Buy a rack server with 4 or 8 NVIDIA RTX PRO 5000 Blackwell cards in Europe

Who buys a rack server with the RTX PRO 5000

We build rack servers with four or eight NVIDIA RTX PRO 5000 Blackwell cards on chassis whose maker lists the card for that model. As of October 2026 that is a short list: six Supermicro models and one 2CRSi server, while most eight-card GPU servers list only passive Server Edition cards. You can buy this system from us: our team sizes the configuration to your workload and budget and replies within one business day.

The RTX PRO 5000 is a workstation card. NVIDIA gives it 48 or 72 GB of GDDR7 with ECC, 1,344 GB/s of memory bandwidth, 300 W, an active cooler and four DisplayPort outputs in a dual-slot board of 4.4 × 10.5 inches. Its product page, read on 24 October 2026, gives no count of cards per system and no server configuration. That leaves the chassis maker's GPU list as the document that decides where the card may go.

Three kinds of buyers ask for it in a rack. A development team that already works on RTX PRO 5000 workstations wants the same card in the server room, so that containers, drivers and MIG layouts move from the desk to the rack unchanged. A company whose models fit 72 GB per copy, such as a 4-bit 70B model or a 32B model in FP8, wants one model per card and a 300 W budget per slot. And a team that needs graphics virtual machines on Red Hat Enterprise Linux with KVM may choose the 72 GB card, because it is the only RTX PRO 5000 on NVIDIA's vGPU list. Most of these systems run bare-metal Linux for one team or a few.

Chassis whose makers list the RTX PRO 5000

We read the makers' datasheets on 24 October 2026. Each system below names the "NVIDIA RTX PRO 5000 Blackwell" in its GPU list; none of the datasheets gives a separate maximum for this card, so the count in the table is the chassis maximum for double-width GPUs.

Supermicro datasheets of the SYS-212H-TN, SYS-222H-TN, SYS-741GE-TNRT, SYS-551A-T, SYS-322GA-NR and AS -4125GS-TNRT2; 2CRSi Godì 1.8SP5-D product page; all read on 24 October 2026. Lenovo PSREF ThinkStation PX specification of 9 October 2026, as read for our workstation guide on 24 October 2026.

Two of the 2U entries need a close look at power. Their datasheets list two 1,200 W supplies in a 1+1 arrangement, and four RTX PRO 5000 cards alone draw 1,200 W, so by our arithmetic a fully loaded four-card build exceeds one supply and the pair would no longer be redundant. We confirm the supply options and the per-card count with the maker before we quote such a build. The SYS-741GE-TNRT has two 2,000 W supplies and the 5U SYS-551A-T, a full tower on rails, one 2,000 W supply for 200 to 264 V; we check the power budget for four cards on each in the same way.

For eight cards, 2CRSi, a French maker, lists the RTX PRO 5000 with 48 or 72 GB at 300 W for its Godì 1.8SP5-D, with eight double-width slots "for GPU TDP up to 350W", four PCIe 5.0 x16 slots per processor, 24 memory slots and 4+1 supplies of 2,000 W. Its operating temperature is 10 to 25 °C, against the 10 to 35 °C on every Supermicro datasheet above. Supermicro's 3U SYS-322GA-NR takes up to eight double-width cards with two Xeon 6900 processors; its 4U AS -4125GS-TNRT2 connects up to ten double-width cards to two EPYC processors through dual-root PCIe 5.0 switches.

Several widely used GPU servers do not list the card. Supermicro's 4U SYS-422GA-NRT names the H200 NVL, the RTX PRO 4500 Server Edition and the RTX PRO 6000 Max-Q and Server Edition; its SYS-422GL-NR and 5U SYS-522GA-NRT name only the two Server Editions, plus the H200 NVL on the first. Lenovo's ThinkSystem and ThinkAgile GPU summary, updated on 12 September 2026, has no RTX PRO 5000 option for any server, and HPE's ProLiant Compute DL380a Gen12 QuickSpecs, as read for our Max-Q rack guide, list the RTX PRO 6000 Server Edition, H200 NVL, L40S, L20 and L4 among their GPU options and no RTX PRO workstation card. We build RTX PRO 5000 rack systems only on chassis whose maker lists the card; name the chassis your data centre standardises on, and our team checks its GPU list, supplies and cables before quoting.

