UK’s trusted IT infrastructure partner since 2003
sales@servnetuk.com
0800 987 4111
Servnet
FinanceToolsConfiguratorGet in Touch
Composable GPU · PCIe Gen5

Liqid GPU pooling
SmartStack and UltraStack, specified

Two ways to stop buying GPUs server by server. UltraStack composes 10, 20 or 30 GPUs to a single host for jobs that outgrow any one chassis. SmartStack shares a pool of up to 30 GPUs across as many as 16 servers, so GPUs follow the work. Both run on LIQID Matrix software, and SmartStack is built on the 600 W-ready Gen5 EX-5410P chassis.
GPUS TO ONE HOST (ULTRASTACK 30)
4–16[src]
HOSTS PER SMARTSTACK
600 W[src]
PER GPU SLOT
512 GB/s[src]
EXTERNAL BANDWIDTH PER CHASSIS
5U–15U[src]
SMARTSTACK FOOTPRINT
36 / 60[src]
MONTH WARRANTY OPTIONS
Two models

Scale up to one server, or share across many

Both families put GPUs in Liqid 10-slot expansion chassis on a PCIe fabric. The difference is where the GPUs go. UltraStack aims all of them at one operating system: 30 GPUs on one host, with GPU-to-GPU traffic kept inside the fabric[src]. SmartStack keeps a pool that several hosts draw from and give back to.
Liqid SmartStack 20: eight hosts sharing twenty GPUsComposed GPU counts are an example. Bundle composition from Liqid's datasheet.LIQID Matrix softwareMoves GPUs between hosts by UI, CLI or REST APImanagement networkHOST SERVERSHost 1 · 6 GPUsLLM inferenceHost 2 · 4 GPUsRenderingHost 3 · 2 GPUsVDIHost 4 · 2 GPUsData science+ 4 more hosts, 0 GPUs each right nowGen5 ToR switch18 ports · Gen5 x16Gen5 x16 per hostEX-5410P #110 of 10 composedEX-5410P #24 composed, 6 freeComposed GPU counts are an example. Bundle composition from Liqid's datasheet.
SmartStack 20 as Liqid specifies it: two EX-5410P chassis (20 GPUs), one 18-port Gen5 x16 top-of-rack switch and up to eight host servers, each with a Liqid Gen5 HBA. Here 14 of the 20 GPUs are composed and 6 sit free in the pool. Source: EX-5410P SmartStack datasheet.
SmartStack

SmartStack specifications

SmartStack is Liqid's validated, shared-pool bundle, sold in four sizes[src]. Every size includes a 1U Liqid Director management appliance and LIQID Matrix. Each host server needs one Liqid Gen5 HBA, bought per host.
SmartStack 10SmartStack 20SmartStack 30SmartStack 30+
Maximum GPUs10203030
Maximum host servers48616
EX-5410P chassis1233
Gen5 fabric switchNone1 × 18-port Gen5 x161 × 18-port Gen5 x162 × 18-port Gen5 x16
Rack units5U10U14U15U
GPU form factorFHFL 10.5″ dual-slotFHFL 10.5″ dual-slotFHFL 10.5″ dual-slotFHFL 10.5″ dual-slot
Management1 × Liqid Director 1U1 × Liqid Director 1U1 × Liqid Director 1U1 × Liqid Director 1U
Warranty & support36 or 60 months36 or 60 months36 or 60 months36 or 60 months

Optional add-on on every size: Liqid AX-5500 IO Accelerator, up to 64 TB NVMe on a PCIe Gen5 x16 FHFL card. Source: EX-5410P SmartStack datasheet, dated 10 Dec 2025. Liqid's GPU web page quotes 20 hosts for its largest SmartStack; we publish the datasheet's 16.

