The names SXM, PCIe and NVL describe three ways NVIDIA packages one GPU family: SXM is a module on a server board, PCIe is a card in a slot, and NVL is a PCIe card that takes NVLink bridges to its neighbors. The SXM vs PCIe difference shows up in memory bandwidth, power and how the GPUs in one VM are linked, and the names matter because an H100 or H200 offer can be any of them.
What SXM, PCIe and NVL mean
SXM is a module form factor. NVIDIA's Hopper whitepaper describes the H100 SXM5 as a custom board that houses the GPU and its HBM3 memory stacks and provides fourth-generation NVLink, available on HGX server boards with 4 or 8 GPUs. PCIe is the dual-slot card that goes into a standard server slot, with PCIe Gen5 at 128 GB/s on the H100 and H200. In NVIDIA's pages, NVL cards are PCIe cards that take NVLink bridges: the H100 page calls the H100 NVL "the PCIe-based NVIDIA H100 NVL with NVLink bridge", and the H200 page says up to four H200 NVL cards connect by NVLink. NVL also appears in rack names such as GB300 NVL72, which is 72 GPUs in one NVLink domain and a different thing.
What NVLink is and how it compares with PCIe
NVLink is NVIDIA's direct GPU-to-GPU interconnect. NVIDIA's NVLink page lists 900 GB/s per GPU for the fourth generation on Hopper and 1,800 GB/s for the fifth generation on Blackwell. In NVLink vs PCIe terms, NVIDIA's H100 and H200 tables list NVLink at 900 GB/s and PCIe Gen5 at 128 GB/s, about 7 times the bandwidth (calculated). The gap matters when a model is split across GPUs, because tensor parallelism exchanges data between the GPUs at every layer, as the multi-GPU page explains.
SXM vs PCIe: memory, power and links side by side
| Variant | Memory | Memory bandwidth | Maximum power | GPU-to-GPU link |
|---|---|---|---|---|
| H100 SXM | 80 GB HBM3 | 3.35 TB/s | Up to 700 W | NVLink, 900 GB/s |
| H100 PCIe | 80 GB HBM2e | 2 TB/s | 350 W | PCIe Gen5 at 128 GB/s, plus an optional NVLink bridge for 2 GPUs at 600 GB/s |
| H100 NVL | 94 GB HBM3 | 3.9 TB/s | 350 to 400 W | PCIe Gen5 at 128 GB/s, plus an NVLink bridge for 2 GPUs at 600 GB/s |
| H200 SXM | 141 GB HBM3e | 4.8 TB/s | Up to 700 W | NVLink, 900 GB/s |
| H200 NVL | 141 GB HBM3e | 4.8 TB/s | Up to 600 W | PCIe Gen5 at 128 GB/s, plus a 2-way or 4-way NVLink bridge at 900 GB/s per GPU |
| A100 SXM | 80 GB HBM2e | 2,039 GB/s | 400 W standard | NVLink, 600 GB/s |
| A100 PCIe | 80 GB HBM2e | 1,935 GB/s | 300 W | PCIe Gen4 at 64 GB/s, plus an NVLink bridge for 2 GPUs at 600 GB/s |
Specifications from NVIDIA's H100, H200 and A100 pages, checked 12 November 2026. The H100 page lists only the SXM and NVL columns today, so the H100 PCIe row comes from NVIDIA's H100 PCIe product brief and the Hopper whitepaper, and the HBM types come from the whitepaper and the H100 NVL product brief. The H100 page quotes 188 GB for an NVL pair, which is 2 x 94 GB. The PCIe brief gives the bridge as 600 GB/s in its table and 900 GB/s bidirectional in its overview, and this page uses the table figure, which matches the whitepaper.
1, 2, 4 or 8 GPUs in one VM
The link matters more as the GPU count grows. This is what NVIDIA's documents support for each VM size:
| GPUs in the VM | SXM (H100 and H200) | PCIe and NVL cards |
|---|---|---|
| 1 | No GPU-to-GPU link is used | No GPU-to-GPU link is used |
| 2 | NVLink | NVLink if a bridge joins the pair, which H100 PCIe, H100 NVL, A100 PCIe and H200 NVL all take, otherwise PCIe |
| 4 | NVLink, on 4-GPU HGX boards that NVIDIA lists without NVSwitch | H200 NVL takes a 4-way bridge. H100 PCIe, H100 NVL and A100 PCIe bridge in pairs, so two pairs are joined by PCIe |
| 8 | NVLink through NVSwitch, 900 GB/s between any pair on Hopper | Bridged pairs, or groups of 4 for H200 NVL, joined by PCIe |
NVIDIA's HGX H100 and H200 datasheet lists NVSwitch on the 8-GPU boards and none on the 4-GPU boards, and the Hopper whitepaper says the 8-GPU configuration gives 900 GB/s between any pair of GPUs. The H100 NVL brief says an NVL card bridges to a single adjacent card, which is why those bridges come in pairs. Whether a PCIe VM has its bridges fitted is the provider's choice, so read the Interconnect column. For a model split across GPUs without NVLink, vLLM's parallelism guide suggests pipeline parallelism instead of tensor parallelism, and vLLM performance tuning covers --tensor-parallel-size.
