Putting an H200 NVL into a server you already own: the checks to do before you order

Check 1: slots, risers and host requirements

The card is 141 GB of HBM3e at 4.8 TB/s on a full-height, full-length, dual-slot PCIe board, ten and a half inches long, drawing up to 600 W through one 16-pin auxiliary connector. A retrofit has to pass seven checks in this order, and a free slot is the easiest of them.

NVIDIA H200 NVL Product Brief PB-12128-001_v01, April 2025, and HPE QuickSpecs NVIDIA Accelerators for HPE, c04123180 v83, September 2026, which gives the same dimensions. The enhanced extender is an option, not the default fitting.

Then count the slots the way the vendor counts them. Dell allows two H200 NVL in a PowerEdge R770 and names the risers that must be present, RC 6-2 and RC 11-2; Lenovo allows two in a ThinkSystem SR650a V4 and only in slots 21 and 23; HPE offers double-wide GPUs in a DL380a Gen12 in quantities of two, four or eight. The usable count is a property of the riser and cable set, not of the empty space at the back of the chassis.

The host is the last mechanical check. NVIDIA’s reference architecture calls for one Gen5 x16 link per GPU, accepts one per two GPUs, and sets minimums for the rest of the machine: two sockets, 2.0 GHz, seven physical cores per GPU and 128 GB of system memory per GPU. A server with 256 GB of RAM meets that minimum with two cards and falls below it with three.

Check 2: power, the 16-pin cable and its sense pins

Power has the most clearly documented failure of all the checks. The card takes up to 600 W from the 16-pin auxiliary connector; the slot itself contributes little, and Dell puts slot power at up to 75 W without an auxiliary cable. That cable carries sense pins declaring what the supply side will deliver, and NVIDIA states that if the power level identified by Sense0 and Sense1 is less than the default power cap of the card, the card will not boot.

Product Brief PB-12128-001_v01, sense-pin table, labelled PCIe CEM 5.1. Only the highest strap is supported: there is no lower-power cabling option for this card.

A cable strapped for a lower class still fits the connector. The card then does not start at all, although the server boots normally; it does not fall back to a lower power limit. The cable is also chassis-specific and slot-specific: Dell names DCNHT for slot 2 and K8FW0 for slot 7 of the same R770, and Lenovo names a 320 mm cable for the SR650a V4 and a 235 mm one for the SR675 V3. Order it by server and slot, by part number.

The configurable range is 600 W by default, with a 350 W power compliance limit and a 200 W minimum in the same table. The 350 W figure is a compliance limit inside the 600 W row, not a variant you can order. A cap set in band with nvidia-smi is lost on the next driver load; only the out-of-band SMBPBI setting survives a reboot.

Next, calculate the power budget with your facilities team, using the power the server maker’s GPU configuration matrix allows for the exact model and slot. Two cards at 600 W, as in Lenovo’s SR650a V4, are 1,200 W. Four at 600 W are 2,400 W, the entire output of one 2,400 W supply before processors, memory, drives and fans have drawn anything, and eight are 4,800 W. HPE’s DL380a Gen12 QuickSpecs accordingly require 2,400 W or 3,200 W supplies whenever the card is configured, with five power supplies for a two- or four-GPU node (two for the system board and three for the GPUs) and eight for an eight- or ten-GPU node. Once GPU draw exceeds the output of one supply, N+N redundancy needs more than two supplies. Dell’s instruction for the R770 is to verify consumption with its planning tool rather than a rule of thumb. At rack level the power feed runs out long before the rack space does, as our rack power and cooling guide shows.

Check 3: cooling and airflow

We found no published airflow figure for this card. The heat sink is passive and the card has no fan, yet the only technical content of the System Airflow Requirements chapter in NVIDIA’s product brief is that the sink accepts air left to right or right to left, plus a figure. It gives no rate in linear feet per minute or in cubic feet per minute, and neither does the equivalent chapter of the H100 PCIe brief. We found none in Lenovo’s product guide, HPE’s QuickSpecs, Dell’s technical guide and thermal matrices or Supermicro’s system-guidance paper either.

