HPE ProLiant servers for AI: which NVIDIA GPUs each model takes, DL380a Gen12 to XD685

HPE ProLiant GPU support, model by model

Which GPUs an HPE ProLiant server takes is set per model in HPE's QuickSpecs. As read on 12 November 2026, the 4U HPE ProLiant Compute DL380a Gen12 is the accelerator server of the line: most double-wide GPUs can be ordered in quantities of up to ten, the NVIDIA RTX PRO 6000 Blackwell Server Edition up to eight. The 2U models DL380 Gen12, DL345 Gen12 and DL385 Gen11 take three or four double-wide cards, and the systems with eight GPUs on an HGX baseboard are the HPE ProLiant Compute XD685 and the HPE Compute XD690.

We supply HPE servers per project across the EU; tell us the model, the cards and how many of them and you receive a written quote within one business day. The facts below come from the QuickSpecs of each model, the "NVIDIA Accelerators for HPE QuickSpecs" and HPE's product pages. Three QuickSpecs, those of the DL380 Gen12, the DL345 Gen12 and the DL385 Gen11, could be read only in part, so their version, date and option lists are missing here.

HPE QuickSpecs a00047453enw (DL380a Gen12), a50009233enw (DL345 Gen12), a50004300enw (DL385 Gen11), a00073553enw (XD685; change log Version 16, footer V15), a50009220enw (HPE Compute XD690), a50009209enw (DL384 Gen12) and a50014668enw (DL585a Gen13); HPE's product page for the DL380 Gen12. Read 12 November 2026.

A QuickSpecs is cited by its document number and version, because the dates inside can differ: the accelerator QuickSpecs carry Version 84 of 8 October 2026 in the change log and 19 October 2026 in the footer, and the XD685 document shows Version 16 in the change log and V15 in the footer. Product pages can differ from the QuickSpecs: the comparison table on HPE's AI servers page does not mention the H200 NVL.

HPE DL380a Gen12 GPU configurations with eight or ten cards

The DL380a Gen12 runs with two Intel Xeon 6 processors only. Its QuickSpecs speak of "up to 10 cutting-edge double-wide GPUs" and set the quantities: "0, 1, 2, 4, 8, and 10 GPU configurations are allowed" and "Most double-wide GPUs can be selected in quantities of 2, 4, 8, and 10." The RTX PRO 6000 Blackwell Server Edition (HPE SKU S6A73C) is the exception.

"HPE ProLiant Compute DL380a Gen12 QuickSpecs", a00047453enw, Version 21, 02-Nov-2026, read 12 November 2026.

For the H200 NVL (S3U30C) this means ten cards without NVLink bridges, or at most eight with HPE's 2-way and 4-way bridges. The base model matters as well. The configure-to-order server P76706-B21 is the one for the RTX PRO 6000 Server Edition; the second model, P74461-B21, is "ONLY available with iLO6" and does not support that SKU, and the card is also excluded from the "8DW NIC FIO Config" (P75011-B21). HPE lists a GPU Upgrade Kit, P81677-B21, as "field upgradable to support up to 8 DW GPUs".

Sixteen single-wide cards appear in the name of the model P76706-B21 ("8 Double Wide/16 Single-Wide") and on HPE's product page ("Up to 16 Single-Wide NVIDIA L4"). The part of the QuickSpecs we could read gives no rule for sixteen, and the accelerator QuickSpecs, as we read them, do not name the DL380a Gen12 in their list for the L4, so have that configuration confirmed in writing.

The chassis is 80.26 cm deep and weighs 37.5 to 62.5 kg, and it takes NVMe drives only. The QuickSpecs list direct liquid cooling as a kit for the processors; if it is selected, the ten-card configuration cannot be.

HPE's entry in MLPerf® Inference v6.1 uses eight RTX PRO 6000 Server Edition cards, the maximum the QuickSpecs allow.

MLPerf® Inference: Datacenter v6.1, closed division, Llama 2 70B (llama2-70b-99.9) and gpt-oss-120b, offline and server scenarios, tokens per second, entry 6.1-0050 (HPE ProLiant Compute DL380a Gen12, eight RTX PRO 6000 Blackwell Server Edition), retrieved from mlcommons.org on 24 October 2026. Result verified by MLCommons Association. 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.

