PrimeMars is targeting the memory bottleneck market in AI data centers with its CXL-based memory expansion solution, 'JBOM'. As demand for GPUs surges due to the proliferation of generative AI, the memory capacity and access structures required to run large-scale models are emerging as new infrastructure challenges.

PrimeMars announced on the 28th that it will begin mass production and supply of 100 to 100 TB-class JBOMs starting in the second half of this year. The company is pursuing supply cooperation with Micron and is jointly developing next-generation CXL memory solutions with Samsung Electronics.

JBOM is a solution that combines multiple high-capacity memory modules to enable a server to utilize them like a single large memory pool. It overcomes the DRAM capacity limitations of existing CPU servers and expands memory scalability to the tens to hundreds of TB levels through a CXL-based architecture.

Prime Mars plans to supply products ranging from 40TB to a maximum of 120TB this year, and expand to over 240TB next year. This is a strategy to mitigate AI infrastructure efficiency issues, which are difficult to resolve solely by adding GPUs, through memory expansion.

In AI servers, even with sufficient computing equipment, GPU utilization drops if data is not loaded in a timely manner. Particularly in large-scale language model inference, the memory capacity for processing model parameters and cache data is cited as a decisive factor in performance.

CXL is a next-generation interconnect technology that enables memory expansion and sharing by connecting CPUs, GPUs, and memory devices at high speeds. Major memory companies such as Samsung Electronics, SK Hynix, and Micron are also working to expand the CXL ecosystem in response to changes in AI server architecture.