Buy a workstation with 2 to 4 NVIDIA RTX PRO 5000 Blackwell 48 GB cards in Europe

Who the RTX PRO 5000 48 GB workstation suits

We build workstations with two, three or four NVIDIA RTX PRO 5000 Blackwell 48 GB cards to order and deliver them anywhere in the EU. Each card has 48 GB of GDDR7 with ECC at 1,344 GB/s and 300 W, so four cards bring 192 GB of GPU memory into one tower at 1.2 kW of board power. You can buy this system from us: our team sizes the configuration to your workload and budget and replies within one business day.

The 48 GB card fits work where each model or each user needs up to about 40 GiB. A development team gets one card per developer, or two MIG instances of 24 GB per card. An engineering office can run CAD, rendering and a 14B or 32B assistant on the same machine. A data science team with several small models in production can keep each on its own card. For 70B models the 48 GB card is the tight choice: the weights alone fill it, so a 70B model always runs split over two or four cards.

Puget Systems tested the 48 GB card in its professional GPU roundup of 18 December 2025: about 50 per cent ahead of the RTX 5000 Ada in Blender, 35 per cent in V-Ray GPU, and 11 per cent ahead of the RTX 6000 Ada in V-Ray. The roundup covered single cards only and did not look at multi-GPU setups. Our RTX PRO 5000 benchmarks collect the published results for both memory versions.

The card and the three build sizes

NVIDIA's datasheet gives the RTX PRO 5000 Blackwell 14,080 CUDA cores, 1,344 GB/s, 300 W, three NVENC and three NVDEC engines and four DisplayPort 2.1b outputs, in an actively cooled dual-slot card of 4.4 × 10.5 inches with one 16-pin connector. The 72 GB version is identical apart from memory, as our comparison of the RTX PRO 5000 with 48 GB or 72 GB shows. Usable memory below follows our users-per-card rule: 90 per cent of the memory the driver reports, less 3 GiB per card. We take 47.8 GiB as the driver figure, in proportion to the 95.6 GiB of a 96 GB card.

NVIDIA RTX PRO 5000 Blackwell datasheet and MIG user guide (13 October 2026); usable memory by our estimate with ECC off. With ECC reserving 6.25 per cent, as NVIDIA's CUDA Best Practices Guide states for GDDR memory, usable memory would be 37.3 GiB per card; NVIDIA publishes no figure for GDDR7.

The cards have no NVLink, so a split model exchanges data over PCIe 5.0. Two cards suit a team with one or two models up to 32B; four cards suit eight developers on MIG, several models side by side, or one 70B model over all four.

Towers whose makers list four RTX PRO 5000 48 GB

Dell Precision 7960 configurator (Canada), Lenovo PSREF ThinkStation PX specification of 9 October 2026 and HP Z8 Fury G6i specifications, read 24 October 2026; BIZON X5500 G2 page, read by our research on 24 October 2026.

Dell offers the card in single, dual, triple and quad configurations of the Precision 7960, with 1,400 W or 2,200 W chassis. Two of its GPU slots are PCIe 5.0 x16 and two PCIe 4.0 x16, so two cards run at half the link bandwidth; for one model per card or per MIG instance that matters little. Lenovo's PX specification gives "up to four" for the RTX PRO 6000 Max-Q and no count for the RTX PRO 5000, so a PX with four RTX PRO 5000 needs Lenovo's confirmation first. HP's current Z8 Fury G6i page does not list the RTX PRO 5000.

We supply the cards for these towers and also build four-card workstations to order on Threadripper PRO or Xeon W platforms, as described on our AI servers page. Tell us which tower your company runs or which platform you prefer, and our team checks slots, power supply and cables before quoting.

Components and options

Our AI servers page (platforms and components); memory and cores by NVIDIA's configuration guide for NVIDIA-Certified Systems (30 September 2026), written for servers and applied here by the same arithmetic.

NVIDIA's guide asks for at least six physical cores per GPU and system memory of at least twice the GPU memory: 24 cores and 384 GB for four 48 GB cards, 12 cores and 192 GB for two. On an eight-channel platform such as Threadripper PRO, 384 GB is eight 48 GB modules and 192 GB eight 24 GB modules, one per channel. Four cards at x16 take 64 of the 128 PCIe 5.0 lanes that AMD lists for the Threadripper PRO 9975WX.

What two and four cards run

The counts follow our users-per-card rule: conversations of 8,192 tokens beside the weights in an FP8 KV cache, at 0.625 GiB per conversation for Qwen3-14B, 1.0 GiB for Qwen3-32B and 1.25 GiB for Llama 3.3 70B; gpt-oss-120b at 32K tokens with a 16-bit cache of 1.125 GiB.

Our estimates with 40.0 GiB usable per card and ECC off; weights 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, as in our model guides. Memory arithmetic, not measured throughput.

The table shows where 48 GB per card draws the line. Models up to 14B in FP8 run as one copy per card with plenty of cache. A 32B model in FP8 fits one card with room for only eight conversations, so a team of more than a handful of simultaneous users runs it over a pair. A 70B model in FP8 leaves almost nothing for cache on a pair, 9 conversations, and becomes practical only over all four cards. Four cards also hold Qwen3-32B as two copies on pairs for 96 conversations. The example values in our guide to a private ChatGPT server by company size put about 80 requests in flight at peak for 2,000 staff, so by memory this build covers a company of that size.

