A listing that says "H100 80GB" can mean two different boards, and "H100 94GB" can mean two others. They differ in memory type, bandwidth, power and how the GPUs talk to each other, and so does the price. This post sets out the variants from NVIDIA's own documents, then shows how to tell which one you have from a shell, with real output.
I run VoltageGPU, which rents H100s by the hour. The H100 outputs below come from an 8x H100 machine we rented on our own platform on 24 October 2026; the raw files are public and linked.
The variants
| H100 SXM | H100 NVL | H100 PCIe | |
|---|---|---|---|
| Memory | 80 GB | 94 GB | 80 GB HBM2e |
| Memory bandwidth | 3.35 TB/s | 3.9 TB/s | 2,000 GB/s |
| Max power | up to 700 W, configurable | 350 to 400 W, configurable | 350 W (310 W in its 300 W power-connector mode) |
| Form factor | SXM | PCIe, dual-slot, air-cooled | PCIe, dual-slot, passive cooling |
| GPU-to-GPU | NVLink, 900 GB/s | NVLink, 600 GB/s | NVLink bridges join two cards |
| MIG | up to 7 x 10 GB | up to 7 x 12 GB | 7 instances |
| PCI device ID | 0x2330 ("H100 SXM5 80GB") | 0x2321 ("H100L 94GB"), see below | 0x2331 |
Sources: NVIDIA's H100 page for the SXM and NVL columns (it no longer lists the PCIe card), the H100 PCIe product brief (PB-11133-001_v02, November 2022) for the PCIe column and its device ID, and the pci.ids database for the two other IDs, all read on 19 November 2026.
What the differences mean in practice:
- Bandwidth. The SXM reads memory about 1.7 times as fast as the PCIe card (3.35 / 2.0). For LLM inference with small batches, that ratio tends to matter more than any TFLOPS figure.
- GPU-to-GPU. On the 8-GPU SXM machine below, every GPU reached every other one through NVLink. PCIe cards talk over PCIe unless a bridge pairs them. For tensor-parallel work across many GPUs, this is the line that matters.
- Power. The SXM can draw up to 700 W, and NVIDIA calls that figure configurable: the host can set a lower cap, and a lower cap can mean lower sustained clocks.
"94 GB" is not one board
pci.ids lists two 94 GB H100s, and only one is the NVL:
curl -fsSLO https://raw.githubusercontent.com/pciutils/pciids/master/pci.ids
awk '/^10de/{f=1;next} /^[0-9a-f]{4} /{f=0} f' pci.ids | grep -E '^\s(2321|2330|2331|2339|22a3)\s'
Output on 19 November 2026 at 21:06 UTC (Windows 11, Git Bash 5.2.26, pci.ids version 2026.11.19):
22a3 GH100 [H100 NVSwitch]
2321 GH100 [H100L 94GB]
2330 GH100 [H100 SXM5 80GB]
2331 GH100 [H100 PCIe]
2339 GH100 [H100 SXM5 94GB]
NVIDIA's page gives the NVL 94 GB, but pci.ids also has 2339, a 94 GB SXM5 module. So a "94 GB" listing does not by itself mean an NVL: the device ID tells you which. I have not found an NVIDIA document that confirms 2321 for the NVL, so treat that pairing as the database's, not NVIDIA's. The awk step keeps only NVIDIA's block: a plain grep on the file also matches other vendors' devices with the same digits.
Check 1: name and device ID, with the driver loaded
nvidia-smi --query-gpu=name,pci.device_id,memory.total,power.limit --format=csv
I have no H100 on my desk, so here is the same query on the GPU I do have, to show the format (19 November 2026, 21:04 UTC, driver 610.88):
name, pci.device_id, memory.total [MiB], power.limit [W]
NVIDIA GeForce RTX 3060 Ti, 0x248910DE, 8192 MiB, 200.00 W
pci.device_id is the device ID followed by NVIDIA's vendor ID, 10DE. An H100 SXM 80 GB should print 0x233010DE. Note that the ID tells you more than the name: pci.ids lists 2489 as GA104 [GeForce RTX 3060 Ti Lite Hash Rate], a variant the name does not mention.
