Buy a server with 8× NVIDIA RTX PRO 6000 Blackwell Server Edition in Europe

Who needs eight cards in one server

You can buy a server with eight NVIDIA RTX PRO 6000 Blackwell Server Edition cards from us, built to order and delivered anywhere in the EU. It has 768 GB of GPU memory in eight 96 GB pools, two AMD EPYC processors and 1,536 GB of DDR5 ECC memory in a 4U or 5U chassis, and it is the largest Server Edition configuration among our AI servers. Our team sizes the configuration to your workload and budget and replies within one business day.

Eight cards suit two kinds of buyer. The first serves many users from one large open model that needs four to eight cards, such as Qwen3-235B-A22B, DeepSeek-V4-Flash or Llama 4 Maverick. The second serves a whole company from one machine: several assistant copies, a coding model, a document model and the small models of a RAG pipeline on cards of their own, as our eight-card GPU allocation plans for 1,000 staff. Where the service must survive the loss of a host, two four-card servers are the safer layout, as one 8-GPU server or two 4-GPU servers compares. For eight 300 W cards, our comparison of Max-Q or Server Edition explains why we recommend the Server Edition in a rack. Training and full fine-tuning of large models belong on our training node with eight H200 NVL and NVLink bridges.

The eight-card build, component by component

The table shows what we build and what we change per project.

Our AI servers page; NVIDIA-Certified Systems configuration guide (30 September 2026) for cores, memory, NVMe, Redfish and TPM; AMD's EPYC 9555 page; Dell PowerEdge XE7745 technical guide for the power supplies, read on 24 October 2026.

1.5 TB of system memory and the processor count

NVIDIA's configuration guide for NVIDIA-Certified Systems, updated on 30 September 2026, asks for at least two sockets, six physical cores per GPU, system memory of at least twice the total GPU memory and one NVMe drive per socket. Eight 96 GB cards make 768 GB, so the guide asks for 1,536 GB and 48 cores. Two EPYC 9555, 64 cores and 360 W each, have 12 memory channels per socket, and 24 modules of 64 GB fill all 24 channels with one module per channel.

NVIDIA's RTX PRO AI Factory reference architecture of 18 May 2026 sets a lower memory floor, 128 GB per GPU, and asks for seven cores per GPU plus two per GPU for the operating system or virtualisation, 72 cores for eight cards, and two more cores per MIG instance. A server that splits all eight cards into 32 MIG instances would need 136 cores by that rule, so we size the processors to how the cards will be used. The memory also holds what the GPUs do not: model files in the page cache for fast restarts, a RAG index, and KV cache offloaded from the cards, as our guide to KV cache offloading explains.

4U or 5U, and which makers list eight cards

The card runs only in a server whose maker lists it for that model, slot, power cable and fan set. These eight-card configurations are in the makers' documents.

Manufacturer documents read on 24 October 2026. HPE DL380a Gen12: eight 2,400 or 3,200 W supplies for eight double-wide GPUs, C19 or C19-C20 cords only. Cisco C845A M8: up to four 3.2 kW supplies with N+1 redundancy.

A 4U chassis is the common size; a 5U chassis gives more room for air and network cards. The servers differ in how the cards reach the processors: Dell's XE7745 connects its GPU slots through four Broadcom PCIe switches, and Supermicro describes the SYS-422GA-NRT as a dual-root design with PCIe 5.0 x16 switches. That matters once a model is split over cards, because the Server Edition has no NVLink. Our list of RTX PRO 6000 Server Edition compatible servers covers cables and fan kits.

What 768 GB of GPU memory holds

By our users-per-card rule, 0.9 × 95.6 GiB − 3 GiB per card, eight cards hold 664.3 GiB of weights and KV cache. The table gives examples; our guide to what an 8× RTX PRO 6000 Server Edition server runs has every layout and the arithmetic.

Our estimates by the users-per-card rule: concurrent conversations at full context length, 16-bit KV cache unless stated. Qwen3-235B-A22B-2507 FP8: 219.8 GiB of weights and 5.9 GiB of cache per 32K conversation from its config.json; Llama 4 Maverick about 417 GB in FP8.

For a large model, two copies serve more requests at once than one. Helix reported on 14 August 2026 that on eight Server Edition cards, two TP=4 instances of DeepSeek-V4-Flash-0731 delivered 481.7 tokens per second at 32 concurrent requests with SGLang, against 236.9 for a single TP=4 instance. Our guide to RTX PRO 6000 without NVLink compares TP=8, two TP=4 copies and pipeline parallelism.

Published results for eight cards

MLPerf® Inference v6.1 of 16 September 2026 holds several eight-card Server Edition systems from different makers.

MLPerf® Inference: Datacenter v6.1, closed division, available category, offline scenario, Llama 2 70B (llama2-70b-99.9) and gpt-oss-120b, entries 6.1-0070, 6.1-0013, 6.1-0036 and 6.1-0050, retrieved from mlcommons.org on 24 October 2026. Results verified by MLCommons Association. Per card is our division of the system result, not an MLPerf metric. 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.

The four Llama 2 70B results lie within 2.6 per cent of each other, so the chassis matters less for throughput than for power, airflow and service. In the server scenario, where queries arrive at random intervals under latency limits, HPE's system reached 28,643.3 tokens per second on Llama 2 70B. Our RTX PRO 6000 benchmarks put these results next to the two- and four-card ones.

