NVIDIA H200 NVL specs: TFLOPS per precision, 141 GB HBM3e, power, NVLink and MIG

H200 NVL specifications at a glance

The NVIDIA H200 NVL is the PCIe version of the H200: 141 GB of HBM3e at 4.8 TB/s on a dual-slot, full-height, full-length card, passively cooled, with up to 600 W of board power. Up to four cards connect over an NVLink bridge at 900 GB/s per GPU, and servers take up to eight. NVIDIA lists 3,341 TFLOPS of FP8 and 1,671 of BF16 with sparsity, 30 TFLOPS of FP64 and no FP4 rate.

NVIDIA H200 product page (updated 27 August 2026, marked "Preliminary specifications"), NVIDIA H200 NVL product brief PB-12128-001_v01 (11 April 2025) and NVIDIA's CUDA GPU list (16 October 2026), read 24 October 2026.

NVIDIA's H200 page shows the NVL and the SXM version side by side and marks the whole table as preliminary specifications that may change. The product brief, an engineering document for server makers, adds clocks, power modes, mechanical and thermal data. This page combines both; benchmarks for the card are in our article on H200 NVL benchmarks.

TFLOPS per precision: sparse and dense

NVIDIA footnotes every Tensor Core figure on the H200 page "With sparsity". That rate counts on weights pruned to a 2:4 pattern; ordinary model weights run at the dense rate, which is half. The FP64 and FP32 figures carry no such footnote.

NVIDIA H200 product page, H200 NVL column, footnote 2 "With sparsity", read 24 October 2026; dense values are the listed sparse values halved.

For the query behind many searches, BF16: 1,671 TFLOPS with sparsity, 835.5 dense. The SXM version in the same table lists 1,979 and 3,958 TFLOPS for BF16 and FP8 at up to 700 W, about 18 per cent more than the NVL, as our comparison of the H200 NVL and H200 SXM explains. Our GPU TFLOPS comparison in FP4, FP8 and BF16 sets these rates against the RTX PRO cards.

FP64 is where the H200 NVL stands apart from the RTX PRO cards: 30 TFLOPS, or 60 on the Tensor Cores, for simulation codes that need double precision. Our page on FP64 GPU servers covers which codes use it.

FP4 on the H200 NVL

NVIDIA's table lists no FP4 rate for the H200 NVL, and the support matrix of NVIDIA's Transformer Engine 2.22 lists BF16, per-tensor FP8 and FP8 block scaling for Hopper (compute capability 9.0), with MXFP8 and NVFP4 not supported. The H200 NVL therefore has no FP4 arithmetic; FP8 is its lowest Tensor Core precision. An FP4 checkpoint can still run: vLLM loads it and computes in 16-bit, which saves memory but not compute, as our article on FP4 models on the H200 NVL and Ada GPUs explains with the engines and their limits.

Memory: 141 GB of HBM3e, 140.4 GiB in nvidia-smi

The product brief gives 141 GB of HBM3e, a memory clock of 3,201 MHz, a peak bandwidth of 4,813 GB/s and ECC as enabled. NVIDIA's blog of 18 November 2024 puts it at 1.5 times the memory and 1.2 times the bandwidth of the H100 NVL. The product page rounds the bandwidth to 4.8 TB/s; we plan with that lower figure. The driver reports 140.4 GiB.

By our users-per-card rule, 90 per cent of that less 3 GiB, 123.4 GiB per card holds weights and KV cache. Llama 3.3 70B in FP8, 67.7 GiB of weights, then leaves room for 44 conversations of 8K with an FP8 KV cache on one card. The bandwidth sets the ceiling for one stream of generated tokens: about 70.6 GB read per token for that model gives about 68 tokens per second, by our arithmetic, an upper bound that real runs stay below. Our GPU memory bandwidth comparison puts HBM3e next to GDDR7.

Compute capability, clocks and SM count

NVIDIA's CUDA GPU list, dated 16 October 2026, gives the H200 compute capability 9.0, in one row with the H100 and GH200 and without a separate NVL entry. The product brief gives a base clock of 1,230 MHz and a boost clock of 1,785 MHz.

An SM count for the H200 NVL is in neither the product page nor the product brief. NVIDIA's MIG user guide describes the card's profiles in sevenths of its streaming multiprocessors, so one 1g.18gb instance holds a seventh of them, but it does not give the total. Figures in third-party databases are not NVIDIA's, and we leave them out.

Power: 600 W by default, 200 W minimum

The product brief lists 600 W as the maximum and default board power and 200 W as the minimum. The card takes one 16-pin connector, which must be cabled for the 600 W mode: the card checks the sense pins and does not boot on a cable that signals less than its default, even when a lower cap is set. A lower limit is set by software or by the server's management controller, as our article on the H200 NVL at 450 W shows with the servers that cap it.

NVIDIA product brief PB-12128-001_v01; server totals by our estimate as in our article on the H200 NVL at 450 W: cards plus two AMD EPYC 9355 (280 W each) for four cards or two EPYC 9555 (360 W each) for eight, plus 15 per cent for the rest of the server, to be confirmed by a measurement on the delivered system.

In its announcement of the H200 NVL, NVIDIA cites a survey finding that "roughly 70% of enterprise racks are 20kW and below and use air cooling". We build H200 NVL servers with two, four or eight cards; send us the models, users and the power feed at the rack position, and our team replies with a configuration and its power per feed.