SmartStack sizes compared by GPUs, hosts and rack unitsRack units cover the Liqid hardware only; host servers are additional.Max GPUsMax hostsRack unitsSmartStack 101 chassis · no switches1045USmartStack 202 chassis · 1 switch20810USmartStack 303 chassis · 1 switch30614USmartStack 30+3 chassis · 2 switches301615URack units cover the Liqid hardware only; host servers are additional.
The SmartStack ladder: GPUs, host servers and rack units per size. SmartStack 30 trades hosts for density (6 hosts, 14U); SmartStack 30+ adds a second switch to reach 16 hosts in 15U. Source: EX-5410P SmartStack datasheet.
UltraStack

UltraStack: up to 30 GPUs in one operating system

UltraStack composes every GPU in the bundle to one host. Liqid's GPU page lists three sizes[src], and the AMD Instinct MI350P system is a fixed UltraStack configuration of its own.
UltraStack 10UltraStack 20UltraStack 30
Maximum GPUs102030
Host servers111
GPUs per hostUp to 10Up to 20Up to 30
Pooling softwareLIQID MatrixLIQID MatrixLIQID Matrix

Liqid's earlier UltraStack L40S reference design shows what a full build contains[src]. It runs on a Dell PowerEdge R760 (dual Xeon Gold) or R7625 (dual AMD EPYC) host. The 10-, 20- and 30-GPU versions take 7U, 12U and 16U with NVIDIA L40S 48 GB GPUs, and Liqid gives average power of 4,675 W, 8,832 W and 11,232 W. That design used the PCIe Gen4 EX-4410 chassis; Liqid's current SmartStack bundles use the Gen5 EX-5410P, and we confirm the chassis for any UltraStack build on quote. See the PowerEdge R760 and R7625 data sheets for the hosts.

The case for sharing

What a shared pool recovers

The more uneven the demand across your GPU servers, the more a shared pool gives back. A worked example with six three-GPU hosts; run your own in the GPU pooling calculator.
Six fixed three-GPU servers versus a shared SmartStack poolExample demand profile; we model yours from GPU counts and usage per team.Fixed: GPUs locked inside each serverComposable: one shared poolBar = installed · fill = in use · red outline = demand it cannot meetSame 18 GPUs, assigned where the demand isLLM servingneeds 8 GPUsSpare hostneeds 0 GPUsVDI poolneeds 2 GPUsCI runnersneeds 1 GPUsRender nodeneeds 3 GPUsData scienceneeds 1 GPUsLLM 8 · Spare 0 · VDI 2 · CI 1 · Render 3 · Data 156%In use, fixed83%In use, pooled5 → 0Unmet GPUsExample demand profile; we model yours from GPU counts and usage per team.
Illustrative arithmetic, not a Liqid benchmark. Six hosts with three GPUs each (18 GPUs) face 15 GPUs of demand. Fixed, only 10 do work (56%) and 5 GPUs of demand cannot be met, because the LLM service needs eight. Pooled through a SmartStack, the same 18 GPUs meet all 15 (83%) and three stay free for the next job.

Compose

Assign GPUs to a bare-metal host from Matrix's UI, CLI or REST API. Liqid quotes provisioning “in seconds”[src].

Scale

Add GPUs to a running host as a job grows; Liqid says allocations change with no reboots required[src].

Return

Release GPUs to the pool when the job ends, and another host can take them, zero-touch[src].
Dell PowerEdge MX

SmartStack MX: GPUs for MX7000 blades

Blade sleds have little room for GPUs. SmartStack MX connects Liqid GPU pools to the Dell PowerEdge MX7000 through a Fabric B module designed by Amulet Hotkey[src]. It composes up to 20 GPUs to a single MX760c sled and up to 30 per MX enclosure, and the same GPUs can also be composed to Dell R-series and C-series rack servers.
MX10MX20MX30MX30+
Maximum GPUs per MX700010203030
Maximum compute sleds48816
PCIe chassis1 × EX-44102 × EX-44103 × EX-44103 × EX-4410
PCIe switchIntegrated1 × 48-port1 × 48-port2 × 48-port
Host adapterCoreModuleXL, 4 × Gen4 x16CoreModuleXL, 8 × Gen4 x16CoreModuleXL, 8 × Gen4 x16CoreModuleXL, 16 × Gen4 x16
Rack units5U10U14U15U

Compatible sleds per the datasheet: MX740c, MX750c, MX760c (the Aug 2026 HCL lists MX750c and MX760c). The datasheet describes MX30 as both 6 hosts and 8 sleds and says up to 8 hosts connect to one B fabric, so we confirm sled counts on quote. SmartStack MX uses the PCIe Gen4 EX-4410 chassis. Source: SmartStack MX datasheet, dated 12 Sep 2024.

