From scientific computing to pooled-memory infrastructure for data-intensive computing and LLM inference.
At SC26, Micron is showing the next leap in HPC and AI performance is coming from memory, not just compute. Processors need enough memory capacity and bandwidth to stay fully utilized. Abaco places a high-capacity, shared memory pool in the rack, close to compute. It is designed for AI factories, neoclouds, cloud data centers and HPC clusters.
What Abaco is and why it matters
Abaco is a rack-scale reference design that pools DRAM within a rack, enabling multiple servers to access large amounts of memory. Rather than limiting each server to its own locally installed memory, Abaco makes a shared memory pool available across servers, so memory can be composed, shared and sized independently of compute.1
Compute is most useful when memory can keep it supplied with data. In one representative graph analytics run, a job that took about 7 hours 12 minutes finished in about 12.6 minutes once local memory and storage constraints were lifted. The pool keeps large working datasets close to compute resources instead of repeatedly moving to and from storage. Workloads avoid storage round trips, sharding and repeated work so processors stay busy doing useful work.2
From a US Department of Energy (DOE) research project to a commercial reference design
Abaco grows out of a problem Pacific Northwest National Laboratory (PNNL) researchers kept running into with graph, analytics and AI datasets too large for any single server’s memory, forcing jobs to spill to storage or split across servers while processors waited on data. PNNL needed many servers to share a large pool of memory at memory speed. Solving it took a whole-system approach. The Abaco team started from an understanding of the compute workload, then worked through the system architecture, the software applications and the underlying file and storage system layers, so that each layer informed the others.3
That end-to-end view is what made an efficient and practical solution possible. It is also an approach with deep roots in DOE projects such as Frontier at Oak Ridge National Laboratory, Aurora at Argonne National Laboratory, and El Capitan at Lawrence Livermore National Laboratory, where memory, interconnects, and software are designed together with the processors to turn compute into scientific results. Abaco applies that same principle at rack scale.4
Abaco’s direct lineage is more specific and starts with Crete (shown below), which Micron designed and built for PNNL and delivered as a rack-scale prototype system in July 2025, and placed in operations to be shared and used across multiple US Government labs.5
Crete took on a specific problem: how to make a large pool of DRAM, decoupled from any single server, available to multiple servers. The Crete prototype established the Compute Express Link (CXL) pooled-memory foundation. Abaco moves that work toward a production-oriented reference design that customers can evaluate, qualify and deploy at rack scale.
From the PNNL Abaco Rack to the Abaco reference design, Micron designed and assembled the Abaco Rack as a 120TB shared-memory infrastructure for PNNL scientific research, combining CPU and GPU compute with rack-scale pooled DRAM. That system became the foundation for the Abaco reference design: a modular, pooled memory architecture that can be configured for different capacities and server platforms.
The reference implementation brings together Supermicro® servers for compute, a Liqid® pooled-memory chassis and management software, Primemas CXL add-in cards that aggregate dense DRAM capacity, and Micron® DDR5 RDIMM modules as the underlying memory technology. Together, these components turn the original solution into a production-ready reference design for AI data centers.
Why HPC and AI need a new memory tier
Many leading HPC and AI workloads operate on datasets that are too large to fit efficiently in local server memory. Their access patterns can also be irregular, meaning hard to predict or cache (such as graph traversals that jump across billions of nodes), and data is often reused repeatedly (such as hot database records or the context an AI model re-reads on every request). If the solution supports it (and some don’t), when the active portion, or “working set,” cannot all fit in memory, it spills to storage, data must be moved, and the resulting I/O latency slows the job.
Spreading a dataset across discrete servers brings its own performance penalties, such as cross-server data movement to stitch partitions back together, duplicated copies of shared data, and idle DRAM stranded in servers that cannot lend it.
Abaco fills the gap between local DRAM and storage with an expandable, memory-speed tier that can be pooled and shared to servers. Micron’s Abaco team evaluated the architecture on graph analytics, production-style financial time-series analytics, and large language model (LLM) inference using key-value (KV) cache. The workloads differ, but the underlying problem is the same: active data must stay close to compute.
Real results across data-intensive workloads6
- Graph analytics: In testing with Pometry Raphtory and Graph 500-style analytics, a representative run dropped from 7 hours 12 minutes to 12.6 minutes once local memory and storage constraints were lifted, roughly 34× faster. When four servers shared the same memory pool, each finished the same job in about 12 to 13 minutes, so the pool served four servers without slowing any of them.
- Financial time-series analytics: A production-style tick-analytics study on financial market data showed up to 11.5× the performance of a storage-constrained baseline on a read- and compute-heavy workload, which translated to a 7.8× performance improvement per dollar.
- KV cache LLM inferencing (local CXL attached): Running Llama 3.1 70B with 4K- to 16K-token prompts across two GPU servers in a rack, locally CXL-attached pooled DRAM delivered approximately 99% throughput parity vs. local DRAM alone. When the servers shared retained KV blocks at roughly 89% reuse, output throughput rose 7× versus running without shared KV blocks (from about 30 to 213 tokens per second), and time to first token dropped from roughly 311 seconds to 46 seconds.
What comes next
We invite AI factory operators, neoclouds, cloud data center teams and HPC centers to bring their own models, datasets and applications and test them on an Abaco system, so we can see together where pooled memory delivers the most benefit.
To support that evaluation, we will publish a series of Abaco blogs and customer resources over the next few months:
- Abaco reference design release to customers: Architecture and deployment guidance to support customer evaluation.
- Repository of benchmarks: A consolidated home for benchmark configurations, methods and results.
- KV cache inference benchmark details: CXL sharing across multiple servers in a rack, including setup, reuse rates and performance.
- famfs in Linux: seamless memory sharing and pooling for apps: famfs, the fabric-attached memory file system from Micron and now accepted into the upstream Linux kernel, gives applications direct access to shared, pooled memory.
To test your own workload on Abaco, contact this blog’s author, Luis Ancajas.
References
- See “What Is Abaco” (liqid.com) to learn more.
- All comparative statements and data therein are based on internal Micron and Liqid collaborative testing.
- See “Liqid Powers PNNL’s Abaco System with CXL Memory Pooling for AI Science” (hpcwire.com) to learn more.
- Learn more about these and other projects at “ORNL’s methodical leap into the exascale era” (ornl.gov), “Introducing El Capitan” (str.llnl.gov) to learn more about these projects.
- The Crete specifications are available at “Crete Technical Specifications” (pnnl.gov).
- Benchmark results shown were performed in a Micron lab with the following configuration details: Server configuration: 2-socket x86 CPU, 16x128GB Micron® DDR5 RDIMMs (2TB), 7x2TB + 1x20TB PCIe Gen5 SSDs; CXL memory chassis: CXL 2.0 switch, 10x1TB CXL 2.0 AIC modules (10TB); OS: Red Hat Linux 9.1 with famfs kernel update. Other configurations may produce different results.