Nvidia has allegedly delayed the introduction of its upcoming SOCAMM technology with an upcoming Blackwell enterprise GPU. ZDNet reports that SOCAMM is now scheduled to arrive with next-generation Nvidia GPUs codenamed "Rubin".
SOCAMM was originally supposed to debut with the GB300, an upcoming Blackwell Ultra product aimed at workstations rather than servers. GB300 is a smaller counterpart to the GB200 that squeezes a Blackwell datacenter GPU and Grace CPU into a motherboard package suitable for OEM desktop/workation duties.
Allegedly, part of the reason Nvidia delayed SOCAMM was a change in GB300's motherboard design. Initially, GB300 was going to take advantage of a board design codenamed "Cordelia" but was transitioned to an existing design codenamed "Bianca. "Cordelia allegedly took advantage of SOCAMM memory in addition to embedding two Grace CPUs and four Blackwell GPUs on the board. "Bianca" only features a single Grace CPU and two Blackwell GPUs and lacks SOCAMM support (instead using existing LPDDR memory).
Cordelia's reliability is reported to be the cause of the switch; the newer board design is allegedly unreliable and suffers from data loss. SoCAMM reliability is also problematic and suffers from thermal issues, causing reliability issues.
Nvidia is also suffering from supply chain issues, which are allegedly contributing to SOCAMM's delays as well. The trillion-dollar giant is (unsurprisingly) having difficulty managing yields as it tries to get its supply chain up and running for the upcoming GB300. Switching to existing technologies (including the older board design with traditional LPDDR memory) will help Nvidia with its supply chain issues.
SOCAMM is an emerging memory form factor that Nvidia helped create in partnership with SK Hynix and Micron. The new form factor takes inspiration from CAMM2 (it is essentially a datacenter version of CAMM2), featuring significantly improved memory performance and storage capacity per square mm compared to traditional form factors (such as mainstream DDR5 DIMMs and RDIMMs). A single SOCAMM "stick" measures 14x90mm, boasting four 16-die LPDDR5 memory stacks giving the module a whopping 128GB of capacity and 7.5 Gbps of memory bandwidth.
SOCAMM is now slated to arrive with Rubin, Nvidia's next-generation GPU datacenter architecture that will succeed Blackwell. Little is known about Rubin/Rubin Ultra, but it is projected to support 12 stacks of HBM4E in 2027 (with 13TB/s of bandwidth), taking advantage of 5.5-reticle-size CoWoS interposers and 100mm x 100mm substrates made by TSMC. Rubin will also be drop-in compatible with the existing Blackwell NVL72 infrastructure.
Follow Tom's Hardware on Google News to get our up-to-date news, analysis, and reviews in your feeds. Make sure to click the Follow button.
