SXM or PCIe for your workload

One card or several

Choose SXM when your job spreads across several GPUs that have to talk to each other often, such as a model split over a node for serving or a multi-GPU training run. Choose PCIe when each job fits on one card, because then the fast link between GPUs is something you pay for and do not use. SXM modules sit on a board with a fast GPU-to-GPU link; PCIe cards sit in an ordinary slot. The two are different purchases at different prices, even with the same chip name.

The H100 is the clearest example because it is sold both ways with 80 GB of memory. The model you can load is the same. What differs is how the card connects to its neighbours, and the server around it.

This site holds no measurements of either form factor, so it cannot say how much the link speeds up any job. It can show what each form costs from the listed rates, and the post sets out which workloads use the link at all.

Why the link matters for split models

When a model is too large for one card, a serving engine or training framework divides it. Tensor parallelism splits each layer, so the cards exchange data many times for every token. Pipeline parallelism gives each card a block of layers, which needs less traffic but leaves cards waiting on each other. Data-parallel training copies the model and syncs gradients after every step.

All of these move data between GPUs, and the path that data takes is what SXM is built for. On PCIe the same traffic goes through the host's slots and switches. This site holds no figures for how much slower that is for any workload, and the honest answer is that it depends on the model, the split and the framework.

A 13 billion parameter model in 16-bit, about 26 GB, split across 2 GPUs only to give the cache more room is a typical case to test on both form factors before committing.

The H100 both ways, priced

On the H100 SXM page, the cheapest hourly listing on record is $1.68 and the dearest $6.16 (sellers with a listed rate: 13). On the H100 PCIe page the same pair reads $1.66 and $3.50 (sellers with a listed rate: 7).

Do not assume PCIe is always the cheaper listing. Rates come from different sellers with different hardware, terms and regions, and it is possible to find an SXM card from one seller under a PCIe card from another. Read both pages, and check that the rates you are comparing are per GPU and on-demand.

Single-card work belongs on either

A 7 billion parameter model in 16-bit, a fine-tune with adapters on a quantised 13 billion parameter base, a batch embedding job: none of these leave one card, so none of them use the GPU-to-GPU link. For that kind of work the form factor matters less than the rate and the terms, and a PCIe card at a lower rate is a sensible first test.

Watch the instance shape, though. Some providers rent SXM cards only as full nodes of 8 GPUs. For a one-card job, the smallest unit you can rent can matter more than the form factor, so ask before comparing.

A full node is also a larger commitment in practice. Even at an hourly rate, 8 GPUs left running while you debug a single-card job means paying for 7 idle cards, so match the unit to the work before comparing rates.

Which cards come in which form

The A100 is also sold both ways, and some sellers list it with no form factor at all; those rates sit on the A100 page and need a question to the seller before they can be compared. The H200, the B200 and the B300 are listed here as SXM parts. The L4, the L40S and the RTX cards are PCIe cards.

That means the form factor question only arises for some purchases. For an L40S it is settled by the card; for an H100 or an A100 it is a choice, and the listing has to say which. A rate without a form factor is a rate you cannot place, so ask before treating it as the cheaper option.

It also means that a model too large for one PCIe card has two routes: several PCIe cards with the slower path between them, or a move to an SXM part with more memory, so the split is avoided entirely. Price both routes and test the one that wins on paper.

What to ask before renting

Ask how the GPUs in a multi-GPU instance are connected, in the provider's own words, and whether that link is the same on every machine of that type. Ask whether you can rent one card, or only a node. Ask whether multi-node jobs have a fast network between machines, which is a separate question from the link inside one.

Then run your split workload for an hour on each form factor and time it. The rental cost calculator turns the result into a monthly cost; what you are comparing is the hours each takes at each rate, not the rate alone.

When the split job runs nearly as well on PCIe, the cheaper form wins. When it runs much worse, the gap between the two rates is what the link costs, and the timing tells you whether it pays for itself.