TP=4 has a cost the memory table does not show. By default vLLM's source disables its custom all-reduce on more than two PCIe-only GPUs, so all-reduce over four cards goes through NCCL; vLLM 0.30.0 (release notes, 22 September 2026) added an opt-in FlashInfer PCIe IPC all-reduce for systems without NVLink (#53576). vLLM's parallelism guide advises GPUs without NVLink to prefer pipeline parallelism for throughput. We found no published measurement of tensor-parallel scaling on four RTX PRO 5000. By our arithmetic one card is held to about 39 tokens per second for a single user of Qwen3-32B in FP8, which reads about 34.4 GB per token; that is a ceiling from the bandwidth, not a measurement. Our guide to four RTX PRO 5000 48 GB in one workstation and the article on RTX PRO 6000 without NVLink: TP=2, TP=4 or replicas explain the layouts.

Eight MIG instances for a development team

NVIDIA's MIG user guide, updated on 13 October 2026, lists the RTX PRO 5000 48 GB with the profiles 1g.24gb and 2g.48gb, each also as a +gfx variant with graphics support. Four cards therefore give eight isolated GPUs of 24 GB, each with its own memory and compute, or any mix of whole cards and halves. A 24 GB half holds gpt-oss-20b, or Qwen3-14B in FP8 with cache for about five conversations of 8K tokens by our rule, which suits one developer or one test instance.

MIG on this card needs Linux, driver 575.51.03 or later, a vBIOS of 98.02.73.00.00 or later and the display mode switched to compute with DisplayModeSelector 1.72 or later, which turns the card's display outputs off. Our RTX PRO 6000 MIG profiles describe the same procedure on the larger card.

The 48 GB card is not on NVIDIA's vGPU list. Of the RTX PRO 5000 cards only the 72 GB version is, from vGPU 20.2 and on Red Hat Enterprise Linux with KVM only. For a deskside vGPU host the RTX 6000 Ada, on the list from vGPU 15.2, is the 48 GB workstation card that NVIDIA supports; our page to buy an RTX 6000 Ada workstation covers it.

Power, heat and the circuit

Our estimate at full load from NVIDIA's 300 W board power and the 350 W TDP of the Threadripper PRO 9975WX, plus 15 per cent for memory, drives, fans and power-supply losses; a 16 A circuit planned at 80 per cent carries 2.94 kW.

Every build fits one 230 V 16 A circuit by our estimate, though a four-card tower should not share it with other large loads. At about 2 kW a four-card tower gives off some 6,800 BTU/h, which a closed office does not absorb for long; it belongs in a technical room or an air-conditioned space. The cards' blowers exhaust through the slot bracket, so the back of the tower needs free space. The cords matter too: a C13 appliance cord carries 10 A, and Dell's 2,200 W unit for the Precision 7960 takes a C19 cord. Our article on RTX PRO 6000 Max-Q power on 230 V works through the same 300 W class in detail.

Comparing the 48 GB build with 72 GB, RTX 6000 Ada and Max-Q

All four cards in this table draw 300 W and take two slots, so the same tower and power supply carry any of them.

NVIDIA datasheets, MIG user guide (13 October 2026) and vGPU supported GPU list (2 October 2026).

Against the RTX 6000 Ada, the 48 GB RTX PRO 5000 has the same memory and power with 40 per cent more bandwidth, PCIe 5.0, FP4 Tensor Cores and MIG. The Ada card keeps vGPU support and 18,176 CUDA cores against 14,080. Against the 72 GB version, the 48 GB card gives up the 4-bit 70B model per card and 36 GB MIG halves. Against the Max-Q, four 48 GB cards have half the memory and three quarters of the bandwidth per card at the same board power. When the workload fits 48 GB per copy, the choice between these comes down to price, which we quote for each configuration. Our pages to buy an RTX PRO 5000 72 GB workstation and to buy a 4× RTX PRO 6000 Max-Q workstation cover the larger builds, and several RTX PRO 5000 or 4500 against fewer RTX PRO 6000 compares twelve builds.

Software installed before delivery

On request we deliver the workstation with a Linux release or your company's image, the NVIDIA driver branch you run, CUDA and a container runtime with the NVIDIA Container Toolkit, and set up MIG instances where you plan them. The card has compute capability 12.0 and the datasheet lists CUDA 12.8, so images built only for older architectures need checking first. 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 covers pinning the branch.

From enquiry to delivery

1. You describe the workload: models, users, MIG plans, the room and the timeline. Sizing is free of charge.
2. We propose the configuration and check cards, tower, power supply and cooling against the room and the circuit.
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 workstation 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.

The checks in our GPU server acceptance testing guide work for a four-card tower too.

Information for the quote

• Models with quantisation and context length, and how many run at once
• Users at peak, and whether they share one service or get their own MIG instance
• Other work on the machine: CAD, rendering, simulation, data preparation
• The tower your company standardises on, if any, and the operating system
• Circuit rating, cooling and the place where the tower will stand
• Delivery site, invoicing entity and timeline

Send us these details through the form below, and we reply with a configuration and a written quote within one business day.

FAQ