A general statement that the racks have good airflow therefore cannot be checked, because there is no published figure to compare it with, and an LFM number from a forum or a reseller page has no primary source behind it. NVIDIA’s brief gives no maximum die temperature or throttle threshold either; the card reports its own slowdown and shutdown thresholds in nvidia-smi. Instead, the OEMs publish a list of parts and limits, and that list serves as the thermal specification.

Dell R770 and R7725 thermal restriction matrices and the R770 Technical Guide; Lenovo SR650a V4 thermal rules; Product Brief PB-12128-001_v01 Table 2-4. None of these documents states an airflow rate.

Compare the last row with the others. The card itself runs in ambient air up to 45 °C, but the servers certified to hold it limit the room to 30 or 35 °C, so the server’s limit is the one to plan for.

Check 4: the server vendor’s support list

Four primary documents answer this: Lenovo Press LP1944, the ThinkSystem NVIDIA H200 141GB GPUs Product Guide; the HPE QuickSpecs NVIDIA Accelerators for HPE, c04123180 v83 of September 2026; the Dell PowerEdge R770 Technical Guide, part number E111S, Rev. A08 of May 2026; and the Dell PowerEdge Server GPU Matrix. Behind them sit the HPE ProLiant Compute DL380a Gen12 QuickSpecs and Supermicro’s System Guidance for GH200 and H200.

The vendor lists are short. Dell shows the card on seventeenth-generation systems only, at a maximum of two in the R770 and the R7725 and three in the R7715, and its PowerEdge XE7740 specification sheet lists eight; it is absent from the fifteenth and sixteenth-generation tables, so an R750 or R760 already in the rack is not on Dell’s list. HPE lists the ProLiant DL380a Gen12, the ProLiant DL380 Gen12 and the ProLiant DL385 Gen11; Lenovo lists the SR650a V4 at two and the SR675 V3 at eight, the latter with 5th Gen AMD EPYC processors only. NVIDIA frames the category as MGX H200 NVL partner and NVIDIA-Certified Systems with up to eight GPUs.

Lenovo’s guide shows how precise this is: the card is supported on the SR650a V4, model 7DGDCTO2WW, and explicitly not supported on the SR650 V4, model 7DGDCTO1WW. An order placed by family name rather than by full model number can therefore end in an unsupported configuration. Before you order, we check the platform, power and cooling of the server you already run. Send us the exact server model and our team will help.

The documents say little about fitting the card in the field rather than at the factory. None says the PCIe card is factory-install only, and none states a warranty consequence for fitting one in the field, so we will not claim either. Lenovo marks the HGX H200 baseboards configure-to-order only, while the PCIe adapter is a normal orderable option on ServerProven. They do scope support to a configuration rather than to a server: a machine type, a slot, a riser, a fan module, an air duct, a cable per slot and a drive backplane on the supported list. HPE puts the difference in qualification terms, saying accelerators sold for NVIDIA-Certified HPE servers undergo thermal, mechanical, power and signal-integrity qualification.

Dell’s R770 thermal restriction matrix, for example, marks the card not supported with whole classes of drive backplane: 24 × 2.5-inch SAS/SATA, 16, 32 and 40-bay EDSFF E3.S NVMe, and rear E3.S. The backplane sets the air impedance in front of the card, so the drive configuration chosen when the server was bought can rule out the GPU. The power the card may draw is set per model and slot as well, so check the server maker’s GPU configuration matrix for the exact model and slot rather than assuming 600 W. Dell also sets a minimum operating system for the R770, Ubuntu 24.04.2 with kernel 6.11 or later, caps the H200 NVL at 450 W in this server and requires High Performance Platinum fans with it, as its installation and service manual states.

Check 5: the NVLink bridge, if the workload needs one

For serving, most retrofits do not need a bridge: if each card holds its own copy of the model, the cards never talk to each other. Bridge topology, bandwidth and the checks on a delivered server are in our NVLink bridge guide. The bridge is worth fitting when a model is split across GPUs at every layer, because tensor parallelism inserts two all-reduces per layer in the forward pass and PCIe Gen5 at 128 GB/s is then the bottleneck. Training that exchanges gradients between the cards gains from the bridge too.