Why a workstation card needs the maker's listing

A GPU server pushes air from front to back through passive heat sinks, and its fans and ducts are designed for that. The RTX PRO 5000 brings its own blower, which, as our RTX PRO 5000 benchmarks article reports from the board partner's specification, exhausts through the slot bracket. In a dense riser the intake of each blower sits against the next card, so the slot layout, fan profile and power cables have to be qualified for it. The same reasoning applies to the RTX PRO 6000 Max-Q, a 300 W blower card that NVIDIA positions for workstations; our guide to the Max-Q in a rack server lists the checks, and they carry over to this card.

Every chassis in the table above also lists the RTX PRO 6000 Max-Q. Where both are listed, the Max-Q gives 96 GB at the same 300 W and 1,792 GB/s, so the RTX PRO 5000 has to earn its place on vGPU on Red Hat KVM, on a team that already standardises on it or on the price of the configuration. MIG is no argument for it: the Max-Q splits into two 48 GB or four 24 GB instances, the 72 GB card into two of 36 GB. Our page on buying a Max-Q rack server covers that card.

What the configuration contains

Makers' datasheets cited above; cores and memory by NVIDIA's configuration guide for NVIDIA-Certified Systems (30 September 2026): two sockets, at least six physical cores per GPU and system memory of at least twice the GPU memory. First figure for 48 GB cards, second for 72 GB.

On the Godì's 24 memory slots, 24 modules of 48 GB give exactly the 1,152 GB that eight 72 GB cards call for under NVIDIA's rule, one module per channel. Each card has one 16-pin power connector and needs a cable that matches the chassis, which we take from the maker's parts list. For storage we plan a mirrored system drive and a mirrored model store; for the network, 25 GbE serves chat traffic and 100 GbE or more pays off where models load from shared storage.

What four and eight cards run

Our users-per-card rule takes 90 per cent of the memory the driver reports, subtracts 3 GiB per card for the runtime and activations, and leaves the rest for weights and KV cache. We take the driver figure of a 72 GB card as 71.7 GiB and of a 48 GB card as 47.8 GiB, in proportion to the 95.6 GiB of a 96 GB card, which gives 61.5 and 40.0 GiB per card. The table counts conversations held at the same moment, with an FP8 KV cache: 1.25 GiB per 8,192-token conversation for Llama 3.3 70B, 1.0 GiB for Qwen3-32B and 0.625 GiB for Qwen3-14B; gpt-oss-120b is counted at 32K tokens with a 16-bit cache of 1.125 GiB.

Our estimates by the users-per-card rule, ECC off: 61.5 GiB usable per 72 GB card, 40.0 GiB per 48 GB card. Weights as in our model guides: Qwen3-14B FP8 15.2 GiB, Qwen3-32B FP8 32.0 GiB, Llama 3.3 70B NVFP4 39.8 GiB and FP8 67.7 GiB, gpt-oss-120b 60.8 GiB. Memory arithmetic, not measured throughput.

One 72 GB card holds a 4-bit 70B model with room for 17 conversations, but not the same model in FP8, which needs a pair. gpt-oss-120b loads on one 72 GB card with only 0.7 GiB to spare, so it runs over pairs. These are requests in flight: with the example values in our guide to a private ChatGPT server by company size, 2,000 employees produce about 80 at peak, which four 72 GB cards cover with Llama 3.3 70B in FP8. Our guide to what an 8× RTX PRO 5000 server runs adds larger models, MIG layouts and the limits of 576 GB.

No RTX PRO Blackwell card has NVLink, so a split model exchanges data over PCIe 5.0. vLLM keeps its custom all-reduce for TP=2 pairs and disables it by default on more than two PCIe-only GPUs, which is why the table prefers pairs; vLLM 0.30.0 (release notes, 22 September 2026) added an opt-in FlashInfer PCIe IPC all-reduce for systems without NVLink (#53576). We found no published measurement of tensor-parallel scaling on four or eight RTX PRO 5000 cards. Pairs belong on the same processor: on the Godì, four slots hang off each socket.

Power, feeds and inlet temperature

The table uses the method of our other server guides: the cards, the processors at their default TDP and 15 per cent for memory, drives, fans and power-supply losses. Feeds are planned at 80 per cent, 2.94 kW for a 230 V 16 A feed and 5.89 kW for 32 A.

Our estimate from NVIDIA's 300 W board power; AMD EPYC 9355 (280 W) and EPYC 9555 (360 W) default TDP, a 350 W workstation processor, and the 500 W ceiling Supermicro gives for the Xeon 6900 in the SYS-322GA-NR. A measurement on the delivered system replaces the 15 per cent with a measured value.

