We keep getting requests for an "H200 SXM5 card." That part does not exist as a card. The H200 SXM is a module that NVIDIA ships on HGX baseboards, and server makers build those boards into 4-GPU and 8-GPU systems. The H200 you can install in a PCIe slot is the H200 NVL. This post covers how the two differ, which servers carry each one, and how to word a request for quote so a dealer prices the right part.
NVIDIA's own data center line card (August 2025) lists exactly two H200 products: HGX H200 and H200 NVL. Lenovo's H200 product guide shows the same split in its ordering table: its three HGX H200 boards are "CTO only," configured into a server at order time, and the H200 NVL card has its own orderable part number.
The two parts side by side
| Specification | H200 SXM | H200 NVL |
|---|---|---|
| GPU memory | 141 GB HBM3e | 141 GB HBM3e |
| Memory bandwidth | 4.8 TB/s | 4.8 TB/s |
| FP8 Tensor Core (with sparsity) | 3,958 TFLOPS | 3,341 TFLOPS |
| BF16 Tensor Core (with sparsity) | 1,979 TFLOPS | 1,671 TFLOPS |
| Max TDP | Up to 700 W, configurable | Up to 600 W, configurable |
| Multi-Instance GPU | Up to 7 MIGs at 18 GB each | Up to 7 MIGs at 16.5 GB each |
| Form factor | SXM module on an HGX baseboard | PCIe, dual-slot, air-cooled |
| GPU-to-GPU link | NVLink, 900 GB/s | 2-way or 4-way NVLink bridge, 900 GB/s per GPU |
| Host link | PCIe Gen5, 128 GB/s | PCIe Gen5, 128 GB/s |
| Server options (NVIDIA's wording) | "NVIDIA HGX H200 partner and NVIDIA-Certified Systems with 4 or 8 GPUs" | "NVIDIA MGX H200 NVL partner and NVIDIA-Certified Systems with up to 8 GPUs" |
| NVIDIA AI Enterprise | Add-on | Included |
Source: NVIDIA H200 specifications, which NVIDIA labels preliminary. Lenovo's guide states the NVL card includes a 5-year NVIDIA AI Enterprise subscription.
Memory capacity and bandwidth are identical. The NVL runs at a lower power ceiling, and NVIDIA lists its peak Tensor Core throughput about 16% below the SXM part. The bigger difference for most buyers is how the GPUs connect to each other.
How the GPUs talk to each other
H200 SXM on an 8-GPU HGX board
NVIDIA's HGX reference architecture lists 900 GB/s GPU-to-GPU bandwidth and 7.2 TB/s total aggregate bandwidth for the HGX H100 and H200 8-GPU baseboards, with up to 1,128 GB of GPU memory on the H200 version. Those baseboards use third-generation NVSwitch, and NVIDIA's Fabric Manager guide describes four NVSwitches on the 8-GPU board, with the same topology for H200 systems. Every GPU reaches every other GPU at full NVLink bandwidth.
The same guide states that two baseboards cannot be joined over NVLink. Scaling past eight GPUs goes over the network fabric.
The 4-GPU HGX H200 board is winding down. Lenovo's SR675 V3 guide says it "is planned to be withdrawn," and Supermicro marks its 4-GPU HGX H200 system, the SYS-421GU-TNXR, as end of life. Plan on the 8-GPU board when you quote SXM.
H200 NVL in a PCIe server
H200 NVL cards use one 2-way NVLink bridge per pair or one 4-way bridge per set of four, as Lenovo's SR675 V3 guide puts it. NVIDIA rates the link at 900 GB/s per GPU inside the bridged group. Four bridged cards hold 564 GB of HBM3e between them.
With 4-way bridges, an 8-card server holds two separate 4-card groups. Dell's PowerEdge XE7740 technical guide lists an "H200 x 8 (NVL4)" configuration built from two 4-way bridges, and a separate "H200 x 8 (No Bridge)" configuration. The two groups are not NVLink-connected to each other.
What that means for the workload
- A model you shard across all eight GPUs with tensor parallelism wants the HGX board, where every GPU has a full NVLink path to every other.
- A model that fits in 282 GB (two cards) or 564 GB (four cards) can run on an NVL group, and an 8-card NVL server can host two such replicas.
- Training that spans several servers runs over the network fabric between them, so quote the NICs and switches with the same care as the GPUs.
Benchmark your own model before you commit. NVIDIA's throughput figures are peak numbers with sparsity.
Which servers carry each one
These are confirmed on each vendor's own page or document as of November 5, 2026.