The topology rules are strict. Each H200 NVL carries exactly one bridge connector; the 2-way bridge joins two cards, the 4-way joins four, and in both cases the cards must be physically adjacent. NVIDIA’s product brief says bridged cards should be within the same CPU domain and lists bridging across CPUs as allowed but not recommended. In a server whose risers split four GPUs two and two across the sockets, a 4-way bridge would span both sockets, which NVIDIA does not recommend. The chassis must also leave at least 2.5 mm of clearance above the north edge of the card and 2.67 mm behind it.

Product Brief PB-12128-001_v01 Tables 4-1 and 4-2, and the H100 NVL Product Brief PB-11773-001_v01. NVIDIA labels the parts “2-slot” and “4-slot” while describing them as spanning two and four cards; since the card is itself dual-slot, read the labels as card counts and derive no pitch from them.

Keep the two bandwidth figures apart, because NVIDIA’s own material blurs them. Per-GPU NVLink bandwidth is 900 GB/s and it is the same on both bridges: 18 links at 25 GB/s in each direction, 900 GB/s in total; NVIDIA’s product brief prints 50 GB/s per lane “each direction”, which contradicts its own “900 GBps bidirectional”. NVIDIA’s product brief prints the 900 and 1,800 GB/s in a column headed “Total NVLink BW”, while NVIDIA’s H200 page gives 900 GB/s per GPU for both bridges; sizing uses 900 GB/s per GPU, and 4 × 141 GB gives the 564 GB of pooled memory NVIDIA quotes for a four-way set. NVIDIA’s web datasheet writes the row as 900 GB/s per GPU while the PDF version drops the words “per GPU”, which is where the confusion starts.

This check differs most for anyone upgrading from an H100 NVL, and one figure needs care. NVIDIA’s corporate blog calls the H200 NVL a 1.5 times memory increase and a 1.2 times bandwidth increase over H100 NVL; its developer blog, three weeks later, says 1.4 times bandwidth. The product briefs give 4,813 GB/s and 3,938 GB/s, a ratio of about 1.22, so against the H100 NVL the uplift is 1.2 times; 1.4 times matches the H200 against the H100 SXM, 4.8 against 3.35 TB/s. The interconnect uplift is the larger one, 900 against 600 GB/s, or 1.5 times. The old card also used three bridges per pair where this one uses a single bridge, and HPE’s DL380a Gen12 QuickSpecs state that the bridges for the H200 NVL “can only be supported with H200 NVL GPUs” and the older Ampere 2-way bridge only with the H100 NVL, so an existing set does not carry over. A bridged H100 NVL pair is 2 × 94 GB, or 188 GB, which is where the old “188 GB H100 NVL” figure comes from; the equivalent H200 NVL pair is 282 GB, as our comparison of the two cards works through. Note also that “up to eight GPUs” is a platform statement, not a bridge statement, because the NVLink domain stops at four cards.

We supply the H200 NVL with its two-way and four-way NVLink bridges; for an H100 NVL upgrade, the new bridges go into the same quote as the cards. Tell us how many cards you plan to link and in which server.

Check 6: firmware, BIOS and virtualisation settings

Two firmware minimums are documented, and both are easy to miss on a machine that has run for two years. For Secure Boot, NVIDIA names CEC firmware 2.0185 or later and NVFlash 5.842 or later; for the platform, NVIDIA-Certified Systems 2.8 or later.

The BIOS setting that decides whether the card initialises is Above 4G decoding, and NVIDIA’s knowledge base warns that vendors name it differently: 64-bit MMIO, Memory Hole for PCI MMIO, Above 4G Decoding. When it is wrong the symptoms are a failure to initialise NVML with an unknown error, or a message that the PCI I/O region assigned to the device is invalid. On some Supermicro boards the option is enabled globally but excluded per slot, so a setting marked enabled can be off for exactly the slot the GPU is in. On Dell sixteenth-generation and later systems the MMIO base setting is gone and the default is 2,048 TB; on Lenovo the setting is MM Config Base, and the documented fix for insufficient PCI resources is counter-intuitively to lower it.

Resizable BAR is often listed as an H200 NVL requirement, but we found no NVIDIA or OEM document that names it as one. NVIDIA’s product brief does publish the size of the card’s large memory window, 256 GiB, which the server has to map above 4 GB. The setting that matters is therefore the server-side one described above: 64-bit MMIO decoding, with enough MMIO space for a 256 GiB window per card.

