A separate fabric for memory
EX-5410C chassis
FX-5015C switch
HX-5160C-1C HBA
Stranded DRAM, quantified
Liqid claims that dynamically consolidating memory can cut the servers, licences and networking needed by up to 75%, and talks of 100% utilisation[src]. Those are vendor figures. In practice you keep local DIMMs for latency-sensitive data and use the pool as a large second tier, so size the pool from the peaks that currently force you to over-buy every server. To size the conventional side, see our server memory guidance.
EX-5410C SmartStack specifications
| SmartStack 10TB (Small) | SmartStack 20TB (Medium) | SmartStack 40TB (Large) | |
|---|---|---|---|
| Memory modules | 10 (8 × 128 GB DDR5 each) | 20 (8 × 128 GB DDR5 each) | 40 (8 × 128 GB DDR5 each) |
| Maximum hosts | 5 | 8 | 12 |
| EX-5410C chassis | 1 | 2 | 4 |
| CXL fabric switch | None | 1 × CXL ToR, 15 × x16 Gen5 uplinks | 2 × CXL ToR, 15 × x16 Gen5 uplinks |
| Rack units | 5U | 10U | 19U |
| Warranty & support | 36 or 60 months | 36 or 60 months | 36 or 60 months |
Add-on: Liqid AX-5500 IO Accelerator, up to 64 TB NVMe on a PCIe Gen5 x16 card. Source: EX-5410C SmartStack datasheet, dated 10 Dec 2025. Liqid's memory web page also mentions sharing 100 TB across 32 servers; that exceeds the validated SmartStacks above, so we treat it as a custom fabric and size it on request.
Micron Abaco: a memory-first rack powered by Liqid
| Workload (Micron lab testing) | Result | Detail |
|---|---|---|
| Graph analytics | Up to 30× faster | Graph query set cut from about 8 hours to 15 minutes |
| KV cache LLM inference | Up to 7× tokens/s | With NVIDIA Dynamo tiering to the Liqid pool |
| In-memory analytical cache | 3.9× throughput | Single server; 10× is the target for the shared configuration |
| Agentic LLM inference | 5.5× tokens per dollar | Projected, not measured |
Vendor results from Micron lab testing as published on Liqid's Abaco page. The first deployment is at Pacific Northwest National Laboratory.
Planning for an EX-5410C
| Item | EX-5410C | FX-5015C switch |
|---|---|---|
| Form factor | 4U, 175 × 447 × 790 mm, 32.58 kg | 1U, 41.7 × 438 × 628.5 mm, 11.6 kg |
| Power | 4 × 1500 W / 3000 W Titanium, N+1 | 2 × 800 W (N+N); 300 W typical |
| Inputs | 100–120 V and 200–277 VAC; Saf-D-Grid 400 V connectors | 100–277 VAC; Saf-D-Grid 400 V |
| Operating temperature | 0–35 °C, 10–90% humidity | 0–35 °C, 10–90% humidity |
| Management | 1 × GE; OpenBMC out-of-band | 1 × GE; OpenBMC out-of-band |
Sources: EX-5410C and FX-5015C datasheets. The 790 mm chassis depth needs a deep rack; check yours before ordering.
Size a CXL memory pool
- Tell us which workloads run short of memory today, and their peak and typical footprint.
- We check your hosts are CXL-capable and on Liqid’s list, or propose ones that are.
- We size the pool against local DRAM and pick the SmartStack, or a custom fabric for larger estates.
- We quote the chassis, modules, switches, HBAs and any new hosts together.
Liqid CXL memory pooling: common questions
What is CXL memory pooling?
CXL (Compute Express Link) lets a server use memory attached over a PCIe-based link as if it were its own. Pooling puts that memory in a shared chassis so several servers draw allocations from one pool instead of each carrying its own surplus. Liqid’s implementation uses CXL 2.0 memory modules in an EX-5410C chassis, a CXL fabric switch and a CXL host bus adapter in each server, orchestrated by LIQID Matrix. For the protocol background, read what is CXL.
How much memory can one server get?
Liqid says one EX-5410C chassis holds 1 TB to 20 TB (ten modules of up to 2 TB each), and that adding chassis to one fabric lets you compose up to 100 TB of memory to a single server, using five chassis.
Which servers support Liqid CXL memory?
The host needs a CXL-capable platform with CXL/PCIe native mode enabled in BIOS. Liqid’s compatibility list (August 2026) names four CXL hosts: ASRock GNR2D16-2T, ALVRES KR2280-X3-A0-R0-00, Fsas Technologies PRIMERGY RX2530 M8 and Supermicro SYS-222C-TN11. We confirm any other platform with Liqid before quoting. See Liqid compatibility.
What workloads benefit most?
Liqid aims its CXL SmartStack at AI inference, retrieval-augmented generation, vector databases and graph databases, where larger memory means bigger context windows, faster embedding lookups and less sharding. Its memory page also lists in-memory databases such as SAP HANA and Redis, HPC and VDI. Confirm application support for CXL-attached memory with the software vendor first.
Is CXL memory as fast as local DIMMs?
No. It sits further from the CPU: Liqid quotes 120 ns latency for its CXL fabric switch alone, and its Abaco brief puts pooled-memory latency at roughly 430 to 500 ns. It is best used as a large second tier behind local DRAM, for data that is big rather than latency-critical. Micron and Liqid’s Abaco results show where that trade pays off.
Can memory be moved between servers while they run?
Liqid says memory can be scaled up or down to match demand with no reboots and no downtime, and that Matrix can power off slots and virtually replace failed DIMMs without disruption. How a given operating system and application handle memory being added or removed still needs testing in your environment.
What is Micron Abaco?
A memory-centric rack architecture designed by Micron, with Liqid supplying the memory fabric, chassis and orchestration. One Abaco rack presents up to 160 TB of DDR5 as a pool that up to 16 servers can claim, share or release without a reboot. The first deployment is at the US Department of Energy’s Pacific Northwest National Laboratory.
More Liqid products we supply and quote
Related Servnet guides, tools and hardware
Sources
Every specification on this page comes from Liqid's own product pages and datasheets, listed below with their dates. Where Liqid's web copy and a datasheet differ, we use the datasheet. Figures are confirmed against the current revision when we quote.
- LIQID Memory Pooling Solutions · web page, read 29 Oct 2026
- EX-5410C SmartStack datasheet (CXL memory) · dated 10 Dec 2025
- EX-5410C CXL 2.0 expansion chassis datasheet · dated 11 Dec 2025
- FX-5015C CXL 2.0 fabric switch datasheet · dated 11 Dec 2025
- HX-5160C-1C CXL 2.0 HBA datasheet · dated 27 Oct 2025
- Liqid Hardware Compatibility List · dated 17 Aug 2026
- Micron Abaco, powered by Liqid · web page, read 29 Oct 2026
- What Is Abaco? Liqid technical brief · Aug 2026
LIQID, LIQID Matrix, SmartStack and UltraStack are trademarks of Liqid Inc. Servnet is an independent UK IT reseller; this page is our own guide to the range.
Talk to a UK specialist
Get expert advice or a no-obligation quote — servers, storage, networking, maintenance, finance and cloud. We reply the same working day.