Monday, March 16th 2026
Micron Announces Mass-Production of HBM4, SOCAMM2 Memory Modules and PCIe Gen 6 SSDs
At NVIDIA GTC 2026, Micron announced it is in high-volume production of three products at once, all timed around and designed for the Vera Rubin platform. The headline is HBM4 with the 36 GB 12-high stack started shipping in volume in Q1 2026, built for NVIDIA Vera Rubin. It hits over 11 Gb/s pin speeds for more than 2.8 TB/s of bandwidth, 2.3 times what HBM3E offered, while also improving power efficiency by over 20%. Micron has also shipped early samples of a 16-high 48 GB variant to customers, a 33% capacity bump per HBM placement over the 12H stack. Micron also announced that its 192 GB SOCAMM2 memory modules are now in high-volume production. Designed for Vera Rubin NVL72 systems and standalone Vera CPU platforms, it enables up to 2 TB of memory and 1.2 TB/s of bandwidth per CPU, with the broader SOCAMM2 portfolio spanning from 48 GB to 256 GB.
On the storage side, Micron 9650, the industry first PCIe Gen 6 data center SSD designed specifically for NVIDIA BlueField-4 STX architecture, is in mass production as we already reported here. It boasts sequential read speeds of up to 28 GB/s and can handle 5.5 million random read IOPS, essentially doubling the read performance of its Gen 5 predecessor. Furthermore, it offers a performance-per-watt ratio that is twice as efficient.
Source:
Micron
On the storage side, Micron 9650, the industry first PCIe Gen 6 data center SSD designed specifically for NVIDIA BlueField-4 STX architecture, is in mass production as we already reported here. It boasts sequential read speeds of up to 28 GB/s and can handle 5.5 million random read IOPS, essentially doubling the read performance of its Gen 5 predecessor. Furthermore, it offers a performance-per-watt ratio that is twice as efficient.
12 Comments on Micron Announces Mass-Production of HBM4, SOCAMM2 Memory Modules and PCIe Gen 6 SSDs
Laptops could use LPCAMM2, if the market could bring itself to use the standard.
So Apple in this case is basically trying to take the best and shorter way to get better performance.
In practice, the M4 used LPDDR5X @ either 7500 or 8533MT/s, same as what's found in many other devices with regular soldered modules.
M1 & M2 & M3 used LPDDR5 6400
The M5 uses LPDDR5X @ 9600, which is even going to be available as LPCAMM2, and also supported by Intel's PTL.
So there's no performance benefit whatsoever.
However, I would like to switch on optical fiber anyway to link RAM & CPU, I guess you already understood - why.