4× NVIDIA RTX PRO 4500 Blackwell Server Edition in 2U: models, users, VDI and power

128 GB on four single-slot cards

A 2U server with four NVIDIA RTX PRO 4500 Blackwell Server Edition cards runs small and mid-size models for a department: by our users-per-card rule each card holds a model of up to about 20 GiB of weights with room for its cache, and the server as a whole about 102.8 GiB of weights and cache. That is four copies of a 14B model for 64 conversations at once, one copy of a 70B model split over all four cards, eight isolated MIG instances or up to 64 virtual desktops, at about 1.4 kW for the whole server.

The card is the passive, single-slot data-centre version of the RTX PRO 4500. NVIDIA's product page gives 32 GB of GDDR7 on a 256-bit bus at 800 GB/s, 165 W, a PCI Express 5.0 x16 interface, one 16-pin power connector and three NVENC and three NVDEC engines. The four cards do not pool their memory. The card has no NVLink, so a model larger than one card is split over PCIe.

NVIDIA RTX PRO 4500 Blackwell Server Edition product page, read on 24 October 2026; the 100 W minimum from NVIDIA's product brief PB-12790-001_02, as cited in our card guide. Usable memory is our estimate.

The single slot is what puts four cards into 2U. On Supermicro's page, the 2U SYS-222C-TN takes four of these cards but only two RTX PRO 6000 Server Edition, and the SYS-222H-TN takes four of them and none of the larger card. The card itself is described in our guide to the RTX PRO 4500 Blackwell Server Edition; this page works out what four of them do together.

How much of 128 GB holds models and cache

Our users-per-card rule takes 90 per cent of the memory the driver reports and subtracts 3 GiB per card for CUDA graphs, activations and the runtime. NVIDIA publishes no driver-visible figure for the 32 GB card, so we scale the 97,887 MiB that a 96 GB RTX PRO 6000 reports to about 31.9 GiB, which gives 0.9 × 31.9 − 3 = 25.7 GiB per card and 102.8 GiB for four. Check the figure with nvidia-smi -q on your own card. NVIDIA's vPC sizing guide states that Blackwell's GDDR7 ECC is "always enabled without reducing usable VRAM". The derivation of the rule is in our guide to how many users one RTX PRO 6000 can serve.

What remains after the weights is KV cache. With an FP8 cache, one conversation of 8,192 tokens takes 0.5625 GiB on Qwen3-8B, 0.625 GiB on Qwen3-14B, 1.0 GiB on Qwen3-32B and 1.25 GiB on Llama 3.3 70B, by the layers, KV heads and head size in each model's configuration. A 16-bit cache doubles these.

Models and layouts on four 32 GB cards

Four cards give three ways to place a model: one copy per card, two TP=2 pairs or one TP=4 copy over all four. More cards per copy leave more room for conversations, and add traffic over PCIe on every layer.

Our estimates by the users-per-card rule, not measurements: concurrent conversations of 8,192 tokens at full length with an FP8 KV cache; gpt-oss-120b at 32,768 tokens with a 16-bit cache of 1.125 GiB. Weights are the Qwen, NVIDIA and OpenAI checkpoints as in our sizing guides.

Two rows show the arithmetic. One copy of Qwen3-14B on one card leaves 25.7 − 15.2 = 10.5 GiB, and 10.5 ÷ 0.625 gives 16 conversations. One TP=4 copy of Llama 3.3 70B in FP8 has 102.8 GiB; less 67.7 GiB of weights that leaves 35.1 GiB, and 35.1 ÷ 1.25 gives 28 conversations, or 14 with a 16-bit cache.

The table also shows where four cards stop. A TP=2 pair holds 51.4 GiB, which is not enough for Llama 3.3 70B in FP8 or for gpt-oss-120b, so both need all four cards as one copy, and the server then runs no other model. Models with several hundred GB of weights do not fit at all, by the sizes in our LLM hardware requirements by model.

Users per model, from a team to a department of 1,500

Conversations are requests in flight at the same moment, not named users. Our guide to a private ChatGPT server by company size uses about 20 requests in flight at the busiest moment for 500 staff and about 40 for 1,000; at the same ratio, 60 correspond to about 1,500 staff.