We supply this server with the layout planned per model; send us the models, context length and peak requests in flight, and our team replies with a configuration and the conversations per layout.

Power supplies, feeds and A/B for 6 to 8 kW

Eight cards at 600 W and two 360 W processors are rated at 5.52 kW together. We add 15 per cent for memory, drives, network cards, fans and conversion losses, an assumption that a measurement on the delivered system confirms or replaces. Dell's datasheet for its AI Factory with RTX PRO 6000 gives an estimated maximum of about 8 kW at full load for its eight-card XE7740 and XE7745 nodes and calls its estimates indicative.

Our estimate: board power from NVIDIA's product brief, 360 W per processor from AMD's EPYC 9555 page, plus 15 per cent; Dell AI Factory with NVIDIA RTX PRO 6000 datasheet, read on 24 October 2026.

Our arithmetic: 230 V × amps, and √3 × 400 V × amps for three-phase; 80 per cent is the planning margin of our guide to GPU rack power and cooling.

At 600 W the server is above the 5.89 kW we plan for a 32 A single-phase feed, so we plan a 400 V 16 A three-phase feed, which also covers Dell's 8 kW. At the 450 W cap, about 5.0 kW, a 32 A single-phase feed is enough by our estimate. With A and B feeds, each side has to carry the whole server, so the rack position needs two such feeds.

The power supplies follow the maker's rules. In Dell's XE7745, supplies 1 and 2 feed the processor zone and supplies 3 to 8 the GPU zone. For eight of these cards at 600 W, Dell's technical guide supports a 3+3 A/B layout in the GPU zone only with 3,200 W supplies, not with 2,400 W ones. The 3,200 W supply delivers its rating at 220.1 to 240 V AC and 2,900 W at 200 to 220 V, so European 230 V mains get the full output. Dell connects these supplies with C19 cords and states that the XE7745 "does not support user power cap". HPE fits eight 2,400 or 3,200 W supplies for eight cards in the DL380a Gen12, also with C19 cords only, and our guide to GPU server power on 16 A, 32 A and three-phase compares the feed types.

Cooling, inlet temperature and weight

All the power becomes heat. About 6.3 kW needs roughly 1,000 CFM of air at an 11 °C rise, and 8 kW about 1,300 CFM, by the 160 CFM per kW in our guide to GPU rack power and cooling. Dell's XE7745 guide lists a maximum ambient temperature of 30 °C with RTX PRO 6000 Server Edition cards in its thermal tables; Supermicro rates the SYS-422GA-NRT for eight 600 W cards at 35 °C. Our guide to GPU server inlet temperature explains what to measure at the rack.

Dell's datasheet notes that customers typically air-cool racks with a heat load up to 30 kW, which by our arithmetic is three such nodes at 8 kW each; denser racks need the measures in our guide to liquid cooling for GPU racks. The XE7745 weighs up to 68.5 kg with all drives, so plan the lift into the rack and the rack's load rating. We check rack, feed and airflow before we quote; tell us the feed type, A/B requirement, inlet temperature and free height units, and we choose the chassis and power supplies to match.

Software, drivers and NVIDIA licences

On request the server arrives with the operating system, NVIDIA driver, CUDA and a container runtime installed. The Server Edition needs a Linux driver from R575 and CUDA 12.9 or later on x86, by NVIDIA's product brief SP-12355-001_v02. As of October 2026, R580 is NVIDIA's long-term branch with support to June 2028, and NVIDIA's security bulletin of 30 September 2026 names 580.178.04 as the first fixed R580 driver, as our article on NVIDIA driver branches and CUDA versions explains.

MIG splits each card into up to four 24 GB instances, up to 32 per server, set per card, and NVIDIA's MIG guide states that NCCL is not supported with MIG, so split models run on whole cards. The Server Edition supports vGPU from release 19.0; compute vGPU profiles and NIM need NVIDIA AI Enterprise, licensed per GPU, which we supply on the same invoice as the hardware. Serving on bare metal with open-source engines needs no NVIDIA licence.

Ordering, testing and warranty through us

The order runs in the same steps for one server or several.

1. You describe the workload, the rack position and the timeline. Sizing is free of charge.
2. We propose the configuration and check that cards, chassis, power supplies and cooling fit your rack and feeds.
3. You receive a written quote within one business day, with price and warranty terms per component; quantities and lead time are confirmed in writing with it.
4. You order on one EU contract with an EU invoice from Eurokommerz in Vienna; NVIDIA licences are lines on the same invoice.
5. We assemble the server and test it 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 on request 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.

For your own acceptance test on delivery, use the checks in our GPU server acceptance testing guide; for a public tender, our GPU server tender specification helps to word the requirements.

What your enquiry should contain

• The models, quantisation and context length, or the departments and use cases the server will carry
• Peak requests in flight, or the number of staff and how they use the service
• Whether cards will be split by MIG or given to virtual machines, and the platform with its version
• Rack position: free height units, rack depth and load rating, feed type and current, A/B feeds, PDU outlets and inlet temperature
• Network uplinks, shared storage and whether the server joins a cluster
• Delivery site, invoicing entity and timeline

Where a model fits four cards, two four-card servers give failover that one eight-card server cannot; our page to buy a 4× RTX PRO 6000 Server Edition server covers that version. We also supply the RTX PRO 6000 cards on their own and NVIDIA AI Enterprise licences on the same EU contract.

FAQ