NVIDIA experiments with SOCAMM – new memory standard for compact AI supercomputers?

The fact that NVIDIA likes to establish its own hardware standards is nothing new. Now the company also appears to be taking on the memory market for compact AI supercomputers. According to South Korean reports, Team Green is working with Samsung, SK Hynix and Micron on a new memory module called SOCAMM – smaller, more efficient and of course only intended for special applications.

SOCAMM – because standard is not good enough?

Little is currently known about SOCAMM, but the initial details are certainly interesting. The modules are said to be more compact than previous solutions and still offer up to 694 I/O pins. By comparison, LPCAMM and classic PC DRAM modules lag far behind in this respect. A high I/O port connection means one thing above all: fewer bottlenecks between memory and processor – and this is precisely what is increasingly becoming a problem for AI applications. Another detail that makes you sit up and take notice: SOCAMM remains replaceable, which is rather unusual in this segment. Instead of permanently soldered memory, there is a module that can be replaced with a more powerful version if required – a feature that could be very useful in high-end notebooks or compact workstations.

Why all this?

The answer may lie in NVIDIA’s push towards personal AI computing. With projects such as DIGITS, the company apparently wants to establish powerful mini AI supercomputers for developers, companies and presumably also for end customers at some point. A customized memory standard naturally fits into this concept.

When will SOCAMM arrive?

The modules are currently still in the test phase, but the rumor mill assumes that mass production could start in 2025. If the standard becomes established – which is not unlikely given NVIDIA’s market power – this could also have consequences for other manufacturers. The question remains: will SOCAMM really be a step forward, or rather another attempt to further expand control over the AI hardware market? The answer will probably come when the first devices with this memory appear.

Source: Sedaily