By those values, four copies of Qwen3-14B in FP8 with 64 conversations cover about 1,500 staff, and Qwen3-32B in FP8 as one TP=4 copy with 70 about 1,750. Llama 3.3 70B in NVFP4 as one TP=4 copy holds 50, a department of about 1,250. Llama 3.3 70B in FP8 holds 28, which suits about 700 staff, and leaves no card for anything else.

For most departments the more useful layout mixes models. An assistant model on two cards, a coding model on one, and an embedding model and a reranker for RAG in two MIG halves on the fourth give each engine its own memory, restart and update cycle. Our guide to serving several models on one GPU server covers the engine settings. We build this server to order and size the split per model; send us your models and the peak requests in flight, and our team replies with a layout and the conversations each model gets.

Splitting a model over PCIe on four cards

Every split copy exchanges partial results between its cards at every layer, and here that runs over PCIe. By default, vLLM's custom all-reduce is disabled on more than two GPUs without NVLink and falls back to NCCL, so a TP=2 pair keeps the faster path by default and a TP=4 copy does not. vLLM release 0.30.0 of 22 September 2026 added an opt-in FlashInfer PCIe IPC all-reduce for systems without NVLink (pull request #53576), and we found no published gain for it on this card. vLLM's parallelism guide of 15 October 2026 advises GPUs without NVLink to "leverage pipeline parallelism instead of tensor parallelism for higher throughput and lower communication overhead".

Bandwidth sets the ceiling for one conversation. By our estimate, a card at 800 GB/s that reads 70.6 GB of FP8 weights per token is limited to about 11 tokens per second, and an ideal TP=4 copy to about 45 before the cost of the all-reduce, which nobody has published for this card. A 14B model in FP8 on one card, at about 16 GB per token, has a ceiling near 49 tokens per second. In a two-socket server, nvidia-smi topo -m shows which cards share a processor; a pair under one socket avoids the link between the sockets. Our article on one model on several GPUs over PCIe or NVLink explains the options.

Published results for one card, and four copies

No publisher has measured four RTX PRO 4500 Server Edition cards in one server in a way we can use. StorageReview tested one card with vLLM 0.27.1 in a Dell PowerEdge R770 and published the results on 14 October 2026. On Ministral 3 14B Instruct in FP8 with 512 input and 512 output tokens, the card produced 1,107 output tokens per second at 32 concurrent requests, against 995 for an L40S and 364 for an L4. With 8,192 input tokens the curve flattened at about 180 tokens per second from 8 concurrent requests; StorageReview wrote that "The flattening is KV cache running out, and 32GB sets that limit".

By our arithmetic, four independent copies reach at most about 4 × 1,107 = 4,428 output tokens per second at short context and 4 × 180 = 720 at 8K, if the host keeps up. These are upper bounds from one card, not a four-card measurement. Exxact's one- to four-card tables of May 2026 contradict their own text, as our round-up of RTX PRO 4500 benchmarks shows, so they give no usable scaling figure.

Eight MIG halves

NVIDIA's MIG guide of 13 October 2026 lists the 32 GB RTX PRO 4500 with up to two instances: 1g.16gb, with half the memory and half the SMs each, or 2g.32gb for the whole card. Four cards therefore offer up to eight isolated GPUs of 16 GB, and MIG is set per card, so two cards can serve a TP=2 pair while the other two run in halves. NVIDIA's MIG guide states that NCCL is not supported with MIG, so a split copy runs on whole cards.

By our rule, a half that reports about 15.9 GiB holds 11.4 GiB of weights and cache. NVIDIA's FP8 checkpoint of Qwen3-8B, 8.8 GiB, fits with cache for about four conversations of 8K; Qwen3-14B in FP8 does not fit. Halves suit embedding models, rerankers, classifiers and isolated environments for developers or students, as in our guide to a GPU server for a university lab. With two plain halves per card, each half gets one decoder and one encoder, so eight halves use eight of the twelve decoders on the cards.

