Buy a server with 8× NVIDIA RTX PRO 4500 Blackwell Server Edition in Europe

Who buys an eight-card RTX PRO 4500 Server Edition server

You can buy a server with eight NVIDIA RTX PRO 4500 Blackwell Server Edition cards from us, built to order and delivered anywhere in the EU. It gives eight separate 32 GB GPUs, 16 MIG instances and 24 video decoders at 1,320 W of board power, for companies, public bodies and labs that run many workloads of moderate size side by side. Our team sizes the configuration to your workload and budget and replies within one business day.

Three kinds of buyer choose this server. IT teams that run virtual desktops take it as a vGPU host: NVIDIA's vPC sizing guide counts up to 128 desktops on eight cards. Platform teams use it for several models at once, such as an assistant, a coding model, embedding and reranking for RAG and speech-to-text, each on its own card or MIG half. Universities and development teams give each user or course an isolated GPU on it, 16 halves of 16 GB per server.

Video analytics is a fourth case. Each card has three NVDEC and three NVENC engines, so the server carries 24 decoders, and a detection model that fits 32 GB runs next to its own decoders on every card. How many camera streams one decoder takes depends on codec, resolution and frame rate, so that figure has to come from the analytics vendor for your streams; our guide to a GPU server for video analytics shows how to count them.

It is the wrong server for one large model with many users. Llama 3.3 70B in FP8 needs four of these cards per copy, while one RTX PRO 6000 Server Edition holds it alone; for that workload our page to buy a 4× RTX PRO 6000 Server Edition server is the better start. Training and full fine-tuning of large models belong on our training node with eight H200 NVL and NVLink bridges.

What is in the configuration

NVIDIA's configuration guide for NVIDIA-Certified Systems, updated on 30 September 2026, sets the floor: two sockets, six physical cores per GPU and system memory of at least twice the total GPU memory. For eight cards of 32 GB that is 48 cores and 512 GB.

NVIDIA-Certified Systems configuration guide (30 September 2026) for cores, memory, NVMe, Redfish and TPM; AMD EPYC 9355 page for 12 memory channels; Supermicro and Puget Systems listings for form factors; our AI servers page; all read on 24 October 2026.

The memory follows the processors' channels. Two AMD EPYC 9355 with 32 cores each give 64 cores and 24 memory channels, so 24 modules of 32 GB make 768 GB with one module per channel, above NVIDIA's 512 GB floor. The chassis decides the processor family: Supermicro's RTX PRO systems page lists eight cards for its 3U SYS-322GA-NR, 4U SYS-422GL-NR and SYS-422GL-FNR2 and 5U SYS-522GA-NRT, all with two Intel Xeon 6900 (the SYS-322GA-NR datasheet does not name the RTX PRO 4500, so for that chassis we have Supermicro confirm the card before we quote), and Puget Systems configures up to eight cards in a 4U server with two AMD EPYC 9005. Supermicro's largest 2U listing for the card is six, so we build eight-card servers in 3U to 5U chassis. Our guide to how many GPUs fit in one server explains why lanes, power and airflow set these limits.

Variants for VDI, several models or MIG

The hardware stays almost the same across the variants; software, licences and memory change.

• VDI host. vGPU 20 on VCF 9.0.1 or later, ESXi 8.0 Update 3g or later, RHEL KVM or Ubuntu, with vPC or vWS licences. NVIDIA's vPC sizing guide gives 16 desktops per card on 2 GB profiles and 10 on 3 GB profiles, as maxima. Extra system memory goes to the virtual machines' own RAM.
• Inference host. Bare metal or Kubernetes with a serving engine per model. By our users-per-card rule, each card holds about 25.7 GiB of weights and cache, so models up to about 20 GiB of weights run one copy per card and larger ones in TP=2 or TP=4 copies.
• Mixed host with MIG. Some cards split into two 16 GB halves for embedding models, rerankers or student environments, the rest whole for larger models. MIG is set per card.
• Power cap. NVIDIA's product brief allows a cap from 100 W to 165 W per card, which brings board power from 1,320 W down to 800 W where a rack position is tight.
• Start with four. A server can start with four cards in an eight-card chassis, with supplies, cables and risers chosen for eight; our page to buy a 4× RTX PRO 4500 Server Edition server covers the four-card version.

