NVIDIA H200 NVL at 450 W: the power range, the servers that cap it and what the cap costs

The H200 NVL at 450 W in NVIDIA's documents

The NVIDIA H200 NVL can run at 450 W as a power cap inside the range NVIDIA documents for it, 200 to 600 W, while its power cable still has to signal the full 600 W class. Dell's 2U PowerEdge R770 caps the card at 450 W, and StorageReview's October 2026 measurements of the card were taken at that level. How much throughput the cap costs against 600 W has not been published for this card as of October 2026, so any plan built on 450 W should include a measurement on your own model.

The figures come from Table 2-1 of NVIDIA's product brief PB-12128-001_v01 of 11 April 2025, for a card whose 16-pin cable is "strapped for 600 W power mode". That is the only mode the brief lists.

NVIDIA H200 NVL product brief PB-12128-001_v01 (11 April 2025), Tables 2-1, 2-4 and 4-3, read on 24 October 2026 from a third-party copy and cited by document number.

NVIDIA's H200 page puts it more briefly: "Up to 600W (configurable)". The brief's first chapter says the card "operates unconstrained up to its maximum thermal design power (TDP) level of 600 W". It lists a base clock of 1,230 MHz and a boost clock of 1,785 MHz, and says nothing about performance at lower power. Its section on programmable power gives the purpose of a cap: "for system power and thermal budget or performance-per-watt reasons".

Why a 450 W cable does not run the card

The 16-pin connector carries two sense pins that tell the card what the cable can deliver. Table 4-3 of the brief maps them to power classes, and only the top class is supported.

NVIDIA H200 NVL product brief PB-12128-001_v01, Table 4-3, read on 24 October 2026.

The brief states that if the signalled power is below the card's default cap, "the card will not boot", and that the card "requires this sense pin logic to boot, even if using the programmable TGP feature to reduce TGP". A cap of 450 W is therefore set on a card that sees a 600 W cable. Supermicro's support FAQ of 13 February 2026 describes the failure. A customer swapped H100 cards for H200 NVL in a SYS-521GE-TNRT, and the new cards did not appear in nvidia-smi or lspci. Supermicro's answer: "Power Cable SKU: CBL-PWEX-0962Y-30 only supports up to 450W", "H200 NVL GPUs require 600W each", and cable CBL-PWEX-1364Y-30 supports up to 600 W.

The RTX PRO 6000 Blackwell Server Edition behaves differently: its brief lists a 450 W cable mode in which the card runs, as our article on Server Edition power at 600, 450 and 300 W shows. A cable that suits that card in a given slot can leave an H200 NVL dark. Our checklist for an H200 NVL in an existing server covers the cable part numbers by server.

Setting the cap: nvidia-smi or the management controller

The brief's section on programmable power lets the server maker or operator set the card's power cap in two ways. In band, through the driver with nvidia-smi, "power cap adjustment must be reestablished after each new driver load". Out of band, through SMBPBI, the interface the server's management controller uses, "power cap adjustment remains in force across driver loads and system boots". The brief's own example enables persistence mode with nvidia-smi -pm 1 and sets a cap with nvidia-smi -pl 310; the default returns with nvidia-smi -pl 600 or a reload of the driver.

For 450 W the command is nvidia-smi -pl 450, and nvidia-smi -q -d POWER shows the current, default, minimum and maximum limits. Two practical consequences follow. A feed sized for capped cards depends on the cap being set again at every boot and driver update, by a systemd unit or by the management controller. And a cap set out of band by the server maker, as in Dell's R770, is part of the platform and not something the operator lifts with nvidia-smi. Our guide to the RTX PRO 6000 power limit with nvidia-smi shows the unit file and DCGM for a fleet; the commands are the same for the H200 NVL.

Which servers cap the H200 NVL at 450 W

In the server makers' documents we read on 24 October 2026, one server caps the card, and the others list it at 600 W.

Dell PowerEdge R770 technical guide Rev. A09 (June 2026); Lenovo Press LP2128 (20 October 2026) and LP1611 (8 October 2026); Dell PowerEdge XE7740 technical guide Rev. A04 (April 2026); Cisco UCS C845A M8 data sheet; read on 24 October 2026.

Dell's guide for the R770, a two-socket 2U server with Intel Xeon 6, states that "The NVIDIA H200 NVL and NVIDIA RTX Pro 6000 are designed to support up to 600 W" and that "they are power capped to 450 W support on R770". It lists at most two H200 NVL, in riser positions RC 6-2 and RC 11-2, and its summary table reads "Up to 2x double width 450 W and 6 x single width 75 W accelerators". The configuration also requires high-performance Platinum fans in every position, iDRAC firmware 1.20.60.55 or later, without which a fan mismatch error appears and fan speed is held at 50 per cent, Ubuntu 24.04.02 with kernel 6.11 or later, and slot-specific power cables, DCNHT for slot 2 and K8FW0 for slot 7. The guide does not mention NVLink, so the R770 holds two capped cards as separate GPUs.

We found no maker document that offers a 450 W option for four or eight H200 NVL, comparable to the 450 W slot-capped Server Edition part Lenovo sells. For any other server, the maker's GPU configuration matrix for the exact model and slot decides; our list of H200 NVL compatible servers gives the documents per maker. We supply the cards for servers like these and build our own servers with two to eight of them; send us the server model and the slots you plan to use, and we check cable, power and cooling before you order.

