Buy the NVIDIA RTX PRO 4500 Blackwell Server Edition: passive 32 GB card, price by quote

Price of the RTX PRO 4500 Server Edition by written quote

The NVIDIA RTX PRO 4500 Blackwell Server Edition is a passive single-slot PCIe card for rack servers, with 32 GB of GDDR7 memory and a board power of 165 W. We supply it to companies across the EU. A price is not printed here: we answer a request within one business day with a written quote that names the price, the lead time and the warranty terms for the number of cards you need.

This is the page for the card on its own. Complete machines are separate orders with their own pages: buy a 4× RTX PRO 4500 Server Edition server for a 2U system and buy an 8× RTX PRO 4500 Server Edition server for the larger chassis. The full sheet, with the servers that take the card, is in our RTX PRO 4500 Server Edition specs.

Who orders the card without a server

Two kinds of buyers ask for cards alone. System builders and integrators fit them into chassis they assemble and qualify themselves, and they order in the quantities of their production run. IT teams add cards to servers that are already in the rack, where the server maker lists the card for that model and free slots, power and the cable kit are at hand.

For both, the first check is the edition. NVIDIA also makes the NVIDIA RTX PRO 4500 Blackwell Workstation Edition, an actively cooled dual-slot card with 200 W and display outputs. The two have the 32 GB and the 10,496 CUDA cores in common, yet one cannot do the other's job: the Server Edition has no fan and stays without airflow in a tower. The workstation card is on our page to buy the RTX PRO 4500 Blackwell, with its figures in the 32 GB spec article.

We supply the RTX PRO 4500 Blackwell Server Edition across the EU; name the server model and the number of cards per chassis and we reply with a written quote within one business day.

RTX PRO 4500 Server Edition specs from the datasheet

NVIDIA datasheet 5192801 (April 2026) and NVIDIA's product brief PB-12790-001_02 (20 April 2026); vGPU supported-GPU list (2 October 2026) and product page (9 July 2026); "CUDA GPU Compute Capability" page (16 October 2026).

Three remarks belong to the table. NVIDIA publishes no AI TOPS figure for this card; the 1,617 AI TOPS of the Workstation Edition, an FP4 figure with sparsity, must not be carried over to it. The datasheet, the brief and the product page state no display output, so plan the card for compute and for graphics delivered remotely through vGPU. And the 800 GB/s are the Server Edition's own figure: the workstation card is specified at 896 GB/s.

Our sizing rule gives one card a budget of 25.7 GiB for model weights plus KV cache: 0.9 times the memory the driver reports, less 3 GiB. The reported memory is an assumption, scaled from the 95.6 GiB the driver shows on the 96 GB card; the driver on the delivered card decides.

What the server has to provide

Airflow. The card has no fan of its own. The product brief describes a bidirectional heat sink that accepts air from left to right or from right to left, and an ambient temperature of 10 °C to 45 °C. It gives no figure for the airflow required. The practical rule follows from that: fit the card only into a server whose maker lists it, because the maker has matched fans, ducts and inlet limits to it.

Slot and length. One full-height, full-length slot of single width for a PCI Express 5.0 x16 card; with the extender the card measures 312 mm. The board weighs 1,055 g, the bracket with screws 21 g and the extender with screws 48 g. Converted, 4.4 by 10.5 in is about 112 by 267 mm for the board itself.

Power. One 16-pin connector per card. The brief states that the card takes up to 165 W through it and describes a sense-pin setting in which the connector delivers at most 150 W and the rest comes through the slot. Take the 16-pin cable from the server maker's parts list for this card. Four cards add 660 W of board power to a server, eight add 1,320 W.

Power cap. The board power can be set between 100 W and 165 W, with nvidia-smi, where the setting is reset when the driver loads, or over SMBPBI, where it persists.

If the cards go into servers you already run, send us the chassis model, the free slots and the power supplies fitted, and we quote the cards for that configuration.

Drivers from R595, operating systems and CUDA

The driver decides whether the card is recognised at all. NVIDIA's product brief names R595 or later for Linux and for Windows. By our reading of NVIDIA's product lists, 595.58.03 is the first release of the R595 branch that names the card, and it is in the data-centre driver release notes of 595.112.07 / 597.30 (22 October 2026) and of 620.70.06 / 621.14 (4 November 2026). Do not plan it on R580: one release of that branch, 580.142, listed the card, but the current one, 580.196.07 of 21 October 2026, does not.

NVIDIA Data Center GPU Driver release notes of 22 October and 4 November 2026 and NVIDIA's table of supported drivers and CUDA Toolkit versions (4 November 2026).