Supply & quote

Size a SmartStack or UltraStack for your GPUs

  • Tell us GPU counts per team or workload, and how busy they are across the day.
  • We compare UltraStack and SmartStack sizes against that profile, including the next size up.
  • We check every host server and GPU model against Liqid’s compatibility list.
  • You get one quote for the Liqid bundle, GPUs, HBAs and any new host servers.

Get a Liqid quote

Tell us what you want to pool and which servers you run. We size it against Liqid's hardware compatibility list and quote the complete build.

Business enquiries only — please use your company email. Gmail, Hotmail and other personal addresses are not accepted.

We only ship hardware to the countries in the dropdown. Software and maintenance services are global.

We use these details to reply to your enquiry and prepare any quote. No marketing emails without your consent. Privacy Policy

No obligation · response within 4 business hours

FAQs

Liqid GPU pooling: common questions

What is Liqid GPU pooling?

GPUs are installed in Liqid EX-5410P expansion chassis rather than inside servers. Each server gets a Liqid Gen5 host bus adapter, a PCIe Gen5 fabric connects servers and chassis, and LIQID Matrix software assigns GPUs to servers on demand. A GPU no longer needed on one server returns to the pool and can be composed to another without anyone touching the hardware.

SmartStack or UltraStack: which do I need?

Choose UltraStack when one workload needs more GPUs than a server can hold: it composes 10, 20 or 30 GPUs to a single host. Choose SmartStack when several servers should share a pool: SmartStack 10 serves up to 4 hosts, SmartStack 20 up to 8, SmartStack 30 up to 6 and SmartStack 30+ up to 16, holding 10, 20, 30 and 30 GPUs respectively. If unsure, tell us how many GPUs each team uses and when, and we’ll model both.

Can GPUs be moved without rebooting the server?

Liqid lists GPU hot-plug and hot-unplug as a SmartStack feature and says GPU allocations can be adjusted with no reboots required. Its Gen5 HBA supports surprise hot-plug on its downstream ports under Windows and Linux. Your application still has to release a GPU before it is removed, so plan reallocation around job boundaries.

What GPUs fit in the EX-5410P chassis?

Ten full-height, full-length, double-wide PCIe Gen5 x16 cards of up to 600 W each. Liqid’s compatibility list includes NVIDIA H200 NVL, L40S and RTX PRO 6000, AMD Instinct MI210 and Intel Flex GPUs among many others, with conditions on some entries. The chassis also takes NVMe add-in cards, FPGAs and DPUs. See Liqid compatibility.

How much rack space does a SmartStack use?

From 5U (SmartStack 10: one EX-5410P, no switch) to 15U (SmartStack 30+: three EX-5410P chassis and two Gen5 top-of-rack switches), per Liqid’s datasheet, including the 1U Liqid Director management appliance. The host servers are additional.

Is there a version for Dell PowerEdge MX blades?

Yes. SmartStack MX connects Liqid GPU pools to the Dell PowerEdge MX7000 through a Fabric B module, composing up to 20 GPUs to a single MX compute sled and up to 30 per MX enclosure. The SmartStack MX datasheet (September 2024) lists the MX740c, MX750c and MX760c sleds as compatible; Liqid’s August 2026 compatibility list names the MX750c and MX760c.

What warranty comes with a SmartStack?

Liqid’s SmartStack datasheet lists 36- or 60-month software and hardware warranty and support. We quote the term you want and confirm the support path on the quote.

The Liqid range

More Liqid products we supply and quote

Keep exploring

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.

  1. LIQID GPU Pooling Solutions · web page, read 29 Oct 2026
  2. EX-5410P SmartStack datasheet (GPU) · dated 10 Dec 2025
  3. EX-5410P PCIe Gen5 expansion chassis datasheet · dated 11 Dec 2025
  4. UltraStack L40S reference design (Dell PowerEdge) · dated 13 May 2024
  5. SmartStack MX datasheet (Dell PowerEdge MX) · dated 12 Sep 2024
  6. HX-5160P-1C PCIe Gen5 HBA datasheet · dated 27 Oct 2025
  7. Liqid Hardware Compatibility List · dated 17 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.

or call 0800 987 4111