If the card will be virtualised, two more settings become requirements. NVIDIA’s vGPU documentation states that VT-d or IOMMU must be enabled and SR-IOV switched on in advanced options; the card needs SBIOS cooperation for SR-IOV and exposes 32 virtual functions. For pass-through, 141 GB of frame buffer breaks MMIO defaults that worked for 24 and 48 GB cards. Broadcom’s knowledge base says to set pciPassthru.use64bitMMIO to true and to give the MMIO size in pciPassthru.64bitMMIOSizeGB as a power-of-two number of GB. NVIDIA’s vGPU documentation lists 512 GB of MMIO space for the H200, and the same 512 follows for one card in pass-through: its 256 GiB memory window plus the smaller register windows comes to just over 256 GiB, so the next power of two is 512. Two cards come to just over 512 GiB, so by our arithmetic they need 1,024.

One conflict cannot be solved by configuration. NVIDIA’s NCCL troubleshooting guide says IO virtualisation can redirect peer-to-peer PCI traffic through the CPU root complex, causing significant performance loss or a hang, and that the bare-metal fix is to disable it; check with sudo lspci -vvv | grep ACSCtl for SrcValid+, as NVIDIA’s NCCL guide writes it. The same page states that virtual machines require ACS to function, so a host gives you VM isolation or unrestricted bare-metal peer-to-peer, not both.

Check 7: driver, vGPU and licence

The software minimums are driver R565 TRD1 or later and vGPU 18.1 or later in the Virtual Compute Server edition. NVIDIA AI Enterprise documentation lists the H200 NVL vGPU types; confirm hypervisor support with NVIDIA for the release you deploy. NVIDIA’s brief also writes “CUDA 12.7 or later”, which is the CUDA version of the R565 driver: we found no CUDA 12.7 toolkit release, and applications built with CUDA 11.8 or any later toolkit support Hopper. NVIDIA does not designate a current branch for this card in its driver release notes, so treat the branch as a lifecycle decision: of those listed in the September 2026 data centre driver document, R580 LTS has the longest published life, to June 2028, against March 2027 for R595 Production and August 2026, now past, for R610. On a server that has run for years, also check the installed version: NVIDIA’s security bulletin of 30 September 2026 names 580.178.04 and 595.91.07 as the fixed versions of these branches, and the newest releases are 580.196.07 of 21 October 2026 and 595.112.07 of 22 October 2026.

The last item is the licence entitlement. Every H200 NVL carries a five-year NVIDIA AI Enterprise subscription, where H200 SXM boards outside NVIDIA’s DGX systems need it as an add-on. It is licensed per GPU rather than per server and activated with the GPU serial numbers, but NVIDIA counts the five years from the board’s shipment to the server maker plus 90 days, activated or not. The serials are printed on the boards, which are hard to reach once the cards sit in a racked server, so read them off before the cards go in.

Three checks that are easy to miss

Each of these three shows up at a different stage.

1. The 16-pin cable is strapped for the wrong power class. The card does not start although the server boots. This shows at first power-on and looks like a dead GPU, but the cause is a wrong cable part number, not a fault in the card.
2. The chassis is not on the OEM’s thermal list for this card. This shows at order and validation time rather than at boot, and fan tuning does not fix it, because the list is the specification. The drive backplane is easy to overlook.
3. MMIO and Above 4G decoding, especially once a hypervisor is involved. A 141 GB card makes this far more likely than the 24 to 48 GB cards most existing servers were sized for, and the per-slot exclusion can remain even after the global setting has been checked.

What we supply

We supply the NVIDIA H200 NVL with its NVLink bridges and its five-year NVIDIA AI Enterprise subscription, and AI servers built to order around it, on one EU contract and invoice with manufacturer warranty. For a server you already run, we check the platform, power and cooling before you order and tell you in advance if the card is not supported in it. The same checks apply to the RTX PRO 6000 Server Edition, which our compatible servers guide covers. If you want the deployment done for you, our engineering partner Vixen.UNO handles it under the same contract.

FAQ