Buy a server with 4 to 8 NVIDIA L4 GPUs in Europe: video, small models and low power

Who buys a server with four to eight L4

You can buy a server with four to eight NVIDIA L4 cards from us, built to order and delivered anywhere in the EU. The L4 is NVIDIA's 72 W Ada Lovelace card for video, small AI models and light virtual desktops, so a server with eight of them draws less power for its GPUs than a single large card. Our team sizes the configuration to your workload and budget and replies within one business day.

NVIDIA's product page lists 24 GB of GDDR6, 300 GB/s of memory bandwidth, 30.3 TFLOPS of FP32 and 485 TFLOPS of FP8 with sparsity, half of that without. The card is single-slot and low-profile, passively cooled and powered from the PCIe slot, so it needs neither a GPU power cable nor a full-height slot. For server options NVIDIA names "Partner and NVIDIA-Certified Systems with 1–8 GPUs". Four kinds of buyer choose it.

• Video analytics and transcoding. Each card has four NVDEC decoders, two NVENC encoders and four JPEG decoders, with AV1. In NVIDIA's DeepStream 7.1 documentation one L4 handled 68 H.264 or 81 H.265 camera streams at 1080p and 30 frames per second with a ResNet18 detector, a tracker and two classifiers.
• Small language models and embeddings. Models of 8B parameters in FP8 fit one card with room for conversations, and a 14B model fits with a few. Classification, extraction, routing, embeddings and rerankers for RAG are typical work.
• Light virtual desktops. NVIDIA's vGPU types reference allows up to 24 desktops per card on the 1 GB L4-1B profile and 12 on L4-2B.
• Racks with little power. Branch sites, edge rooms and full racks often have one small feed. Eight L4 draw 576 W at their rated power.

The L4 is not the card for 70B models serving a team, for training, or for hardware isolation between tenants. Our comparison of the L4 and the L40S covers the 48 GB card, and the L4 and the RTX PRO 4500 Server Edition the Blackwell card that NVIDIA says delivers "over 5x the performance of the previous-generation L4 GPU", a claim its page states without test conditions.

Configuration of a four-card and an eight-card L4 server

Most L4 servers we quote fall into two sizes. The table shows what goes into each and where we adapt it to the project.

Our AI servers page; NVIDIA-Certified Systems configuration guide (30 September 2026) for sockets, cores per GPU, memory, Redfish and TPM; Lenovo Press LP1717 for the card counts per chassis, read on 24 October 2026.

NVIDIA's configuration guide for NVIDIA-Certified Systems asks for two sockets, at least six physical cores per GPU and system memory of at least twice the GPU memory. For four L4 that is 24 cores and 192 GB, for eight 48 cores and 384 GB. A video server often needs more cores than that minimum, because the processors receive and parse the camera streams. In NVIDIA's DeepStream run, 81 H.265 streams on one L4 kept an AMD EPYC 7763 at 46.1 per cent, against 8.06 per cent for 68 H.264 streams, so the codec changes the processor load a great deal. We size the processors from your codecs and stream counts, as our guide to EPYC or Xeon for a GPU server explains, and our guide to how much RAM a GPU server needs covers memory beyond the rule.

The card measures 69 by 169 mm, in NVIDIA's form factor 5.5 according to Lenovo, and runs on PCIe 4.0 x16, also in a PCIe 5.0 slot. It has no display outputs. Lenovo states that the L4 supports neither MIG nor NVLink, and that it offers SR-IOV with up to 256 virtual functions for vGPU.

Chassis that list four to eight L4

A passive card depends on the server's fans, so we build only into chassis whose maker lists the L4 for that slot and count. These are listings we read.

Lenovo Press LP1717, server support tables 3 and 4; Dell PowerEdge XE7745 technical guide (Rev. A09, September 2026), table 23; both read on 24 October 2026.

Lenovo's tables also show limits that matter: its 1U SR630 V3 takes three L4, and several entry towers and 1U servers take none. Dell lists the L4 in the XE7745 with sixteen single-width Gen4 x16 slots at 75 W each, and its restriction table allows the card up to 35 °C inlet temperature in the listed configuration. Lenovo's ThinkEdge SE455 V3 is 438 mm deep, as our guide to small language models on the L4 notes, which suits branch racks with little depth. We check the maker's listing for the card count, the processors and the riser layout before we quote.

What four or eight L4 run

We size with our users-per-card rule: 90 per cent of the memory the driver reports, less 3 GiB per card for the runtime, less the weights, leaves the KV cache for conversations. Scaled from the 97,887 MiB a 96 GB card reports, a 24 GB card shows about 23.9 GiB. NVIDIA's CUDA Best Practices Guide states that ECC on GDDR memory takes 6.25 per cent of the capacity, so one L4 keeps about 17.2 GiB with ECC on and 18.5 GiB with ECC off. The table uses ECC on.

Our estimates by the users-per-card rule, not measurements: concurrent conversations of 8,192 tokens with an FP8 KV cache, ECC on. Desktops are NVIDIA's maximum vGPUs per card in equal-size mode from its vGPU types reference (29 September 2026); engines from NVIDIA's L4 product page.

