Nutanix and GPUs: building an AI cluster
- Six NX platforms take GPUs today, across the G9 and G10 generations. The per-node ceiling: four L40S or H100 NVL in the NX-9151-G9
- The model code does not tell you: the NX-8150G-G10 takes three cards, the NX-8155-G9 in the same 2U1N chassis is not on Nutanix’s GPU list at all
- The L40S does not split with MIG at all; the RTX PRO 6000 Server Edition splits into four 24 GB instances
- MIG-backed vGPU is listed for vSphere, Red Hat KVM and Ubuntu KVM; AHV is not named
- MIG instances cannot talk peer-to-peer across GPUs and NCCL does not work with MIG, so distributed training will not assemble on slices
Which nodes take cards, and which do not
A hyperconverged node is a server where disks, compute and networking are fused into one block under AOS. Room for a full-size card is not always there: drive bays and fans eat it. So some NX models take up to four accelerators and others take none.
| NODE | FORM FACTOR | GPUs |
|---|---|---|
| NX-9151-G9 | 2U1N | up to 4× L40S or H100 NVL 94 GB |
| NX-3155-G9 | 2U1N | up to 2× L40S, or up to 4× L4 |
| NX-8150G-G10 | 2U1N | up to 3× L4, A16 or L40S |
| NX-8155AS-G10 | 2U1N | up to 3× L4 or A16 |
| NX-1175S-G10 | 1U1N | up to 2× L4 |
| NX-1150S-G9 | 2U1N | GPU option for the edge (L4) |
| NX-8170, NX-3060, NX-3035, NX-8155, NX-8155A, NX-8173AS, NX-1120AS | 1U1N / 2U1N / 2U2N / 2U4N | not on the GPU list |
Per Nutanix’s platform FAQ and the G10 launch material (G10 nodes ship since October 2025); confirm the spec sheet of the exact generation before ordering
Reading GPU capability from the model code does not work; the table shows it directly. There are two reliable checks: the GPU line in the model’s spec sheet, and the card list validated for that chassis. The NX-9151-G9 accepts the H100 NVL; the NX-3155-G9 with the same 2U1N form factor stops at two L40S; the H100 80 GB PCIe is qualified but not sold by Nutanix. Power and airflow set the limit, not case size.
What blocks a fifth card
The power spread across the line is large: an edge node draws a few hundred watts, while an NX-9151-G9 with four 350 W cards needs a power budget of several kVA and a matching cooling budget, a separate conversation with your data centre about the phase and row cooling.
The second reason: server cards are passive. The L40S and RTX PRO 6000 Server Edition have no fans of their own and live off front-to-back chassis airflow. NVIDIA’s forum has a telling case: an L40S in a Dell Precision workstation idling at 100 °C. Outside a server, this cannot be fixed.
Three ways to hand a card to a VM
| METHOD | HOW IT DIVIDES | ISOLATION & LICENCE |
|---|---|---|
| Passthrough | whole card to one VM | full; no vGPU licence needed |
| Time-sliced vGPU | time division, processes take turns | dedicated frame buffer per VM, shared compute, no fault isolation; licence mandatory |
| MIG-backed vGPU | hardware slices working in parallel | hardware isolation; NVIDIA’s sizing guide measured 20% higher performance than time-sliced at the 48 GB profile size |
Then comes the limit that breaks half of all multi-card node plans. The MIG documentation states it plainly: with MIG enabled there is no peer-to-peer between instances on different GPUs, over PCIe or NVLink, and NCCL is unsupported. NVIDIA’s forum position: distributed training over MIG can theoretically be pushed through GPUDirect RDMA over the network, but it will be slower than taking a bigger slice or disabling MIG. Partitioning a card and fast card-to-card exchange are mutually exclusive modes.