8x H200 SXM (HGX H200)
| Server | Notes |
|---|---|
| Supermicro SYS-821GE-TNHR | 8U, Intel Xeon. Lists HGX H100 8-GPU and HGX H200 8-GPU, NVLink with NVSwitch |
| Supermicro AS-8125GS-TNHR | 8U, AMD EPYC. Lists HGX H200 8-GPU |
| Dell PowerEdge XE9680 | 6U. "8 NVIDIA HGX H200 141GB 700W SXM5 GPUs" is one of its GPU options |
| Lenovo ThinkSystem SR780a V3 | Liquid-cooled HGX H200 8-GPU board |
The SYS-821GE-TNHR and XE9680 list both HGX H100 and HGX H200 boards, so the same chassis model can show up in quotes for either. Check which board is actually installed.
Some HGX H200 servers are already withdrawn. Lenovo has withdrawn the SR680a V3 from marketing, and ASUS marks the ESC N8-E11 as EOL. Those can still appear on refurbished quotes, with the support limits that come with a withdrawn product.
H200 NVL (PCIe)
| Server | Max H200 NVL | NVLink bridges |
|---|---|---|
| Dell PowerEdge XE7740 | 8 | 4-way, or no bridge |
| Lenovo ThinkSystem SR675 V3 | 8 (AMD EPYC 9005 models) | 2-way or 4-way |
| Supermicro SYS-422GA-NRT | 8 | Optional |
| Supermicro SYS-422GL-NR (MGX) | 8 | Optional |
| ASUS ESC8000A-E13 | 8 | 2-way or 4-way |
A PCIe GPU server is not automatically qualified for H200 NVL. Check that the card appears on the server's own supported-GPU list, at the GPU count you want, before you compare prices.
Why "H200 SXM5 card" quotes go wrong
A dealer who reads "SXM5" and "card" in the same line has to guess. One may quote a full HGX H200 server, another an H200 NVL, and a third a loose SXM module, which only runs on an HGX baseboard in a server built for it and will not work in a PCIe slot. You end up with quotes for different products that cannot be compared.
Decide first which one you need:
- You want eight GPUs with full NVLink between all of them: quote an HGX H200 8-GPU server.
- You want H200 memory in a standard PCIe server, or fewer than eight GPUs: quote H200 NVL cards in a server that lists them, with the bridge you need.
- You already own a server: send its exact model number and we will check whether it supports H200 NVL at all.
How to word the RFQ
For SXM, name the board and the server:
1x Supermicro SYS-821GE-TNHR (or equivalent HGX H200 server) with NVIDIA HGX H200 8-GPU, 8x H200 SXM 141 GB, NVIDIA part number 935-24287-2740-000 (air-cooled board). CPU, system memory, NICs, storage, and power supplies as listed. New or refurbished, with warranty terms. Delivered to [city, country] by [date].
For NVL, name the card count, the bridge, and the server:
1x Dell PowerEdge XE7740 (or equivalent server that lists H200 NVL) with 8x NVIDIA H200 NVL 141 GB PCIe, NVIDIA part number 900-21010-2740-030, with two 4-way NVLink bridges (900-23946-2700-000) installed. CPU, system memory, NICs, storage, and power supplies as listed. New or refurbished, with warranty terms. Delivered to [city, country] by [date].
The part numbers come from the ordering table in Lenovo's H200 product guide. The liquid-cooled 8-GPU board is 935-24287-2741-000 and the 2-way NVL bridge is 900-23945-2700-000. Ask each dealer to confirm the NVIDIA part number on the quote. That catches a substitution before you pay.
Also include:
- Air or liquid cooling, and the rack power you have per server
- Whether you need NVIDIA AI Enterprise (included with NVL, an add-on for SXM)
- The destination country and the buying entity. H200 is export-controlled, and the license question turns on where the server ships and where the buyer's ultimate parent company is headquartered. See our export controls guide.
Frequently Asked Questions
Can I buy a single H200 SXM GPU?
NVIDIA and the server makers sell H200 SXM as HGX boards with four or eight GPUs, installed in a server. A loose SXM module only works on an HGX baseboard in a server built for it, so for one or two GPUs in a PCIe server, the part to buy is the H200 NVL.
Is the H200 NVL slower than the H200 SXM?
NVIDIA lists the same 141 GB of memory and 4.8 TB/s of bandwidth for both. The NVL has a 600 W power ceiling against 700 W for the SXM, and NVIDIA's peak FP8 figure is 3,341 TFLOPS against 3,958 TFLOPS, both with sparsity. Memory capacity and bandwidth match, so a workload limited by memory bandwidth should show a smaller gap than a compute-bound one. Benchmark yours.
Can I bridge eight H200 NVL cards together?
The largest bridge joins four cards. An 8-card server runs as two 4-card NVLink groups, and traffic between the groups does not use NVLink.
Does H200 NVL come with NVIDIA AI Enterprise?
NVIDIA lists it as included with H200 NVL and as an add-on for H200 SXM. Lenovo's guide describes it as a 5-year subscription with the NVL card.
Get a sourced quote
gpu.fm does not hold inventory. We source HGX H200 servers and H200 NVL systems through dealers and quote on request, with the part numbers, warranty, and delivery terms written down. Send the configuration you want, or send the workload and we will suggest one.