Desktops with vGPU 20

NVIDIA's vPC sizing guide, updated on 19 August 2026, calls the card "the primary recommended GPU for NVIDIA vPC". It gives 16 desktops per card on 2 GB profiles and 10 on 3 GB profiles and notes that 1 GB profiles are not supported on Blackwell. NVIDIA's RTX vWS sizing guide of the same date puts 6 to 8 light users per card on the DC-4Q profile and 3 to 6 medium users on DC-4Q or DC-8Q; for heavy users it lists only RTX PRO 6000 profiles.

NVIDIA vPC and RTX vWS sizing guides, both updated 19 August 2026, read on 24 October 2026; NVIDIA presents the vPC figures as maximum density. Four-card totals are our arithmetic.

The card is on NVIDIA's list of GPUs supported by vGPU software from release 20.0, and the vPC guide states that it supports MIG-backed vGPU. NVIDIA's vGPU documentation shows vGPU 20 on the R595 branch with an end-of-life date of March 2027, and the long-term vGPU 19 branch does not cover the card, so a host bought now moves to a later branch within its first year. On vSphere, NVIDIA lists the card for VCF 9.1, VCF 9.0 from 9.0.1 and ESXi 8.0 from Update 3g; RHEL KVM and Ubuntu are listed too. A four-card host suits a pilot or one department; our guide to VDI GPU sizing scales hosts to 1,000 desktops, and MIG and vGPU seats per card explains the licences.

Power on one 16 A feed

At 165 W per card, four cards draw 660 W. With two AMD EPYC 9355 at a default TDP of 280 W each, cards and processors are rated at 1,220 W. We add 15 per cent for memory, drives, network cards, fans and conversion losses, an assumption to be confirmed by a measurement on the delivered system.

Our estimate: board power from NVIDIA's product pages and brief, 280 W per processor from AMD's EPYC 9355 page, plus 15 per cent for the rest of the server. SE: Server Edition.

A 230 V single-phase feed at 16 A is rated at 3.68 kW, and we plan 80 per cent of it, 2.94 kW. The four-card server uses less than half of that by our estimate, so one 16 A feed per side carries it with redundant supplies on A and B feeds. Four RTX PRO 6000 Server Edition cards at 600 W do not fit one such feed; they hold three times the memory at almost four times the board power. Two four-card RTX PRO 4500 Server Edition servers on one feed would reach about 2.8 kW, too close to plan without measured figures. About 1.4 kW of heat needs roughly 220 CFM of air at an 11 °C rise, by the 160 CFM per kW in our guide to GPU rack power and cooling, and our guide to GPU server power on 16 A, 32 A or three-phase sets the feeds side by side.

2U servers that list four cards

Supermicro "RTX PRO Blackwell Server Edition" systems page, 2CRSi RTX PRO 4500 Blackwell page and Lenovo Press LP2391 (updated 28 July 2026), all read on 24 October 2026. SE: Server Edition.

The card runs only in a chassis whose maker lists it with the matching risers, fans and cables; Lenovo notes that the card ships without auxiliary power cables because they are server-specific. For any model, the maker's configurator decides. We check the chassis, feed and airflow before we quote; tell us the rack position and feed you have, and we name the 2U chassis that fits. The parts, variants and ordering steps are on our page to buy a 4× RTX PRO 4500 Server Edition server.

When another configuration fits better

Eight of these cards in a 3U to 5U chassis double everything on this page, to 128 desktops, 16 MIG halves and two TP=4 copies of a 70B model, as our page on 8× RTX PRO 4500 Server Edition shows. For one 70B-class model with many users, four RTX PRO 6000 Server Edition cards hold one copy per card, as our guide to 4× RTX PRO 6000 Server Edition works out. Where a model needs 48 GB on one card, our comparison of the L40S and the RTX PRO 4500 Server Edition weighs 48 GB at 350 W. RAG with long documents fills 32 GB with cache quickly, and the larger card holds more requests per copy. Training and full fine-tuning belong on H200 NVL with NVLink bridges.

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 build AI servers to order with four RTX PRO 4500 Server Edition cards in 2U chassis whose makers list them, assembled and tested before delivery, on one EU contract and invoice with manufacturer warranty on every component, and install the operating system, drivers, CUDA and a container runtime on request. vGPU and NVIDIA AI Enterprise licences come on the same invoice.

FAQ