NVIDIA RTX PRO 4500 Blackwell Server Edition benchmarks: published results and limits

What has been measured, and by whom

The RTX PRO 4500 Blackwell Server Edition has one detailed published LLM benchmark: StorageReview's review of 14 October 2026, in which the card matched an NVIDIA L40S at short context, trailed it by 17 to 33 per cent with 8K prompts and delivered about three times the output of an L4. A second StorageReview test of 24 September 2026 compared it with the L4 in an edge server, and Exxact published one- to four-card figures that do not agree with their own text. For the workstation RTX PRO 4500, Puget Systems and Phoronix published rendering and llama.cpp results.

The two editions share the 10,496 CUDA cores and 32 GB of GDDR7 on a 256-bit bus. The Server Edition is passive, single slot, 165 W and 800 GB/s; the workstation card is active, dual slot, 200 W and 896 GB/s. Results from one edition carry over to the other only approximately, and our page on the RTX PRO 4500 or the RTX PRO 4500 Server Edition covers the choice between them.

The publishers' pages, read on 24 October 2026. StorageReview's RTX PRO 6000 comparison ran in a Dell Pro Max Tower T2.

vLLM on the Server Edition against L4, L40S and RTX PRO 6000 Workstation Edition

StorageReview ran the Server Edition, an L40S and an L4 in a Dell PowerEdge R770 with two Xeon 6787P processors, driver 595.71.05, NGC vLLM 26.08 with vLLM 0.27.1 and CUDA 13.4.1. In the R770 the cards ran at 165 W (4500), 350 W (L40S) and 72 W (L4). An RTX PRO 6000 Workstation Edition ran in a separate host, a Dell Pro Max Tower T2, at 600 W and at a verified 450 W cap. The requests had 512 input and 512 output tokens, at 1 to 32 concurrent requests.

StorageReview, NVIDIA RTX PRO 4500 Blackwell Server Edition review, 14 October 2026: output tokens per second at 32 concurrent requests, 512/512. RTX PRO 6000 Workstation Edition at 600 W: 2,283, 2,139, 4,155 and 616. SE: Server Edition; WS: Workstation Edition.

On the three Ministral models the 4500 and the L40S curves overlap from 1 to 32 concurrent requests, at less than half the board power. The 4500 reached 2.8 to 3.6 times the L4, and the RTX PRO 6000 at 450 W about twice the 4500. On the 27B model in NVFP4 the order changes: the L40S, which has no FP4 hardware, was ahead, and the RTX PRO 6000 at 450 W and 600 W landed within 1 per cent of each other. StorageReview calls NVFP4 on this card "a vLLM software story" that every GPU in its test shared, and at one request every card produced 17 to 22 tokens per second on that model. Our L40S vs RTX PRO 4500 Server Edition comparison sets these results next to the specifications.

Long context: where 32 GB runs out

With 8,192 input and 1,024 output tokens the picture changes. The 4500 on Ministral 3 14B in FP8 flattened at about 180 output tokens per second from 8 concurrent requests; the L40S reached 230 and held, and the RTX PRO 6000 kept climbing to 552 at 450 W and 608 at 600 W. Across the three models the 4500 trailed the L40S by 17 to 33 per cent at 32 requests. StorageReview's explanation: "The flattening is KV cache running out, and 32GB sets that limit."

Our users-per-card rule predicts the same. It takes 0.9 × the memory the driver reports less 3 GiB, about 25.7 GiB on this card. A 14B model in FP8, such as Qwen3-14B at 15.2 GiB, leaves about 10.5 GiB for cache, room for only a few requests with prompts of 8K. The rule and its derivation are in our guide to how many users one RTX PRO 6000 can serve. For RAG with long documents, a card with more memory per copy, or a model split over two cards, holds more requests at once. We supply servers with both the RTX PRO 4500 Server Edition and the RTX PRO 6000 Server Edition; describe your prompt lengths and models, and our team says which card holds your requests.

Arithmetic ceilings in the lab

StorageReview also ran maximum achievable matmul throughput (MAMF), which shows what the card's Tensor Cores deliver in practice.

StorageReview, 14 October 2026: maximum achievable matmul TFLOPS. SE: Server Edition; WS: Workstation Edition.

StorageReview puts the 4500 at 67 to 70 per cent of NVIDIA's dense peaks. Its NVFP4 result is 1.5 times the L40S's FP8 ceiling, which matters for prompt processing on long inputs once the serving software uses FP4 well.

Edge server test against the L4

On 24 September 2026, StorageReview put the Server Edition and an L4 into the same HPE ProLiant DL145 Gen11 with one AMD EPYC 8534P and 96 GB of memory, using Metrum AI Bench 3.9.2 with vLLM 0.24.0, nine prompt shapes and up to 256 concurrent requests.

StorageReview, RTX PRO 4500 edge review in the HPE ProLiant DL145 Gen11, 24 September 2026 (Spanish edition read on 24 October 2026): output tokens per second, BF16 models with a BF16 KV cache. SE: Server Edition.

StorageReview summarises the result as about 2.7 times the L4's output at practical concurrency and 3.0 to 3.2 times at 256 requests. For one user on Qwen3.5-4B, each token took 12.4 ms on the 4500 and 35.0 ms on the L4. The whole server drew a median of 301 W at 32 requests with the 4500 against 196 W with the L4, and the GPU averaged 84 °C in the longest runs. The 4500 also ran Gemma 4 26B-A4B in NVFP4, at 1,613 tokens per second at 32 requests, which the L4 cannot run natively. Our comparison of the L4 and the RTX PRO 4500 Server Edition covers slot, power and the form-factor change from a low-profile to a full-height card.