Four cards stay on one 16 A feed per side; eight cards need 32 A. For comparison, eight RTX PRO 6000 Server Edition cards at 600 W come to about 6.3 kW by the same method. Feed types and A/B layouts are in our guide to GPU server power on 16 A, 32 A and three-phase.

Inlet temperature decides between the eight-card chassis. A room that holds 27 °C at the rack front suits Supermicro's 35 °C rating and not the Godì's 25 °C. Measure at the rack position first, as our guide to GPU server inlet temperature describes.

RTX PRO 5000 or a Server Edition card in the rack

For most rack projects we recommend a Server Edition card. The table puts a four-card RTX PRO 5000 server next to our four-card inference node and an eight-card RTX PRO 4500 Server Edition server.

NVIDIA product pages and datasheets; NVIDIA vGPU supported-GPU list (2 October 2026) and MIG user guide (13 October 2026); Supermicro RTX PRO Blackwell Server Edition systems page; usable memory by our rule, as in our comparison of several smaller cards or fewer RTX PRO 6000.

The RTX PRO 6000 Server Edition is passive and listed by HPE, Lenovo, Cisco, Dell and Supermicro in two-, four- and eight-card servers; it is on NVIDIA's vGPU list for VMware, and one card holds Llama 3.3 70B in FP8. The RTX PRO 4500 Server Edition gives eight passive single-slot cards and vGPU on VMware at a similar power. The RTX PRO 5000 fits where the team already standardises on it, where 72 GB per copy is enough and the work runs on bare metal, or where graphics vGPU on Red Hat KVM is the goal. Our pages to buy a 4× RTX PRO 6000 Server Edition server and buy an 8× RTX PRO 4500 Server Edition server describe those builds, and our comparison of several RTX PRO 5000 or 4500 against fewer RTX PRO 6000 sets twelve builds side by side.

MIG, vGPU and software on delivery

NVIDIA's product page gives the RTX PRO 5000 up to two MIG instances. NVIDIA's MIG user guide lists the 48 GB card with two 24 GB instances; the 72 GB card is not in the guide yet, and its datasheet gives two instances of 36 GB, so eight 72 GB cards could offer 16 isolated GPUs of 36 GB. MIG needs Linux, driver 575.51.03 or later and the card switched to compute display mode with DisplayModeSelector, which turns its display outputs off. Only the 72 GB card is on NVIDIA's vGPU list, from vGPU 20.2 and only on Red Hat Enterprise Linux with KVM, with graphics vGPU types; our guide to MIG and vGPU: VMs per GPU compares the cards.

On request the server arrives with a Linux release the chassis maker supports, the NVIDIA driver branch you will run, CUDA and a container runtime with the NVIDIA Container Toolkit, with MIG set up if you plan it. The card has compute capability 12.0, and NVIDIA's datasheet lists CUDA 12.8. As of October 2026 the R580 branch has long-term support to June 2028, and NVIDIA's security bulletin of 30 September 2026 names 580.178.04 and 595.91.07 as fixed versions; our guide to NVIDIA driver branches and CUDA versions explains the choice.

How ordering works with us in the EU

1. You describe the workload, the chassis you prefer or the rack position. Sizing is free of charge.
2. We check the maker's GPU list for the exact model, the slot layout, cables, supplies and the inlet temperature, and say in advance if the RTX PRO 5000 does not fit.
3. You receive a written quote within one business day with price, lead time and warranty terms per component. Quantities and lead time are confirmed in writing with the quote.
4. You order on one EU contract with an EU invoice from Eurokommerz in Vienna.
5. The server is assembled and tested before delivery, with a power-on and basic load test.
6. We deliver anywhere in the EU, with commissioning on site on request; other destinations through our logistics partner.
7. Every component carries the manufacturer warranty, handled through us; DOA units are replaced, and the assembly is covered by us.

Our GPU server acceptance testing guide lists checks you can repeat on delivery.

What to send us for a quote

• The models, with quantisation and context length, and whether each card runs its own model or a model is split
• Peak requests in flight, and whether users share one service, MIG instances or virtual machines
• Four or eight cards, 48 or 72 GB, and the chassis your data centre prefers
• The rack position: feed type per side and inlet temperature
• Operating system or hypervisor
• Delivery site, invoicing entity and timeline

Write these points in the form below, and we reply with a configuration and a written quote within one business day.

FAQ