Skip to content
InfratoApp
InfratoApp / Blog / sxm-or-pcie-name-the-card-exactly

SXM or PCIe: name the card exactly on your price page

An H100 SXM and an H100 PCIe share a name and are sold at different prices. A price page that says only "H100" leaves the buyer to guess which one is on offer.

INFRATOAPP · PUBLISHED 17 OCTOBER 2026 · 5 MIN READ · 4 SOURCES · PRICES LAST CHECKED 18H AGO

One name, two purchases

Name the form factor and the memory on every rate: "H100 SXM 80GB" or "H100 PCIe 80GB", never "H100". Both carry 80 GB of memory, but the SXM module sits on a board with a fast link between GPUs and the PCIe card sits in an ordinary slot. Buyers price them separately because they are different purchases, and a page that drops the form factor makes its rate impossible to place.

The difference shows up in the market straight away. We keep a separate page for each part, with its own sellers and its own range, and a buyer who knows the difference checks which one you meant before reading your number.

The two H100 pages side by side

On the H100 SXM 80GB page, the lowest and highest listed on-demand hourly rates are $1.68 and $6.16, and the count of sellers is 13. On the H100 PCIe 80GB page, the same figures are $1.66, $3.50 and 7. These are two markets for two parts, and the ranges are meant to be read apart.

Now suppose a page offers "H100, 80GB" at one rate. Nobody reading it can tell which range it belongs in. If it is a PCIe card, the buyer comparing it with SXM rates thinks they have found a bargain and later finds the link between cards is not what their training job needed. If it is an SXM module, the buyer comparing it with PCIe rates thinks it is dear and moves on without asking.

Either way the provider loses something. The first case costs a customer after the sale, the second costs the sale itself.

The A100 shows what happens when the name is short

The A100 came with 40 GB and 80 GB of memory and in both module and card form, and some price pages still list it as "A100". We keep a page for exactly that, the A100 as sold with no form factor, where the count of sellers is 12, separate from the A100 SXM 80GB page, where it is 9.

The unlabelled page exists because we refuse to guess. A rate that says "A100" might be any of the four combinations, and putting it beside a named A100 SXM 80GB rate would compare things a buyer would not compare. So it sits in its own group, where it can be compared only with other rates that left out the same detail.

For the seller, that is a smaller audience. A buyer filtering for the 80 GB SXM part will not see a rate that does not say it is one.

What the card name on a price page should include

Four parts, in the order buyers search: the model, the form factor, the memory and, where it applies, the variant. "H100 SXM 80GB", "H100 PCIe 80GB", "A100 SXM 80GB", "L40S 48GB". The H100 NVL with its 94 GB is a third H100 product, and it needs its own name as well.

Use the name the card is sold under, not an internal instance code. An instance type called something like "gpu-large-8" tells a buyer nothing until they open a second page to decode it, and some will not. If you must use your own codes, print the card name next to them every time.

Keep the memory figure on the line, even where only one size exists. It costs a few characters and it answers the first question a buyer sizing a model will ask.

Why the memory figure belongs on the same line

Sizing a card is arithmetic that can be done before reading a rate. Model weights need about the parameter count times the bytes per parameter: 2 bytes in 16-bit, 1 in 8-bit, half a byte in 4-bit. A 45 billion parameter model in 16-bit needs about 90 GB for the weights alone, which does not fit on one 80 GB card, while the same model in 8-bit needs about 45 GB of weights and might.

Might, because the key-value cache and activations need memory on top, and that amount grows with context length and batch size. It cannot be given as one figure. Anyone working this out wants the memory size right beside the rate, so they can rule a card in or out without opening a spec sheet in a third tab.

This is also where the H200, with 141 GB, and the H100 NVL, with 94 GB, enter a buyer’s list. A provider that names its cards with their memory lets the buyer see at once which of its offers can hold the model, and which cannot, before a word is spent on price.

When a buyer should care which one it is

For a single-card job, such as serving a model that fits in 80 GB, the link between GPUs barely enters into it, and a PCIe card at a lower rate may be the sensible buy. For a job spread across several cards in one machine, the fast link on an SXM board is often what the buyer is paying for. The page should let the buyer make that choice, not make it for them by leaving the word out.

Providers sometimes worry that naming the PCIe card exactly will make it look second best. There is a case for the opposite. A buyer who wanted a single card for inference and finds "H100 PCIe 80GB" at a clear rate knows they have found what they came for, and does not have to wonder what the provider left out.

The same reasoning applies to anything an assistant says about your prices. A model asked for the cheapest H100 has no way to tell which card an unlabelled rate is for unless the page says. A precise name makes it harder for a provider to be quoted in the wrong range.

What we could not answer

  • ⚑This does not say which form factor suits a given workload in figures, because we hold no performance or interconnect measurements.
  • ⚑We cannot tell from a provider page which A100 an unlabelled rate refers to, so we cannot say how many of those sellers offer the 80 GB SXM part.
infratoapp, “SXM or PCIe: name the card exactly on your price page”, read 27 October 2026. lowest listed H100 PCIe 80GB rate per hour: $1.66. https://www.infratoapp.com/blog/sxm-or-pcie-name-the-card-exactly

Questions people ask about this

What is the difference between an H100 SXM and an H100 PCIe?

Both have 80 GB of memory. The SXM module sits on a board with a fast link between GPUs, while the PCIe card sits in an ordinary slot. They are sold as separate products at separate prices, so a rate for one should never be compared with a rate for the other.

Why are some A100 rates listed without a form factor?

Because the provider page did not say which A100 it was. The A100 was sold with 40 GB or 80 GB and as a module or a card, so a rate marked only "A100" cannot be placed beside a named variant. It is kept apart rather than guessed.

How should a GPU provider name a card on its pricing page?

With the model, the form factor and the memory, in the words buyers search for, such as "H100 SXM 80GB". Internal instance codes should always have the card name printed beside them, because buyers will not decode them and an assistant quoting the rate will not either.

These render as FAQPage structured data · answers are filled from the same live figures

Read next

How to describe interruptible GPU instances honestly

Before choosing an interruptible GPU a buyer needs the reclaim warning, what happens to the job and how the rate is set, not just the size of the discount.

Egress fees: the GPU cost buyers ask about before they sign

Buyers who have paid to move data out of a cloud ask about egress first. A GPU page that is silent on it is read as hiding a charge, whether or not one exists.

A checklist for a GPU pricing page buyers can compare

A GPU pricing page buyers can compare names the card exactly, quotes per GPU per hour, labels the commitment and states egress, regions and access mode.