Micron posts AI‑optimized memory signals
Micron highlighted new AI‑optimized memory (HBM4, PCIe Gen6 SSD) tailored for NVIDIA workloads and energy‑efficient scaling — a sign memory suppliers are refocusing on AI performance per watt for production inference shared. That ramps pressure on device teams to reconcile on‑device memory constraints with heavier ML models.
Micron began volume shipment of its HBM4 36 GB 12‑high stacks in the first quarter of calendar 2026, explicitly citing design alignment with NVIDIA’s Vera Rubin platform. (publicnow.com) The HBM4 parts are specified to run at over 11 Gb/s per pin and deliver aggregate bandwidth greater than 2.8 TB/s for an HBM stack. (electronicsweekly.com) Micron’s Micron 9650 is the industry’s first PCIe Gen6 data‑center SSD in mass production and advertises up to 28 GB/s sequential read and 5.5 million random read IOPS. (micron.com) The 9650 uses Micron’s G9 NAND and Micron’s product materials claim it can provide roughly 2× the read performance of Gen5 drives at substantially improved perf-per‑watt — Micron quantifies the SSD’s efficiency at up to 100% higher performance per watt versus Gen5. (micron.com) Micron confirmed 192 GB SOCAMM2 modules are now in high‑volume production while also sampling 256 GB SOCAMM2 units, with Micron stating a single 8‑channel CPU can now support about 2 TB of LPDDR5X using these modules. (guru3d.com) Micron timed these announcements to NVIDIA’s GTC 2026 (March 16, 2026) and used the event to rebut supply‑rumor chatter by emphasizing shipping status and ecosystem interoperability for Vera Rubin. (publicnow.com)