The workstation card: rendering, video and llama.cpp

Puget Systems' content creation roundup of 18 December 2025 tested the workstation RTX PRO 4500 Blackwell at 200 W and 896 GB/s, one card per system; Puget wrote that "we won’t be looking at multi-GPU setups in this article". Against the RTX 4500 Ada it reported gains of 73 per cent in V-Ray GPU, 43 per cent in Octane, 32 per cent in Topaz Video AI, 28 per cent in Unreal Engine and about 50 per cent in Blender. Our RTX PRO 4500 Blackwell 32 GB guide uses these rendering figures.

Phoronix tested the workstation RTX PRO 4500 and its siblings on Linux on 21 May 2026, with a Threadripper 9980X, Ubuntu 26.04, Linux 7.0 and driver 595.58.03, and ran llama.cpp on 8B, 7B and gpt-oss-20b models in Q8_0. The figures are in charts only; Phoronix's text says the 4500 came "close to the RTX 6000 Ada Generation performance" in some AI tests. Neither publisher tested the Server Edition, its MIG or its vGPU support.

Results to treat with care

Exxact's benchmark of 20 May 2026 is the only published source with two, three and four Server Edition cards, and it cannot be used for scaling. It names no host, driver or software versions. Its table gives 8.6 tokens per second for Llama 3.3 70B in Ollama on one card where the text says "~14.6 tok/s", and 6.0 on two cards where the text says "~20.4 tok/s". The text reports 6,810 tokens per second for two cards in NVFP4 where the table shows 4,870, and the rows of the four-card section are labelled as two cards. Several efficiency values do not match the table's own throughput divided by power, and single-stream Ollama results sit next to batched vLLM results, which Exxact itself calls "not directly comparable".

NVIDIA's product page states that the card "delivers over 5x the performance of the previous-generation L4 GPU" for small models, with no test conditions on the page; StorageReview's edge review notes that the claim depends on FP4 and NIM. NVIDIA's "up to 50x" vector-search and "up to 100x" video-captioning claims compare an "8x RTX PRO 4500 server" with dual AMD EPYC 9654 CPU-only systems, so they measure a GPU server against a server without GPUs.

StorageReview's review states that "a 2U server that takes four double-wide cards can take eight of these", without naming a server. Supermicro's listing for the card tops out at six in 2U and eight in 3U to 5U, so check the maker's configurator. In the same review, the RTX PRO 6000 that served as reference was a Workstation Edition in a Dell Pro Max Tower T2, while the specification table calls it "SE".

From one card to eight: what the numbers imply

No publisher has measured an eight-card RTX PRO 4500 Server Edition server. NVIDIA's own claims are footnoted against an "8x RTX PRO 4500 server", AWS offers eight of the cards per instance in its g7.48xlarge, and Supermicro's RTX PRO systems page lists eight in 3U, 4U and 5U models (the datasheets do not always name the card), so the configuration exists; its throughput has to be derived.

Our arithmetic from StorageReview's single-card results of 14 October 2026, output tokens per second at 32 concurrent requests per card; not measured on eight or four cards. RTX PRO 6000 values were measured on a Workstation Edition at 450 W.

The multiplication holds only for independent copies, one per card, with no traffic between cards and a host that keeps up with 256 concurrent requests. At short context, eight 4500s reach about what four RTX PRO 6000 cards at 450 W reach (measured on the Workstation Edition), at 1,320 W of board power against 1,800 W. At 8K the four larger cards are ahead by half again, because each copy has three times the memory, and more than three times as much is left for cache.

Split copies are another matter. By default, vLLM's custom all-reduce is disabled on more than two GPUs without NVLink; 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 no publisher has measured TP=2, TP=4 or TP=8 on this card. A model that needs four cards per copy, such as Llama 3.3 70B in FP8, should be planned by memory with our rule, and its speed per user confirmed in a test before it carries production traffic. Our guide to 8× RTX PRO 4500 Server Edition: models, users, VDI works out the layouts. We size eight-card servers from your model list and request shape; send us both, and our team replies with a configuration and the figures behind it.

Power and tokens per rated watt

StorageReview did not meter power in the October review: "the efficiency math available this time is against each card's tested power limit." On that basis the 4500 led every card on the 14B model at short context.

StorageReview, 14 October 2026: output tokens per second per rated watt at 32 concurrent requests. SE: Server Edition; WS: Workstation Edition.

The September edge review did meter the server: about 160 W with either card at light load, a median of 301 W with the 4500 at 32 requests, and about 1.9 times the L4's output tokens per watt at the wall. For an eight-card server, board power is 1,320 W, and by our estimate with two 280 W processors the server draws about 2.2 kW at the wall, inside the 2.94 kW we plan per 230 V 16 A feed. Our article on GPU energy per token explains how to turn such figures into energy per answer.

Which workloads suit the RTX PRO 4500 Server Edition

The published results point one way. The card is strong where a model fits comfortably in 32 GB and requests are short: chat and agent turns on models up to the 14B class, embeddings and rerankers, small vision-language models, and virtual desktops, where NVIDIA's vPC sizing guide calls it "the primary recommended GPU for NVIDIA vPC". It is weaker where context is long or one model needs several cards, because the cache runs out and splits go over PCIe. For those, the RTX PRO 6000 Server Edition holds a copy per card, as its benchmark round-up shows, and the L40S benchmarks cover the 48 GB alternative.

What we supply

We build AI servers to order with two to eight RTX PRO 4500 Server Edition cards, assembled and tested before delivery, on one EU contract and invoice with manufacturer warranty on every component. A measurement on the delivered system confirms the power figure for your configuration. We also supply the card alone from our professional GPU range, next to the L40S, L4 and RTX PRO 6000 Server Edition, with configuration and quote within one business day.

FAQ