Which card splits how far
| CARD | MAX MIG INSTANCES | SMALLEST SLICE |
|---|---|---|
| RTX PRO 6000 Blackwell SE, 96 GB | 4 | 1g.24gb: 24 GB, 46 SM |
| RTX PRO 4500 Blackwell, 32 GB | 2 | 1g.16gb: 16 GB |
| H100 NVL, 94 GB (the H100 Nutanix sells) | 7 | 1g.12gb: 12 GB |
| H200 NVL, 141 GB | 7 | 1g.18gb: 18 GB |
| L40S, L4, A2, A16 | no MIG | passthrough or time-sliced vGPU only |
The last row deserves attention when picking a node: the L40S is the workhorse card for the NX-9151-G9 and NX-3155-G9, and it is exactly the one that does not slice. “How do we partition the L40S then?” is a question asked on NVIDIA’s forum by people who chose the card by TFLOPS and looked at the partitioning model afterwards.
What works on AHV
NVIDIA maintains a dedicated Nutanix AHV support page and classes the hypervisor as generic Linux with KVM. Ordinary time-sliced vGPU works. But in the consolidated feature table, MIG-backed vGPU is ticked only for vSphere, Red Hat KVM and Ubuntu KVM from release 19.0. Verify compatibility for your AOS version before ordering.
Two more things surface after deployment. Live migration of a vGPU VM requires the identical physical card on the receiving host and a driver from the same release branch that is the same or newer, and never works across different MIG profiles. And unlicensed behaviour: 20 minutes at full speed, then a 15 fps cap with degraded CUDA; after 24 hours 3 fps and no CUDA. C-series compute profiles come only with NVIDIA AI Enterprise (licensed per GPU), graphics vWS per concurrent user: two separate budget lines.
How many VMs per card
| vGPU TYPE | MEMORY PER VM | VMs PER CARD |
|---|---|---|
| DC-96C | 96 GB | 1 |
| DC-48C | 48 GB | 2 |
| DC-24C | 24 GB | 4 |
| DC-12C | 12 GB | 8 |
| DC-8C | 8 GB | 12 |
Compute profiles for the RTX PRO 6000 Blackwell Server Edition
For virtual desktops NVIDIA’s sizing guide gives two to four heavy users per RTX PRO 6000 (DC-12Q to DC-24Q profiles) and six to eight light users per RTX PRO 4500 Server Edition (DC-4Q). And the budget-saving caveat: the best-effort scheduler often allows 2–3× higher user density. The monitoring rule is simple: VM video-memory fill should rarely cross 90% and average below 70%.
| WHAT WAS MEASURED | NO PARTITIONING | WITH MIG |
|---|---|---|
| flan-t5-base on A100 40 GB (Red Hat) | ~4,200 tok/s | ~9,800 tok/s on 7 slices |
| P95 latency next to noisy neighbours (vLLM, community test) | ~2,499 ms | ~1,319 ms |
| Single small task on a 1/7 slice (NVIDIA forum microbenchmark) | 0.00533 s | 0.02640 s |
The last row is the price of partitioning: an individual slice is roughly five times slower than the whole card. Aggregate throughput still grows, because a small model does not saturate an A100, and scaling is sublinear.
Where to start and how to grow
The working starting configuration is three nodes: RF2 replication runs on three, and the cluster survives losing one node without stopping VMs. Cards are not needed in every node; the pattern that most often works: three nodes for general workloads, one or two of them with accelerators for inference.
Keep the GPU pool homogeneous, and the reason is live migration. A second node with the same card gives headroom and a place for VMs to move during maintenance. A mixed cluster with an L40S in one node and an RTX PRO 6000 in another works too, but every planned maintenance turns into an evening of service stops.
What we supply
Eurokommerz supplies the Nutanix NX line together with the GPUs Nutanix lists for it: L40S, L4, A16 and H100 NVL, and the RTX PRO 6000 Server Edition on the newest nodes, with EU invoicing and manufacturer warranty. Deployment and virtualisation engineering comes from our partner Vixen.UNO under the same contract.
FAQ
Which Nutanix NX nodes support GPUs?
How many GPUs does an NX-3155-G9 hold?
Does NX-8155-G9 support H100?
The L40S does not support MIG. How do we divide it?
Do we need AI Enterprise to hand a MIG instance to a VM?
The card is capped at 450 W instead of 600. Is it defective?
What do two cards in a node give over one?
Can we mix L40S and RTX PRO 6000 nodes in one cluster?
Tell us which models you run and how many people work with them. An engineer will pick the node and card count, price the licences, and say where the configuration hits the rack’s power limit. We reply within one business day.
Talk to an expertWe reply within one business day