What is published about the cost of 450 W

Three published sources touch the question, and none answers it directly for this card.

StorageReview's ranking of enterprise GPUs, published 14 October 2026 and modified 20 October, tested one H200 NVL at 450 W, which it gives as "the most a Dell PowerEdge R770 GPU slot delivers", with NVIDIA's NGC vLLM 26.08 container (vLLM 0.27.1, CUDA 13.4.1) and driver 595.71.05. It writes: "The card is rated to 600W, and we have not run it there, so its figures on this page are a floor."

StorageReview, Best Enterprise GPUs (14 October 2026, modified 20 October 2026), one H200 NVL at 450 W, Ministral 3 14B in FP8 for the chat rows, read on 24 October 2026.

Even capped, the card ran 50 to 81 per cent ahead of an RTX PRO 6000 Workstation Edition at 600 W on three FP8 and BF16 models at 32 streams, and behind it on an NVFP4 model, 467 against 616 tokens per second, because Hopper has no FP4 arithmetic. StorageReview also found it the only card in its ranking that kept scaling past 32 streams.

The H200 NVL's MLPerf® result was run at 600 W. Every system description in the round that states a board power gives 600 W, as our H200 NVL benchmarks record. In MLPerf® Inference: Datacenter v6.0, closed division, Llama 2 70B (llama2-70b-99), offline scenario, Dell's PowerEdge XE7740 with eight H200 NVL reached 32,004.0 tokens per second, entry 6.0-0021, retrieved from mlcommons.org on 24 September 2026, result verified by MLCommons Association: about 4,000 per card, our division of the system result, not an MLPerf metric. Model, software and load differ from StorageReview's tests, so the two do not combine into a cost of the cap. The MLPerf name and logo are registered and unregistered trademarks of MLCommons Association in the United States and other countries. All rights reserved. Unauthorized use strictly prohibited. See www.mlcommons.org for more information.

For the RTX PRO 6000 Workstation Edition, StorageReview measured both levels and found a loss of 2 to 5 per cent at short context and 4 to 9 per cent at long context at 450 W. That card has a different chip and GDDR7 memory, so we do not carry the figure over to the H200 NVL.

What the sources allow is a reading of where the cap can bind. A cap only takes effect when the card would draw more than the cap. In the Watt Counts study of April 2026, as our guide to GPU energy per token records, one H200 NVL serving models of 20B to 27B parameters through vLLM drew 131.5 W on average at low load and 336.7 W at high load, below 450 W. Token generation at moderate concurrency is limited mainly by memory bandwidth, which a power cap does not change. Work that keeps the Tensor Cores busy, such as prefill of long prompts, large batches, fine-tuning and FP64 solvers, draws the most power and loses clock speed first. For these, measure before deciding.

kW per server at 600 W and at 450 W

Our estimate: board power from NVIDIA's product brief; two AMD EPYC 9355 at 280 W each for two and four cards, two EPYC 9555 at 360 W each for eight; plus 15 per cent for memory, drives, network cards, fans and conversion losses.

We plan a 230 V 16 A feed at 2.94 kW, 32 A at 5.89 kW and 400 V three-phase 16 A at 8.87 kW, 80 per cent of the rating, as in our guide to GPU server power on 16 A, 32 A and three-phase feeds. By our estimate the cap moves two configurations to a smaller feed. A four-card server at 600 W, about 3.4 kW, is above one 16 A feed; at 450 W, about 2.7 kW, it stays below. An eight-card server at 600 W needs a three-phase position; at 450 W, about 5.0 kW, it fits a 32 A single-phase feed. A two-card server fits a 16 A feed either way.

The power supplies follow the maker's rules for the card, which do not change with a cap. HPE requires 2,400 or 3,200 W supplies whenever the H200 NVL is configured in its DL380a Gen12, and Dell's 3,200 W supply for the XE7740 delivers 3,200 W at 220.1 to 240 V and 2,900 W at 200 to 220 V. A cap set in band also lapses after a driver reload until it is set again, so we size the feed on the capped figure only when the cap is set out of band or by a service at boot, and a measurement on the delivered system confirms it. Cooling needs fall with the cap at about 160 CFM of air per kW, by our guide to GPU rack power and cooling.

When to run the H200 NVL at 450 W

Run 600 W wherever the feed and the chassis allow it, and in particular for fine-tuning, long-context prefill, large-batch inference and FP64 work, where the card draws the most. Use 450 W where it decides the installation: a four-card server on a 16 A feed, an eight-card server on a 32 A single-phase feed, or a 2U server whose maker caps the card, such as the R770. For chat and RAG serving at moderate concurrency, the published draw under load suggests the cap often does not bind, but a run of your own model at both settings is the only figure that settles it, and a measurement on the delivered system confirms or corrects our power estimate. We size power mode, feed and chassis together; tell us the feed per rack position and the workload, and our team returns a configuration with the power per feed. The full sizing for each card count is in our guides to the 2× H200 NVL server, the 4× H200 NVL server and the 8× H200 NVL server.

What we supply

You can buy this system from us: our team sizes the configuration to your workload and budget and replies within one business day. We supply the H200 NVL as cards with their NVLink bridges and the power cables the server maker lists, and build AI servers to order with two, four or eight cards at 600 W or with a cap set at the level your feed needs, assembled and tested before delivery, on one EU contract and invoice with manufacturer warranty on every component. Each card includes a five-year NVIDIA AI Enterprise subscription, according to NVIDIA's licensing guide.

FAQ