A result with ten of these cards that turns up in searches belongs to an earlier round. It is a benchmark configuration, described in our article on RTX PRO 6000 benchmarks, and the QuickSpecs do not let you order it. What eight cards hold is worked out in which LLMs fit eight RTX PRO 6000 Server Edition cards, and the other makers' entries are in our MLPerf Inference v6.1 comparison.

DL380 Gen12, DL345 Gen12 and DL385 Gen11 with three or four cards

Three ProLiant models in 2U take the large cards in smaller numbers.

The HPE ProLiant Compute DL380 Gen12 has up to two Intel Xeon 6 processors. HPE's product page states "Up to three double-wide (DW) or eight single-wide (SW) GPUs." and names "Up to 8 Single-Wide NVIDIA L4 or RTX A1000 or 3 Double-Wide NVIDIA L40S, H100 NVL or RTX PRO 6000 Blackwell Server Edition" GPUs. The accelerator QuickSpecs differ from the page on two points: they list the H200 NVL for this server, and they removed the H100 NVL in Version 79 of 20 April 2026. The per-card maximums in the server's own QuickSpecs could not be read, so three double-wide cards is the product-page figure. The RTX A1000 named there is an 8 GB single-slot card that we supply; see buy the RTX A1000.

The HPE ProLiant Compute DL345 Gen12 has one AMD EPYC processor of the fifth generation. Its QuickSpecs announce "New GPU support, up to six (6) single-width or four (4) double-width GPUs." and require the GPU configure-to-order model P81949-B21. Another line of the same document reads "4 Single-Width or 4 Double-Width GPUs", so the single-width count is not consistent; four double-width cards is the figure on which both lines and the product page agree.

The HPE ProLiant DL385 Gen11 has two AMD EPYC processors of the fourth or fifth generation. Its QuickSpecs give "8 single wide (SW) or 4 double wide (DW) GPUs" in the GPU configure-to-order server P54198-B21 and list "4x H200 NVL, 4x RTX PRO 6000 BSE" among the newly supported configurations. The "Maximum Performance Fan kit is only supported when 600W GPUs are in the configuration", and a further entry reads "25C~30C server ambient temperature is now supported."

HPE's portfolio page lists no ProLiant DL385 Gen12. It names the DL385 Gen11 and the DL365 Gen11 as the two-socket AMD models, and the AMD models of Gen12, the DL325 Gen12 and the DL345 Gen12, have one socket. For a ProLiant you already run we supply the cards after a check of the platform, power and cooling; write to us with the server model and its configuration, and we say in advance if something will not work. The order of checks is in our guide to adding a GPU to an existing rack server.

Which NVIDIA card HPE lists for which ProLiant server

The accelerator QuickSpecs turn the question round and list the servers per card.

"NVIDIA Accelerators for HPE QuickSpecs", c04123180, footer "V84 - 19-October-2026", change log Version 84 of 08-Oct-2026, read 12 November 2026. ProLiant servers only; the lists also name edge and scale-up systems.

HPE heads these lists "Supported Servers" and logs additions as "sustaining quals"; neither term is defined. Under the lists it writes: "Please see the platform QuickSpecs for configuration rules including enablement kits." In practice a card can be ordered for a server when the server's own QuickSpecs list the SKU with its rules, down to the count, the kit, the cable and the power supplies. The H200 NVL entry carries the note "Includes a five-year NVIDIA AI Enterprise subscription." The 32 GB single-slot RTX PRO 4500 Blackwell Server Edition is listed for six ProLiant servers, more than either of the two larger cards; we supply it as a card as well, see buy the RTX PRO 4500 Server Edition. Two further SKUs in the document, the RTX PRO 6000D (S6W21C) and the L20 (S4A92C), are headed "(China, Hong Kong, Macau only)". Our lists of servers compatible with the RTX PRO 6000 Server Edition and of H200 NVL compatible servers set HPE's counts next to those of other makers.

HPE ProLiant DL384 Gen12 is a Grace Hopper server

The HPE ProLiant Compute DL384 Gen12 differs in architecture from the card servers above. It is a 2U air-cooled server built on one NVIDIA GH200 Grace Hopper Superchip, or on two of them as "NVIDIA GH200 NVL2" with 144 Arm Neoverse V2 cores. The GPU is part of the superchip: 144 GB of HBM3e in the single version and 288 GB in the NVL2 version, next to 480 GB or 960 GB of LPDDR5X. HPE calls it "the first server from HPE enabled with NVIDIA GH200 NVL2".

