Single-slot GPU for AI: RTX PRO 4000 Blackwell or RTX PRO 4500 Blackwell Server Edition

Two single-slot cards, built for different places

The NVIDIA RTX PRO 4000 Blackwell and the NVIDIA RTX PRO 4500 Blackwell Server Edition are the two air-cooled, full-height single-slot cards of the current Blackwell range. Choose the RTX PRO 4000 for a workstation with four free slots and people at the desk; choose the Server Edition for a rack server, where it gives 8 GB more memory per slot, MIG and vGPU. The two do not replace each other, because one is cooled by its own fan and the other by the server.

NVIDIA RTX PRO 4000 Blackwell datasheet (June 2026) and RTX PRO 4500 Blackwell Server Edition product page, NVIDIA's desktop GPU overview and vGPU supported GPU list (updated 2 October 2026), all read on 24 October 2026; 100 W minimum from NVIDIA's product brief PB-12790-001_02 as cited in our card guide; builds from Dell and Supermicro listings. FHFL: full height, full length.

Which NVIDIA cards take a single slot

We checked the slot width in NVIDIA's own documents rather than in shop listings. NVIDIA's overview of professional desktop GPUs, read on 24 October 2026, lists the RTX PRO 4000 Blackwell as "single slot" and every other Blackwell workstation card as dual slot, including the workstation RTX PRO 4500 at 200 W. The Server Edition's product page states "Single-slot, FHFL (4.4" H x 10.5" L)" and "Passive".

NVIDIA professional desktop GPU overview, RTX PRO 4500 Server Edition and L4 product pages, read on 24 October 2026; L4 cooling from Lenovo's ThinkSystem GPU summary (12 September 2026).

Two points in this table catch buyers out. The workstation RTX PRO 4500 shares the chip and the 32 GB with the Server Edition but needs two slots, as our comparison of the RTX PRO 4500 and its Server Edition explains. And the 70 W low-profile Blackwell cards are dual slot; the low-profile single-slot card is the L4, an Ada data-centre card with 300 GB/s, compared in our article on the L4 and the RTX PRO 4500 Server Edition. The RTX 4000 Ada is the previous single-slot generation; our comparison of the RTX PRO 4000 Blackwell and the RTX 4000 Ada covers the upgrade.

Active or passive: the cooling decides the chassis

The RTX PRO 4000 carries its own fan and draws air from inside a tower, which is what a workstation case provides. Four of them in adjacent slots work only where the case moves enough air past the cards; Dell's Precision 7960 configurator offers the card as a quad option, and our page on four RTX PRO 4000 in one workstation lists the towers.

The Server Edition has no fan. Lenovo describes it as "requiring system airflow to operate", and NVIDIA's product brief gives an ambient range of 10 to 45 °C with airflow across the heat sink. It runs only in a server whose fans and baffles are designed for it, and the maker's configurator sets the count and the inlet temperature. A tower case or a general-purpose server without GPU airflow is not the place for it.

The RTX PRO 4000 in a rack server is the less common case. Lenovo's ThinkSystem GPU summary, updated on 12 September 2026, lists the RTX PRO 4500 Server Edition, the L4 and the active RTX 4000 Ada as single-width server options, and lists no RTX PRO 4000 Blackwell. We put the RTX PRO 4000 into a server only where the server maker lists it. We build workstations with up to four RTX PRO 4000 and servers with the Server Edition; tell us the tower or server you plan to use, and we check its cooling for the card before we quote.

Memory, bandwidth and models per card

By our users-per-card rule, 90 per cent of the memory the driver reports, less 3 GiB per card, holds weights and KV cache. We scale the 95.6 GiB that a 96 GB card reports by capacity, which gives 18.5 GiB for the RTX PRO 4000 and 25.7 GiB for the Server Edition; check the real figure with nvidia-smi -q. The rule is derived in our guide to how many users one RTX PRO 6000 can serve.

Our estimates by the users-per-card rule: concurrent conversations of 8,192 tokens with an FP8 KV cache of 0.5625, 0.625 and 1.0 GiB per conversation; memory-bound, not measurements.

The extra 8 GB is what separates the cards for language models. A 14B model in FP8 fits both, but the Server Edition keeps about three times the room for conversations. Bandwidth adds less: 800 against 672 GB/s puts the ceiling for one conversation of a 14B FP8 model near 49 against 41 tokens per second by our arithmetic.

Published vLLM results exist for the Server Edition only. StorageReview tested one card with vLLM 0.27.1 and published on 14 October 2026: 1,107 output tokens per second on Ministral 3 14B in FP8 at 32 concurrent requests, level with an L40S at 995, and a plateau of about 180 tokens per second with 8,192 input tokens, which StorageReview attributed to KV cache running out at 32 GB. We found no vLLM measurement of the RTX PRO 4000 from a publisher with a stated method. Our round-up of RTX PRO 4500 benchmarks lists the Server Edition results.

Four and eight cards: what the builds hold

Single slot pays off in card count. Four RTX PRO 4000 take the four slot widths that two dual-slot cards occupy, and a 2U server takes four Server Edition cards, a 3U to 5U server eight.

Our estimates by the users-per-card rule, concurrent conversations of 8K with an FP8 KV cache, as on our pages on 4× RTX PRO 4500 Server Edition in 2U and 8× RTX PRO 4500 Server Edition. SE: Server Edition.

