Primemas Inc. showcased CXL-based memory technologies at the Future of Memory and Storage (FMS) 2026 conference, highlighting its collaboration with Micron Technology (NASDAQ: MU) on a U.S. Department of Energy-backed computing project designed to support large-scale artificial intelligence and high-performance computing workloads.
The Abaco Project, being developed for the Department of Energy’s Pacific Northwest National Laboratory (PNNL), is a rack-scale computing system expected to provide more than 100 terabytes of CXL-attached shared and pooled memory.
The project is supported by the DOE Office of Science’s Advanced Scientific Computing Research program and is intended to address growing memory requirements in scientific AI, HPC and other data-intensive applications.
AI systems increasingly require more memory as models become larger, datasets expand and inference workloads handle longer contexts. Conventional servers typically rely on memory installed directly within each system, which can leave capacity underused while requiring operators to over-provision infrastructure for peak workloads.
CXL, or Compute Express Link, provides a mechanism for connecting processors with memory and other resources using a high-speed interconnect. Memory pooling and expansion enabled by the technology can allow multiple compute systems to access shared resources, potentially improving memory utilization.
Under the Abaco architecture, CPU and GPU clusters will be able to access pooled Micron DDR5 memory at rack scale. Primemas said the configuration could enable memory pools exceeding 100TB when three to four Abaco chassis are deployed in a server rack.
Primemas is supplying CXL add-in cards and controller technology for the system.
The company’s PMA14 and PMA16 cards are PCIe-based CXL memory expansion products with 14 and 16 RDIMM slots, respectively. With 256GB RDIMMs, the cards can provide up to 3.5TB and 4TB of DRAM capacity.
Primemas said the cards can provide up to four times the memory capacity per CXL port compared with conventional monolithic solutions. The cards are scheduled to be shipped to Micron in September for installation in Abaco racks, followed by testing, qualification and benchmarking.
The company’s Hublet architecture provides the controller technology used to aggregate large amounts of DRAM. Primemas said the architecture is designed to provide the memory density required for PNNL workloads.
The Abaco system involves several technology providers. Micron is supplying DDR5 memory, while Liqid is contributing memory pooling and fabric management technology. The collaboration is aimed at allowing computing clusters to access large shared memory resources through an orchestrated infrastructure.
“PNNL’s scientific AI and HPC workloads have memory footprints that no single server can satisfy,” said Sumit Puri, co-founder and chief technology officer of Liqid.
Micron executive Luis Ancajas said demand for AI and data-intensive computing was increasing the need for architectures that can scale memory capacity across data centers.
The project comes as semiconductor and data center companies explore ways to overcome memory limitations that can constrain the performance of increasingly powerful processors. While additional compute capacity can accelerate AI workloads, processors can remain underutilized if sufficient memory capacity or bandwidth is unavailable.
Primemas also introduced SLiM, or Switchless Pooled Memory, at FMS 2026. The system is designed to address memory requirements associated with AI inference, particularly key-value (KV) caches used to retain information during model processing.
The 1U system is designed to allow multiple servers to access tens of terabytes of shared DRAM without external CXL switches. Primemas said the architecture is intended to provide low-latency access while reducing the hardware and space requirements associated with switched memory systems.
The company said it is also developing a roadmap for near-memory computing. Its plans include integrating CPU and FPGA acceleration capabilities into the Hublet architecture for AI and database applications.
Joseph Baco, vice president of business development at Primemas, said interest in CXL-based and pooled memory systems at FMS reflected growing demand for alternatives to fixed local-memory architectures.
The commercial prospects for pooled memory will depend on factors including system latency, software compatibility, workload characteristics and cost. But the technology is attracting attention as AI infrastructure operators seek to make memory capacity more flexible and improve utilization across increasingly large compute clusters.
The Abaco Project will provide a test of the approach at a national laboratory, while Primemas’ SLiM system targets commercial AI inference workloads. Both illustrate an industry effort to treat memory as a scalable data center resource rather than a fixed component of individual servers.