What eight cards run

By our users-per-card rule, 0.9 × the memory the driver reports less 3 GiB per card, the eight cards hold about 205.6 GiB of weights and cache. The table gives examples; our guide to 8× RTX PRO 4500 Server Edition: models, users, VDI has every layout, the derivation and the published measurements.

Our estimates by the users-per-card rule, conversations at full length with an FP8 KV cache; desktops from NVIDIA's vPC sizing guide (19 August 2026), a maximum density.

The conversations are requests in flight at the same moment. At the example values of our company-size guide, 128 of them cover about 2,000 staff, and 56 cover about 1,000. We supply the server in each of these variants; send us your models, desktop counts and peak requests in flight, and our team replies with a configuration and the capacity per workload.

Power and room requirements

By our estimate, eight cards at 165 W and two processors at 280 W draw about 2.2 kW at the wall, including 15 per cent for memory, drives, fans and losses. We plan a 230 V 16 A single-phase feed at 80 per cent, 2.94 kW, so one such feed per side carries the server, and with redundant supplies on A and B feeds each side carries the whole load. With the cards capped at 100 W, the estimate falls to about 1.6 kW.

• Cords and outlets. The number of cords follows the chassis' supply layout, which is often shared with configurations for 600 W cards. Puget Systems' 4U eight-GPU server, for example, uses 3+1 supplies of 3,200 W and asks for four 200-240 V connections on separate circuits; it offers C19-to-C20 cords. Check the C19 outlets on the PDU.
• Cooling. About 2.2 kW of heat needs roughly 350 CFM of air at an 11 °C rise. The card is passive and takes its air from the server's fans; the server maker sets the inlet limit for the chassis with this card.
• Room. A server with eight cards under load belongs in a server room or data centre, as our guide to a GPU server in an office explains.

We check rack, feed and airflow before we quote. Tell us your feed type, A/B requirement and inlet temperature, and we choose the chassis and power supplies to match.

Software delivered with the server

On request the server arrives with the operating system, NVIDIA driver, CUDA and a container runtime installed, so your team can start a serving engine or a hypervisor on day one.

Our AI servers page; driver branch from NVIDIA's product brief PB-12790-001_02 and NVIDIA's driver release notes of October 2026, as cited in our card guide.

NVIDIA's security bulletin of 30 September 2026 names 595.91.07 as the first fixed R595 driver. NVIDIA's vGPU documentation lists vGPU 20, on the R595 branch, with an end-of-life date of March 2027, so a VDI host moves to a later vGPU branch within its first year; our article on NVIDIA driver branches and CUDA versions explains the choice. Open-source engines on bare metal need no NVIDIA licence. NVIDIA AI Enterprise, licensed per GPU, is needed for NIM and other NVIDIA AI Enterprise software, so eight cards need eight licences, as our guide to NVIDIA AI Enterprise licensing sets out. Models, RAG and the platform on top are a separate service, Private AI/ML.

How an order with us runs

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

1. You describe the workload: models, desktops, users, data, rack and timeline. Sizing is free of charge.
2. We propose the configuration and check that cards, chassis, power supplies and cooling fit your rack and feed.
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; 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.

We also build on your own chassis or with parts you already own, after a compatibility check of the platform, power and cooling. For your own acceptance test on delivery, use the checks in our GPU server acceptance testing guide; to compare our quote with others line by line, our guide to GPU server quote comparison lists what each line should state.

What to include in a quote request

A short description is enough to start; these points let us quote the exact configuration in the first reply.

• Models you plan to run, with quantisation and context length, or the desktop profiles and seat counts per user type
• Peak requests in flight, or the number of users and how they work
• MIG plans: which workloads need an isolated half
• Software platform: bare metal, Kubernetes, vSphere or RHEL KVM, with the version
• Rack position: free height units, feed type and current, A/B feeds, PDU outlets and inlet temperature
• Network uplinks and where models, images and data are stored
• Delivery site, invoicing entity, timeline, and any tender requirements, which our GPU server tender specification helps to word

What we supply

We supply the server with eight NVIDIA RTX PRO 4500 Blackwell Server Edition cards in 3U, 4U or 5U chassis whose makers list eight cards, together with four-card versions and the cards from our professional GPU range for servers you already run. NVIDIA AI Enterprise and vGPU licences come on the same EU contract and invoice as the hardware.

FAQ