With the example values of our guide to a private ChatGPT server by company size, about 20 requests in flight at peak for 500 staff and 40 for 1,000, four RTX PRO 4000 serve a company of about 500 with Qwen3-14B, and four Server Edition cards about 1,500. Splitting a model over four cards runs over PCIe, because neither card has NVLink. By default vLLM keeps its custom all-reduce for two PCIe-only GPUs and uses NCCL for four; the opt-in PCIe all-reduce of release 0.30.0 (pull request #53576) has no published result on these cards, as our article on TP=2, TP=4 or replicas without NVLink explains.

Supermicro's page for RTX PRO Blackwell Server Edition systems lists four Server Edition cards in its 2U SYS-222H-TN and SYS-222C-TN and eight in the 3U SYS-322GA-NR, the 4U SYS-422GL-NR and the 5U SYS-522GA-NRT; 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. We build servers in chassis whose makers list the card and workstations with the RTX PRO 4000; send us the free slots, the chassis and the models you plan, and our team replies with the card and the count that fit.

MIG, vGPU and video engines

Here the cards differ most. The Server Edition's product page states "Up to 2 MIG @ 16 GB", so one card becomes two isolated GPUs with half the memory each, and a four-card server offers eight. NVIDIA's MIG user guide does not list the RTX PRO 4000; it is always used as a whole card.

For virtual desktops only the Server Edition qualifies. NVIDIA's list of GPUs supported by vGPU, updated on 2 October 2026, includes it from release 20.0 and does not include the RTX PRO 4000. NVIDIA's vPC sizing guide gives 16 desktops per Server Edition card on 2 GB profiles, and NVIDIA's vGPU documentation shows the vGPU 20 branch ending in March 2027, so a host bought now moves to a later branch in its first year; our guide to VDI GPU sizing scales hosts by seat count.

For video, the Server Edition has three NVENC and three NVDEC engines per card, the RTX PRO 4000 two of each. Eight Server Edition cards carry 24 decoders for a video analytics server, as in our guide to a GPU server for video analytics. The RTX PRO 4000 has the outputs: four DisplayPort 2.1b per card, up to four 4K displays at 165 Hz each.

Power per build and per circuit

Our estimate at full load: NVIDIA board power, one 350 W workstation processor for the tower, two 280 W AMD EPYC 9355 for the servers, plus 15 per cent for memory, drives, fans and power-supply losses. SE: Server Edition.

All four builds stay inside the 2.94 kW we plan per 230 V 16 A circuit at 80 per cent. Per watt the two cards are close: 4.6 GB/s of bandwidth per watt for the RTX PRO 4000 and 4.8 for the Server Edition, and 0.17 against 0.19 GB of memory per watt, by our arithmetic. The Server Edition can be capped down to 100 W per card where a feed or a branch office limits power, at a cost in speed that nobody has published.

Drivers and software

Both cards are Blackwell parts with compute capability 12.0 and need NVIDIA's open GPU kernel modules on Linux. The Server Edition's product brief names R595 or later drivers; NVIDIA's data-centre driver 595.91.07 lists it, while the long-term R580 branch does not, as our article on mixing Ada and Blackwell GPUs records. The RTX PRO 4000 is listed in both branches. For NVIDIA AI Enterprise support the Server Edition has to run in a server on NVIDIA's list of NVIDIA-Certified Systems.

When single slot is the wrong choice

Single-slot cards trade memory per card for density. For a 70B model in FP8, gpt-oss-120b or long documents with many users, one 96 GB card holds a whole copy where four single-slot cards hold it split and short of cache: by our rule four Server Edition cards give Llama 3.3 70B in FP8 28 conversations of 8K as one TP=4 copy, and one RTX PRO 6000 Server Edition gives it 12 without any split. In a rack the next steps are the L40S with 48 GB at 350 W, compared in our article on the L40S and the RTX PRO 4500 Server Edition, and the RTX PRO 6000 Server Edition; in a tower, the RTX PRO 6000 Max-Q with 96 GB at 300 W in a dual slot. Training and full fine-tuning belong on H200 NVL with NVLink bridges.

Which card for which use

• Developers and analysts at a workstation, four GPUs in one tower: RTX PRO 4000.
• CAD, rendering or video editing beside small models: RTX PRO 4000, for its display outputs.
• A shared AI service for a department, in a rack: RTX PRO 4500 Server Edition, four in 2U or eight in a larger chassis.
• Virtual desktops or virtual workstations: RTX PRO 4500 Server Edition with vGPU 20 or later.
• Isolated GPUs for students or teams: RTX PRO 4500 Server Edition with two 16 GB MIG halves per card.
• Video analytics: RTX PRO 4500 Server Edition, three decoders per card.
• Low-profile slots in an edge server: neither; the L4.

We supply both cards and build the systems around them; describe the workload and where the hardware will stand, and we propose the card, the chassis and the count.

What we supply

You can buy this system from us: our team sizes the configuration to your workload and budget and replies within one business day. We supply the RTX PRO 4000 Blackwell and the RTX PRO 4500 Blackwell Server Edition as cards, build workstations with up to four RTX PRO 4000 and AI servers to order with four or eight Server Edition cards in chassis whose makers list them, assembled and tested before delivery, on one EU contract and invoice with manufacturer warranty on every component. vGPU and NVIDIA AI Enterprise licences come on the same invoice.

FAQ