's memory innovations and GlobalFoundries' manufacturing excellence," said Robert Giterman, CEO of RAAAM. NXP Semiconductors is described as closely evaluating the solution, and lead-customer design access is targeted for early 2027.

The collaboration plugs into a broader GF memory portfolio strategy. GF already offers eMRAM in production, and in August 2025 announced 22FDX+ with RRAM — an oxide-RAM-based embedded non-volatile memory aimed at storing neural-network weights, with Nordic Semiconductor as a partner and volume production slated for 2026. Where RRAM and eMRAM attack non-volatile storage, GCRAM attacks the working memory that still dominates power and area in inference workloads — a complementary piece of the same puzzle.

Why FD-SOI keeps winning at the edge

Underlying both announcements is GF's long-standing bet on FD-SOI. The 22FDX platform's adaptive body-biasing lets designers tune transistor threshold voltage dynamically, enabling nominal operation at 0.4 volts and cutting power per MHz by roughly 30 percent versus 0.5V operation, while its ULL SRAM cells hold retention leakage down to around 0.35 picoamps per cell — roughly five times lower than a typical 12nm-class process. For battery-powered, always-on devices that must run small inference models continuously, that combination of ultra-low-voltage logic and low-leakage memory is exactly what the market needs.

GF is not alone in seeing it. STMicroelectronics is pushing 18nm FD-SOI with embedded PCM, Samsung has scaled 28FDS into 18FDS, and industry analyses project the FD-SOI process market growing from roughly $930 million in 2022 to over $4 billion by 2027. The edge-AI story is increasingly a memory story — and GF's bet is that by pairing differentiated FD-SOI processes with a portfolio of embedded memories — eMRAM, RRAM and now gain-cell RAM — it can own the parts of the AI wave that the leading-edge foundries are too busy to serve.