Check the link with nvidia-smi
Two commands show what a running VM has. NVIDIA's nvidia-smi documentation says topo -m prints a matrix of connections between the GPUs, where NV followed by a number means a bonded set of that many NVLinks. PIX, PXB, PHB, NODE and SYS are paths over PCIe, from a single PCIe switch (PIX) to a hop between NUMA nodes (SYS). The nvlink --status option lists each GPU's NVLinks, with the bandwidth of the active ones and Inactive for a link that is present but not active.
nvidia-smi topo -m
nvidia-smi nvlink --status
If only some pairs show NV entries, pin a 2-GPU job to one of those pairs with CUDA_VISIBLE_DEVICES, as the CUDA_VISIBLE_DEVICES page explains. The multi-GPU training guide covers training on a node with 2 to 8 GPUs.
H100 and H200 offers with 2 or more GPUs
The Interconnect column in the live tables follows the card, not a measurement of each machine. SXM offers read NVLink (SXM), and PCIe and NVL offers read NVLink bridge when the provider lists one and PCIe when it does not.
| GPU | GPUs | Provider | Region | vCPU / RAM / Disk | Interconnect | Per GPU-hr | Per hour | Launch |
|---|---|---|---|---|---|---|---|---|
| H100 SXM5 80 GB | 8x | Massed Compute | Virginia, US | 126 / 976 GB / 2800 GB | NVLink (SXM) | $3.14 | $25.12 | Launch |
Which variant to pick
For 1 GPU, the H100 PCIe vs SXM choice is memory bandwidth, power and live price, 2 TB/s against 3.35 TB/s and 350 W against up to 700 W, and an H100 NVL adds memory, 94 GB against 80 GB. For 2 GPUs, H100 SXM vs NVL is NVLink on the board against a bridged pair, so look for NVLink bridge on PCIe and NVL rows to keep tensor-parallel traffic on NVLink. For 4 GPUs, choose SXM or an H200 NVL offer whose Interconnect column says NVLink bridge, and confirm with nvidia-smi topo -m that all four GPUs show NV entries. For 8 GPUs, choose SXM if all eight must share NVLink. The A100 follows the same pattern at 600 GB/s. The H100 page, H200 page and A100 page list the offers and what fits in each card's memory, and for more than 8 GPUs you can request a cluster quote.
Frequently asked questions
What is SXM?
SXM is NVIDIA's module form factor: the GPU and its memory sit on a board that plugs into a server baseboard, with NVLink between the GPUs. NVIDIA lists NVLink at 900 GB/s per GPU for the H100 SXM and H200 SXM.
What is NVLink?
NVLink is NVIDIA's direct GPU-to-GPU interconnect. NVIDIA lists 900 GB/s per GPU for the fourth generation on Hopper, against 128 GB/s for PCIe Gen5.
What is the difference between SXM and PCIe?
SXM is a module on a server board with NVLink, while PCIe is a card in a slot that reaches other GPUs over PCIe unless an NVLink bridge joins them. On the H100, NVIDIA lists 3.35 TB/s of memory bandwidth for SXM against 2 TB/s for PCIe.
What are the H100 NVL and H200 NVL?
NVL cards are PCIe cards built to be bridged. The H100 NVL has 94 GB of HBM3 and bridges in pairs at 600 GB/s, and the H200 NVL has 141 GB of HBM3e and takes a 2-way or 4-way bridge at 900 GB/s per GPU.
How do I check whether my VM has NVLink?
Run nvidia-smi topo -m. NV followed by a number means NVLink between that pair of GPUs, and PIX, PXB, PHB, NODE or SYS means the path goes over PCIe. nvidia-smi nvlink --status lists each GPU's NVLinks and their bandwidth.
Does the A100 PCIe have NVLink?
Only through a bridge. NVIDIA lists an NVLink bridge for 2 GPUs at 600 GB/s, over PCIe Gen4 at 64 GB/s, and NVLink at 600 GB/s on the A100 SXM.