Comparison · Same card, different fabric

# H100 SXM5 or H100 PCIe?

Both are dedicated, both are monthly, both open without an identity check. What separates them is memory, bandwidth and $4,592 a month — and which of those matters depends entirely on what you intend to run.

- $6,061/mo H100 SXM5, one card
- $1,469/mo H100 PCIe, one card
- 80 vs 80 GB memory per card
- 1.7× bandwidth gap

## The short answer

Which one, and when

- **Pick the H100 SXM5:** when throughput matters. Same 80 GB, but 68% more memory bandwidth — which is what sets generation speed — for $4,592/month more.
- **Pick the H100 PCIe:** when the H100 SXM5 offers you nothing you need. Same memory, comparable bandwidth, $4,592/month cheaper.
- **Neither, if:** your model does not fit on a single card of either. Sharding costs synchronisation at every layer — see the [interconnect guide](https://gpuserver.io/guides/nvlink-vs-pcie) before buying a multi-GPU node.

## Side by side

**H100 SXM5 compared with H100 PCIe**

|  | [H100 SXM5](https://gpuserver.io/gpu/h100-sxm5) | [H100 PCIe](https://gpuserver.io/gpu/h100-pcie) | Difference |
|---|---|---|---|
| Architecture | Hopper | Hopper | Same generation |
| Memory per card | 80 GB HBM3 | 80 GB HBM3 | Identical |
| Memory bandwidth Predicts generation speed | 3,350 GB/s | 2,000 GB/s | 1.68× in favour of the H100 SXM5 |
| Form factor | SXM on an HGX board | PCIe add-in card | NVLink against PCIe between cards |
| Node sizes | 4×, 8× | 1×, 2×, 4×, 8× | Up to 640 GB in one node |
| From, per month | $6,061 | $1,469 | $4,592/month apart |

## Price at every node size

**Monthly price of both cards at each available node size**

| Node | H100 SXM5 | H100 PCIe | Total VRAM |  |
|---|---|---|---|---|
| 1 × GPU | not offered | $1,469/mo | 80 GB | [H100 PCIe](https://gpuserver.io/configure?c=h100-pcie-x1) |
| 2 × GPU | not offered | $2,892/mo | 160 GB | [H100 PCIe](https://gpuserver.io/configure?c=h100-pcie-x2) |
| 4 × GPU | $6,061/mo | $5,564/mo | 320 GB / 320 GB | [H100 SXM5](https://gpuserver.io/configure?c=h100-sxm-x4) [H100 PCIe](https://gpuserver.io/configure?c=h100-pcie-x4) |
| 8 × GPU | $11,509/mo | $10,568/mo | 640 GB / 640 GB | [H100 SXM5](https://gpuserver.io/configure?c=h100-sxm-x8) [H100 PCIe](https://gpuserver.io/configure?c=h100-pcie-x8) |

## What each one holds on a single card

At 8k context, best precision that fits. This is the difference that decides a purchase — not the specification sheet.

### Only the H100 SXM5

Nothing. Every model the H100 SXM5 holds on one card, the H100 PCIe holds too — the difference between them is speed, not capability.

### Only the H100 PCIe

Nothing. The H100 SXM5 holds everything the H100 PCIe does.

**The same model on both: Llama 3.3 70B.** On the H100 SXM5, roughly 43 tokens/second at 4-bit (AWQ, GPTQ); on the H100 PCIe, roughly 26. Single-stream and deliberately conservative — with continuous batching the aggregate is several times higher on both. Treat the *ratio* as the useful number, not the absolute.

## Cost per GB and per TB/s

Two ratios that cut through the specification sheet. The first tells you what memory costs; the second what speed costs.

**Cost per gigabyte of VRAM and per terabyte per second of bandwidth**

| Card | $/GB of VRAM, per month | $/TB per second, per month | Better at |
|---|---|---|---|
| [H100 SXM5](https://gpuserver.io/gpu/h100-sxm5) | $75.76 | $1,809 | Neither |
| [H100 PCIe](https://gpuserver.io/gpu/h100-pcie) | $18.36 | $735 | Both ratios |

These ratios rank cards; they do not choose one. A card that is cheaper per gigabyte is worthless if it has too few gigabytes to hold your model at all — capacity is a threshold, not a slope. Use them to break a tie between two cards that both fit.

## H100 SXM5 against H100 PCIe, in short

Is the H100 SXM5 worth the extra money over the H100 PCIe?

It costs $4,592 more a month — 313% — for 68% more memory bandwidth. Both hold the same set of models on a single card, so the difference is speed rather than capability.

Which one is faster for LLM inference?

Generation speed is bounded by memory bandwidth, because each token requires re-reading the active weights. The H100 SXM5 has 3,350 GB/s against 2,000 GB/s, so the H100 SXM5 leads by roughly 68% on a model both of them hold.

Do both come as multi-GPU nodes?

H100 SXM5: 4×, 8×. H100 PCIe: 1×, 2×, 4×, 8×. SXM cards live on an HGX board of four or eight — never two, never ten — and are joined by NVLink rather than PCIe.

Can I rent either without an identity check?

Yes, both, on the same terms. No document, no selfie, no phone number and no company registration, at any level of spend. Payment is in cryptocurrency, each invoice gets its own address, and root access arrives under 5 minutes after the first confirmation.

## Other comparisons

- [H100 SXM5 vs A100 SXM4  The generational jump, on the same board Compare](https://gpuserver.io/compare/h100-sxm5-vs-a100-sxm4)
- [H200 SXM5 vs H100 SXM5  More memory, more bandwidth, same architecture Compare](https://gpuserver.io/compare/h200-vs-h100-sxm5)
- [H100 PCIe vs A100 PCIe  The two 80 GB PCIe cards Compare](https://gpuserver.io/compare/h100-pcie-vs-a100-80gb)

## Both are racked and waiting, in all 6 data centres.

Whichever you pick, it is monthly, paid in crypto, opened without an identity check and delivered in under 5 minutes.

[Open the configurator](https://gpuserver.io/configure) [Read the guides](https://gpuserver.io/guides)

---

Source: https://gpuserver.io/compare/h100-sxm5-vs-h100-pcie/. This file is generated from the same data as the website; if a figure here differs from a page, the page is authoritative and this file is stale — the canonical source is https://gpuserver.io/.