en.Wedoany.com Reported - Liqid is providing a pooled memory layer for the Abaco system designed by Micron for the U.S. Department of Energy's Pacific Northwest National Laboratory (PNNL). The platform delivers over 160 TB of coherent memory across multiple servers, targeting AI and scientific workloads where data access efficiency is a greater bottleneck than processor performance.
Computing power is currently expanding faster than memory capacity, bandwidth, and locality can keep pace. Large graph models, scientific simulations, molecular workloads, and AI inference often stall because processors wait for data, or operators split tasks around the memory limits of a single server. The Abaco design treats memory as a shared, programmable resource rather than a fixed component permanently bound to a single machine. Liqid states that its rack-scale design can allocate up to 160 TB of DRAM to a single host or distribute capacity across up to 16 nodes. Micron serves as the project's primary technology contractor, while Liqid provides the expansion hardware, switching fabric, host adapters, and orchestration layer. This deployment remains at the national laboratory level and is not a signal that mainstream data centers are ready to rebuild around disaggregated memory, but it moves the discussion from architecture diagrams to hardware running in practice.
Liqid's EX 5410C platform supports CXL 2.0, offering up to 40 TB of DRAM per chassis, with multiple chassis forming a capacity pool exceeding 160 TB. Liqid Matrix software handles capacity allocation to hosts and integrates with tools such as Kubernetes, Slurm, and Ansible. The core proposition is that expensive memory should not be wasted sitting idle in the wrong server. Operators can assign a large capacity pool to simulations, analytics tasks, or inference workloads, then redirect that capacity to other tasks later. Ideally, this improves utilization and defers server purchases.
In practice, however, the economics depend on workload behavior. Memory pooling introduces fabric components, adapters, management software, and new operational dependencies, while local DRAM remains simpler. Latency-sensitive applications may not tolerate remote access patterns, even if the CXL fabric offers high bandwidth. Buyers need to base decisions on measurements from their own workloads, not just aggregate capacity figures. Fault domains also shift: a memory pool shared across multiple nodes increases flexibility, but a fabric or management failure can impact more systems simultaneously. Capacity planning moves from server procurement to orchestration strategy, requiring corresponding rebuilds of monitoring, access control, and recovery processes.