NVLink bridges and PCIe

Each H200 NVL has 18 NVLink links and takes one bridge. The product brief lists two bridges: a 2-slot bridge, part number 900-23945, for two cards at 900 GB/s, and a 4-slot bridge, 900-23946, for four cards, listed at 1,800 GB/s. NVIDIA's H200 page gives one figure for both, 900 GB/s per GPU, and the brief does not say how its 1,800 GB/s for the four-way bridge is counted. We plan with 900 GB/s per GPU, the figure both documents support. NVIDIA's blog states that NVLink is "7x faster than fifth-generation PCIe"; with 900 against 128 GB/s on the H200 page, that is the ratio.

NVIDIA H200 product page and product brief PB-12128-001_v01, read 24 October 2026.

With four-way bridges, an eight-card server holds two NVLink domains of four, joined over PCIe. Which bridge layouts each server maker allows is in our list of H200 NVL compatible servers, and what the two domains mean for one large model in our guide to the H200 NVL NVLink bridge, 2-way or 4-way.

Thermal and mechanical data

The card has a passive heat sink that takes air in either direction through the server, an operating ambient range of 10 to 45 °C in the product brief, and a board weight of 1,217 g without bracket and bridge. It is 10.5 inches long, full height and dual slot. A server maker can set a lower inlet limit for its own chassis than the card's range; our compatible-servers list records the limits the makers publish.

MIG, vGPU and NVIDIA AI Enterprise

NVIDIA's MIG user guide, updated on 13 October 2026, lists these profiles for the H200 141 GB product. NVIDIA's product page gives up to seven instances "@ 16.5GB"; the guide names the smallest profile 1g.18gb, an eighth of the memory. We plan with the guide's profiles, the ones the driver creates.

NVIDIA MIG user guide, supported MIG profiles, H200 141GB table, updated 13 October 2026, read 24 October 2026; 1g.18gb+me (one instance) not shown.

Virtual machines get the card through compute vGPU (C-series) under NVIDIA AI Enterprise. The vGPU type list of NVIDIA AI Enterprise 8, updated 2 September 2026, names the "NVIDIA H200 PCIe 141 GB (H200 NVL)" with MIG-backed types from H200-7-141C down to seven H200-1-18C per GPU and time-sliced types down to 32 H200-4C per GPU. NVIDIA's list of GPUs supported by its vGPU software, updated 2 October 2026, has no Hopper GPU; its note says that GPUs supporting only C-series types are no longer covered there from vGPU 16.0 and "continue to be supported with NVIDIA AI Enterprise". The type list marks two H200 types as supported on ESXi from vSphere 8.0 Update 3 and names no other hypervisor, so we confirm the hypervisor and release against NVIDIA AI Enterprise's support matrix before planning vGPU. Our H200 NVL MIG profiles and the guide to GPUs in vSphere: passthrough or vGPU cover the setup.

The licence comes with the card. NVIDIA's licensing guide, updated 2 September 2026, states: "Each NVIDIA H200 NVL Tensor Core GPU includes a five-year NVIDIA AI Enterprise subscription." It starts at the ship date of the board to the server maker plus 90 days, a start date that "cannot be modified, as it is tied to the specific card"; the subscription is activated with the GPU's serial number and covers the NVIDIA-Certified system the card is installed in. Our article on NVIDIA AI Enterprise licensing explains the editions.

H200 NVL and H200 SXM in NVIDIA's table

NVIDIA lists both H200 versions in one table, and many published spec sheets mix their columns. The SXM module sits on an HGX board with NVLink switches; the NVL is a PCIe card with bridges. The memory is the same, the compute and power are not.

NVIDIA H200 product page, updated 27 August 2026, "Preliminary specifications", read 24 October 2026.

A figure of 3,958 TFLOPS, 700 W or "4 or 8 GPUs" in a spec sheet therefore belongs to the SXM version. NVIDIA lists server options for the NVL with up to eight GPUs, in MGX and NVIDIA-Certified systems, and for the SXM HGX boards with four or eight. What the shorter NVLink reach of the bridges means for one model across eight cards is the subject of our H200 NVL vs H200 SXM comparison.

Driver, security and release

NVIDIA product brief PB-12128-001_v01; NVIDIA blog "Hopper H200 NVL", 18 November 2024; read 24 October 2026.

NVIDIA announced the H200 NVL on 18 November 2024 and wrote that platforms with it "will be available from NVIDIA's global systems partners beginning in December". Against the H100 NVL, NVIDIA claims up to 1.7 times the LLM inference and 1.3 times the HPC performance; our comparison of the H200 NVL and H100 NVL looks at those figures and their conditions.

What we supply

You can buy the H200 NVL from us: our team sizes the configuration to your workload and budget and replies within one business day. We supply the cards across the EU, with NVLink bridges, and build AI servers to order with two, four or eight H200 NVL, assembled and tested before delivery, on one EU contract and invoice with manufacturer warranty on every component. Quantities and lead time are confirmed in writing with the quote, and our guide to buying the H200 NVL in the EU explains the steps. Ask us for an H200 NVL configuration with your server model or workload, and we check bridges, cables and inlet limits before the quote.

FAQ