The server takes no PCIe GPU cards, so none of the accelerators in the table above is an option for it. The version with two superchips runs on four Flex Slot Titanium supplies of 1,800 to 2,200 W in a 3+1 layout and holds up to eight E3.S NVMe drives; the chassis is 789 mm deep. Its processors are Arm cores, which sets it apart from the Xeon and EPYC models on this page. The QuickSpecs, V17 of 5 October 2026, are current.

HPE ProLiant Compute XD685 and HPE Compute XD690 for HGX B300

Eight GPUs on an HGX baseboard come in two HPE systems, and the cooling decides between them.

"HPE ProLiant Compute XD685 QuickSpecs", a00073553enw (change log Version 16 of 03-Aug-2026, footer V15), and "HPE Compute XD690 QuickSpecs", a50009220enw, V7 of 02-November-2026, read 12 November 2026.

In the XD685 the cooling follows the GPU: HGX H200 with air or direct liquid cooling, HGX B200 and HGX B300 with direct liquid cooling only. The QuickSpecs add: "The B300 DLC accelerator can only be selected with the XD685 DLC High Power server (S6Y07A)". With B300 the default is twelve power supplies, ten "can be deployed as needed", and HPE speaks of "12-PSU N+3 redundancy configurations". The chassis is 1,049 mm long and weighs 103 kg fully loaded.

HPE's air-cooled HGX B300 system is the XD690, with "Eight NVIDIA® Blackwell Ultra SXM6 GPUs" "on the NVIDIA HGX B300 platform". The QuickSpecs give "12 power supplies, each rated at 3200 W, providing N+N redundancy" and a system power of 14.3 kW continuous and 16.4 kW peak. The system does not carry the ProLiant name: its product page sits under HPE's Cray path and it is managed through a BMC with Redfish, while HPE's EMEA store lists it next to the XD685 in the category "ProLiant Compute XD600 Servers". The accelerator SKUs of both systems carry "HGX B300 288GB" in their names. What N+N and N+3 mean at the feeds is the subject of PSU redundancy in GPU servers. Our comparison DGX B300 vs HGX B300 vs DGX B200 shows what the server maker adds around NVIDIA's baseboard, and the same baseboard at two other makers is covered in Dell PowerEdge XE servers and Supermicro's SYS-822GS-NB3RT.

An XD685 or XD690 with HGX B300 is a project order with us: outline what the system is for and where it will stand. The distribution asks for the company, the project and the intended use before a quote, and with these we check price and availability for your project.

DL585a Gen13 and three further models HPE has announced

For the HPE ProLiant DL585a Gen13 a first QuickSpecs exists, V1 of 7 October 2026. It describes a 10U server with two AMD EPYC 9006 processors "designed for up to eight qualified double-wide PCIe GPUs" in "0, 1, 2, 4, and 8 GPU configurations". The RTX PRO 6000 Blackwell Server Edition is listed and the H200 NVL is not, and "10 or 14 power supplies must be selected". HPE announced the server in a press release of the same day.

HPE's AI servers page shows more names: a ProLiant DL525 Gen13 in 1U with "NVIDIA RTX PRO 4500 & 6000 Blackwell", announced on the same day, and a ProLiant DL394 Gen12 with NVIDIA Vera CPUs, for which the page announces up to two double-wide GPUs. Its FAQ states: "HPE Compute XD700 will support the latest NVIDIA Rubin GPUs." We name these models as announced and print no further specifications for them.

Warranty, service and what we need for a quote

For the DL380a Gen12 HPE states the warranty as "3-Year Parts, 3-Year Labor, 3-Year Onsite support with next business day response." For the XD690 the QuickSpecs say: "Hardware diagnostic support and repair are available for three years from the date of purchase." The service levels above that are HPE Tech Care and HPE Complete Care; the DL380a Gen12 QuickSpecs list, for example, the "HPE 3 Year Tech Care Essential DL380a Gen12 HW Service". Our article on GPU server warranty in the EU sets a maker's terms next to an integrator's. HPE marks its QuickSpecs "Worldwide"; apart from the two SKUs for China, Hong Kong and Macau, the documents read contain no limitation specific to the EU for the H200 NVL, the RTX PRO 6000 Blackwell Server Edition or HGX B300.

HPE's own documents decide the order lines, so a request names the model exactly as the QuickSpecs write it, the cards and their number, the power supplies for the voltage at your rack and the inlet temperature of the room. Before quoting we check the rack, power and airflow, and quantities and lead time are confirmed in writing with the quote.

FAQ