The conversations are requests in flight at the same moment. With the example values of our guide to a private ChatGPT server by company size, 500 staff produce about 20 of them at the busiest moment and 2,000 staff about 80. Four copies of an 8B model therefore serve a company of about 1,000 people, and eight copies one of 2,000 with headroom. One answer is limited by the card's 300 GB/s: Qwen3-8B in FP8 reads about 9.4 GB per token, a ceiling of about 32 tokens per second for one stream by our arithmetic.

Qwen's FP8 releases scale their weights in blocks of 128 by 128, a format TensorRT-LLM's support matrix does not list for Ada, as our L40S benchmark review notes, so confirm that your serving engine runs the checkpoint on this card. Ada has no FP4 arithmetic, and 4-bit models usually run as INT4 through AWQ or GPTQ. Our page on what an 8× L4 server runs has the full model list, the camera counts and the published results.

We supply both sizes with the layout planned per model and per camera pipeline; send us your models, cameras and desktop counts, and our team replies with a configuration and the users or streams per layout.

Power and room for an L4 server

The L4 changes the power picture of a GPU server: the processors draw as much as the cards or more. By our estimate, with two 200 W AMD EPYC 9135 for four cards and two 280 W EPYC 9355 for eight, plus 15 per cent for memory, drives, fans and conversion losses, the figures are these. A measurement on the delivered system confirms the estimate.

Our estimate: 72 W per card from NVIDIA's L4 product page, AMD's default TDP of 200 W for the EPYC 9135 and 280 W for the EPYC 9355, plus 15 per cent for the rest of the server.

We plan a 230 V 16 A single-phase feed at 80 per cent of its 3.68 kW, so 2.94 kW. All three servers fit one such feed, and with redundant power supplies on an A and a B feed each side carries the whole server. Two eight-card L4 servers draw less than one server with four RTX PRO 6000 Server Edition cards at about 3.4 kW. Our guide to GPU server power on 16 A, 32 A and three-phase feeds covers larger feeds.

By the 160 CFM per kW in our guide to GPU rack power and cooling, four cards need roughly 130 CFM of air at an 11 °C rise and eight about 210 CFM. Because the card has no fan of its own, the server's fan profile has to match it: Dell's limit of 35 °C applies to the listed XE7745 configuration, and other chassis have their own limits. A 1U or 2U server with eight passive cards runs its fans fast under load, so it belongs in a server room or a closed rack.

Software, drivers and vGPU on the L4

On request the server arrives with the operating system, the NVIDIA driver, CUDA and a container runtime installed. We install a supported driver branch: R580, NVIDIA's long-term branch with support to June 2028, or R595. NVIDIA's security bulletin of 30 September 2026 names 580.178.04 and 595.91.07 as the first fixed drivers. Our article on NVIDIA driver branches and CUDA versions explains the choice. For video, DeepStream and the Video Codec SDK run on the same driver.

The L4 has been on NVIDIA's vGPU list since release 15.2, and NVIDIA's vGPU lifecycle page, updated on 2 October 2026, lists it as fully supported. vGPU 19 on the R580 branch is supported until July 2028. B profiles need a vPC or vWS licence, Q profiles vWS and C profiles NVIDIA AI Enterprise, which the L4 does not include; open-source serving engines on bare metal need no NVIDIA licence. Our guide to L4 vGPU profiles lists every profile with its limits. Licences come on the same invoice as the hardware.

Ordering an L4 server from us

The steps are the same for one server or a series for many sites.

1. You describe the workload: cameras and codecs, models, desktops, sites and racks. Sizing is free of charge.
2. We propose the configuration and check that cards, chassis, processors 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; vGPU and NVIDIA AI Enterprise licences are lines on the same invoice.
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 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 add L4 cards to servers you already run, after a compatibility check of the platform, slots and cooling. For several branch sites, one configuration repeated per site keeps spares and images the same. Our guide to GPU server acceptance testing has checks for your own acceptance test on delivery.

Details that let us quote in the first reply

A short description is enough to start. With these points we can quote the exact configuration straight away.

• Cameras: count, codec, resolution, frame rate and the frames per second the model analyses
• Models with quantisation and context length, and the peak requests in flight
• Desktops per user type, the hypervisor and its version, and the vGPU licence edition you run
• Rack position: free height units, depth, feed type and current, A/B feeds and inlet temperature
• Existing servers that should take L4 cards, with model and generation
• Delivery sites, invoicing entity, timeline and any tender requirements, for which our GPU server tender specification helps

Write these points into the form below, and our team replies with a configuration and a written quote within one business day.

When another card fits better

When a model needs more than 24 GB per card, or one answer must be faster than 300 GB/s allows, the RTX PRO 4500 Server Edition gives 32 GB at 800 GB/s, MIG and FP4 at 165 W in a full-height single slot with a 16-pin cable; our page to buy an 8× RTX PRO 4500 Server Edition server covers it. For 32B models per card, graphics and VDI with heavier profiles, the L40S server has 48 GB at 350 W. For a 70B model for many users, four RTX PRO 6000 Server Edition cards hold one copy per card, as our page to buy a 4× RTX PRO 6000 Server Edition server shows.

What we supply

We supply servers with four to eight NVIDIA L4 cards, or up to sixteen in a 4U chassis that lists them, built to order, assembled and tested before delivery, and the L4 as single cards for servers you already run, on one EU contract and invoice with manufacturer warranty. NVIDIA vGPU and NVIDIA AI Enterprise licences come on the same invoice.

FAQ