Both rows are the data-centre driver release notes that list the card. On NVIDIA's download page it is found under the product type Data Center; the RTX Enterprise release notes for workstation cards do not list it. As operating systems the brief names Windows 10, Windows 11, Windows Server 2019 and Windows Server 2022, and the R620 release notes name Ubuntu 24.04, RHEL 9.8 and SLES 15.7 among the Linux distributions. For CUDA the brief asks for 13.0 U2 or later; CUDA 13.2 pairs with R595 and 13.5 with R620.

MIG halves and vGPU on one card

Of the two RTX PRO 4500 Blackwell cards, only the Server Edition is on NVIDIA's vGPU supported-GPU list of 2 October 2026. Its entry starts with release 20.0, with full support, and NVIDIA's documents for the card name RTX Virtual Workstation and vPC and give SR-IOV with 16 virtual functions. As verified on 10 November 2026, NVIDIA's vGPU index shows vGPU 20 on R595 (newest release 20.3, end of life March 2027) and vGPU 21 on R620, a long-term branch whose release 21.0 dates from November 2026, with end of life in November 2029; its documents list the card. vGPU 19 on R580 does not include it. NVIDIA's vGPU documentation also counts GPUs that support vGPU software as supported for pass-through.

For MIG the datasheet states up to two instances of 16 GB each. NVIDIA's MIG guide of 4 November 2026 has one row for an RTX PRO 4500 Blackwell with two instances and no row that names the Server Edition; with the brief's driver requirement, plan MIG on this card with R595 or later. The profiles are set out in our article on RTX PRO 4500 Server Edition vGPU profiles, and the whole range in MIG and vGPU on RTX PRO and Ada cards.

What has been measured, and under which conditions

StorageReview tested the card on 14 October 2026 in a Dell PowerEdge R770 with driver 595.71.05 and vLLM 0.27.1. At 32 concurrent requests of 512 input and 512 output tokens, Ministral 3 14B in FP8 produced 1,107 output tokens per second, against 995 on an L40S and 364 on an L4 in the same test. That level with the L40S holds for short requests of small models. With 8,192 input and 1,024 output tokens the same model reached about 180 tokens per second against 230 on the L40S, and on the short requests Qwen3.8-27B in NVFP4 came to 320 against 480 on the L40S.

A second StorageReview test, of 24 September 2026, used an HPE ProLiant DL145 Gen11 and vLLM 0.24.0. Qwen3.5-4B in BF16 with 256 input and 256 output tokens gave 529, 1,529 and 2,821 output tokens per second at 8, 32 and 256 concurrent requests, against 192, 532 and 935 on an L4, and the whole server drew a median of 301 W at 32 concurrent requests against 196 W with the L4. NVIDIA itself states over 5x the performance of the L4 in small-model inference, without publishing test conditions. Our round-up of RTX PRO 4500 Blackwell benchmarks and the comparison L4 vs RTX PRO 4500 Server Edition discuss these results.

Quote, contract and delivery

We can quote at once when a request names the number of cards (tell us how many you need), the edition, the server models they are for and the delivery address with the company to be invoiced. The written quote follows within one business day and states the price, the lead time and the warranty terms; quantities and lead time are confirmed in writing with the quote. The card is available to order.

The price depends on the quantity, the delivery address and the date of the order. We print no list price and no MSRP figure, and a card built into a server is a separate line of a system quote, as our page on servers built to order describes. The order runs on one European contract and is invoiced from Austria. Delivery goes anywhere in the EU and, through our logistics partner, to other destinations. The manufacturer warranty is handled through us, and a card that fails on arrival is replaced. The first consultation is free. For production quantities, see our page on NVIDIA RTX cards in volume for system builders.

When another card is the right order

A tower or a desk-side workstation needs the Workstation Edition, with its own fan and four DisplayPort outputs. Where a single-slot card with a fan is wanted, the NVIDIA RTX PRO 4000 Blackwell has 24 GB at 145 W in one slot, but it is on neither NVIDIA's vGPU list nor in its MIG guide; see the page to buy the RTX PRO 4000 Blackwell, its specs and our comparison of single-slot NVIDIA RTX cards.

If a model needs more than 32 GB on one card, a different passive card is the answer, since NVIDIA lists no Server Edition of the RTX PRO 5000 or 4000. Our article RTX PRO 5000 Server Edition: which cards exist lists passive cards we supply, and the choice between this card and the workstation card is the subject of RTX PRO 4500 or RTX PRO 4500 Server Edition. If you are weighing cards against a complete server, ask for both in one request; the quote then shows the cards alone and the server built to order.

FAQ