On the H100 machine, plain nvidia-smi printed this for each of its 8 GPUs (02-nvidia-smi.txt, 24 October 2026, driver 595.112.07; first GPU shown):
| 0 NVIDIA H100 80GB HBM3 On | 00000000:18:00.0 Off | 0 |
| N/A 22C P0 70W / 700W | 0MiB / 81559MiB | 0% Default |
"80GB HBM3" and a 700 W cap fit the SXM column. 81,559 MiB is how the driver reported the 80 GB on this card.
Check 2: before any driver is installed
On a fresh machine with no NVIDIA driver yet, the kernel already knows the IDs. This loop lists every NVIDIA PCI device with its class and device ID; we ran it on that machine before installing the driver (00-probe.txt):
for d in /sys/bus/pci/devices/*; do [ "$(cat "$d/vendor")" = 0x10de ] && echo "${d##*/} $(cat "$d/class") $(cat "$d/device")"; done; true
Two of its twelve lines:
0000:05:00.0 0x068000 0x22a3
0000:18:00.0 0x030200 0x2330
Class 0x0302 is a 3D controller, here an H100 SXM5 80 GB (0x2330). Class 0x0680 with 0x22a3 is an NVSwitch: the machine had four, and their presence alone tells you this is an SXM board with switched NVLink, not PCIe cards.
Check 3: how the GPUs are connected
nvidia-smi topo -m
On the same machine (04-nvidia-smi-topo.txt), every pair of the 8 GPUs read NV18; the legend in the output defines NV# as "Connection traversing a bonded set of # NVLinks". The first row:
GPU0 X NV18 NV18 NV18 NV18 NV18 NV18 NV18 0-23,96-119 0 N/A
Without NVLink, the same matrix shows PCIe path types instead, which the legend lists as SYS, NODE, PHB, PXB and PIX. If you paid for an 8x SXM node and see those between GPUs, ask your provider why.
What it costs, by form factor
RunPod's pricing page, read on 20 November 2026 at 07:49 UTC, prices all three, per GPU-hour on demand:
| Community Cloud | Secure Cloud | |
|---|---|---|
| H100 PCIe | $1.99 | $2.89 |
| H100 NVL (listed with 94 GB) | $2.59 | $3.19 |
| H100 SXM | $2.69 | $4.39 |
On Secure Cloud the SXM costs $1.50 more per hour than the PCIe card. The NVL sits between them, though it has more memory and bandwidth than the SXM; what it gives up is NVLink speed (600 against 900 GB/s) and the SXM's power headroom. Nebius lists "NVIDIA HGX H100", an SXM board, at $4.50 on demand (same reading).
What this proves, and what it doesn't
The name, the device ID and the topology tell you which model the driver and the kernel see. They do not tell you:
- That the card is genuine or healthy. These values are read from the board; a cryptographic proof of a genuine NVIDIA GPU is a separate step, NVIDIA's attestation, which we have run only on GPUs in confidential-computing or Protected PCIe mode. Health needs a load test: bandwidth, NVLink throughput, errors under heat.
-
The sustained clocks. The power cap can be lowered by the host; read
power.limiton your own machine, not on a spec sheet. - Anything about NVL or PCIe output. The H100 outputs above are from one SXM machine. I have no NVL or PCIe H100 output to show, so the IDs for those come from pci.ids and NVIDIA's brief, not from a machine.
The specs in the first table are NVIDIA's figures; I measured none of them.
Drafted with AI assistance, then checked: the two local commands were run on 19 November 2026 and their output is copied as printed; the H100 outputs are copied from the raw files of 24 October 2026; prices were read on the providers' pages on 20 November 2026.
Top comments (0)