GPU & CXL memory pooling calculator
How much does sharing save?

It compares two ways of buying GPUs for the same workloads. Fixed: every team or service gets its own servers, sized for its own peak. Pooled: all workloads share one pool of GPUs, sized for the busiest hour of their combined demand. It reports GPUs needed each way, the GPUs a pool avoids, average utilisation, idle GPU-hours a day and GPU power, then names the smallest Liqid SmartStack or UltraStack that holds the pool.

A fixed estate has to cover each workload’s peak separately, so it needs the sum of the peaks. A shared pool only has to cover the busiest single hour of total demand: the peak of the sum. When peaks happen at different times, the second number is much smaller. When every workload peaks at the same time, the two are equal and pooling recovers nothing; the calculator says so.

For several host servers it picks the smallest validated SmartStack that holds both the pooled GPU count and the host count: 10 GPUs and 4 hosts, 20 and 8, 30 and 6, or 30 and 16, per Liqid’s SmartStack datasheet. For a single host it picks UltraStack 10, 20 or 30, or, with AMD MI350P selected, the UltraStack MI350P system at any count from 4 to 30 GPUs (Liqid’s datasheet; its web page describes growth in pairs). Because that system is sold complete, the results then use its GPU count. Beyond 30 GPUs or 16 hosts it flags a custom fabric.

For each host you enter the DRAM it already has, its peak memory need and when that peak happens. The fixed approach upgrades every host to its own peak; the pooled approach buys a shared CXL memory pool sized for the busiest hour of total shortfall. The calculator then picks Liqid’s 10 TB, 20 TB or 40 TB EX-5410C SmartStack by pool size and host count, measures the saving against the memory that stack actually holds, and flags anything beyond 40 TB or over 100 TB on one host.

We don’t reproduce vendor list prices, and a pound figure that showed only the GPUs saved, without the cost of the chassis, switches, adapters and software that make the pool, would mislead. The calculator gives the quantities that drive cost (GPUs, terabytes, hosts, rack units and kW), and we quote both options from those.

It works at whole-GPU granularity, so it does not model GPU partitioning (MIG or vGPU). It does not model network design, host CPU and RAM, or power beyond the GPUs themselves. It assumes GPUs can be moved between hosts as each workload’s window starts and ends, which suits services and scheduled jobs that release their GPUs better than one long job that holds them all day. Treat the result as a first sizing to refine with us, not a final design.

PCIe cards up to full-height, full-length, double-wide and 600 W, on servers Liqid has tested. Liqid’s chassis take PCIe add-in cards, so in our view SXM GPUs on NVLink baseboards, which are not add-in cards, cannot be pooled this way. Check your hosts and GPUs on our Liqid compatibility page.