RTX PRO 4500 Blackwell Server Edition: specs, MIG, vGPU, power and the servers that take it

RTX PRO 4500 Server Edition specs in short

The RTX PRO 4500 Blackwell Server Edition is the passive data-centre version of NVIDIA’s 32 GB RTX PRO 4500. It keeps the 10,496 CUDA cores, the 32 GB of GDDR7 on a 256-bit bus and the 51 TFLOPS of FP32 of the workstation card, and NVIDIA specifies it at 165 W on a single-slot, full-height, full-length board with 800 GB/s of memory bandwidth and three video encoders and decoders. It splits into up to two 16 GB MIG instances, and NVIDIA’s vGPU software supports it from release 20.0.

NVIDIA’s product page lists 1.6 PFLOPS of FP4 and 811 TFLOPS of FP8 Tensor throughput, 82 RT Cores, a PCI Express 5.0 x16 interface and one PCIe CEM5 16-pin power connector, and calls the card confidential compute capable. The single slot and the 165 W matter for a mid-size server, because they let one chassis carry many cards, each with its own 32 GB and its own video engines.

Server Edition vs workstation RTX PRO 4500

Both cards use the same chip configuration and the same 4.4 by 10.5 inch board outline. The Server Edition changes the parts that matter in a server: passive cooling, a single slot, lower power and no display outputs listed on its page.

NVIDIA RTX PRO 4500 Blackwell Server Edition product page, datasheet (April 2026) and product brief PB-12790-001_02 (20 April 2026); NVIDIA RTX PRO 4500 Blackwell workstation datasheet (April 2026) and product page; NVIDIA vGPU supported GPU list, updated 2 October 2026; all read on 10 October 2026.

The bandwidth gap is 800 against 896 GB/s, about 11 per cent by our arithmetic. Generation speed for a model that fits on one card is mostly limited by memory bandwidth, so the ceiling per card is that much lower than on the workstation card. FP32 is the same on both datasheets. NVIDIA footnotes the workstation card’s 1,617 AI TOPS as “Effective FP4 TOPS with sparsity” and does not state the basis of the Server Edition’s 1.6 PFLOPS, so the two FP4 figures cannot be set side by side. Everything that fits in 32 GB on the workstation card fits on the Server Edition, and our RTX PRO 4500 workstation card guide sizes the model classes.

RTX PRO 4500 release date and announcement

NVIDIA announced the workstation RTX PRO 4500 Blackwell on 18 March 2025, together with the RTX PRO 5000 and 4000, and its press release said these cards “will be available in the summer” through workstation makers. That release did not mention a Server Edition of the 4500.

For the Server Edition we found no NVIDIA press release or announcement date. NVIDIA’s own documents are the product page, a product brief whose first version is dated 2 March 2026 and a datasheet with the code “Apr26”. Supermicro’s press release of 18 March 2026 announced systems with “the new NVIDIA RTX PRO 4500 Blackwell Server Edition GPU”.

vGPU and NVIDIA AI Enterprise support

As of 10 October 2026, NVIDIA’s list of GPUs supported by vGPU software has one RTX PRO 4500 entry, the Server Edition, with release 20.0 as the first supported release. On vSphere, NVIDIA’s support matrix (updated 29 September 2026) lists it for VCF 9.1, for VCF 9.0 from 9.0.1 and for ESXi 8.0 from Update 3g. The product brief names R595 or later drivers, and Lenovo names vPC, vApps and RTX vWS as the editions.

NVIDIA’s vPC sizing guide, updated 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. In its table for a 2U server it counts eight cards and 128 desktops, a table that “reflects maximum supported density rather than a recommended deployment point”. For virtual workstations, NVIDIA’s RTX vWS sizing guide puts 6 to 8 light users per RTX PRO 4500 on the DC-4Q profile and 3 to 6 medium users on DC-4Q or DC-8Q. Our guide to MIG and vGPU seats per card explains the profiles and the licence types.

For AI workloads, the NVIDIA AI Enterprise 8.2 support matrix (2 September 2026) lists the RTX PRO 4500 Blackwell Server Edition among its discrete GPUs, with the condition that GPUs “are supported with servers listed on the NVIDIA-Certified Systems”. If you need NVIDIA’s support for AI Enterprise, specify a server model that NVIDIA lists as certified with this card.

MIG on the RTX PRO 4500 Server Edition

NVIDIA’s MIG user guide, updated 11 September 2026, has a profile table for the “RTX PRO 4500 Blackwell 32GB products”. It does not name an edition, while the Server Edition’s product page states “Up to 2 MIG @ 16 GB”. The profiles divide the card into halves, and the +gfx variants also run graphics.

NVIDIA MIG user guide, supported MIG profiles, RTX PRO 4500 Blackwell table, updated 11 September 2026. L2 cache follows the memory share; each half has one copy engine and the whole card two.

The media engines do not divide evenly. With two plain halves, each instance gets one decoder, one encoder and one JPEG engine, so by our reading of the table the third decoder and encoder and the optical flow accelerator stay unused; a video workload that needs all three decoders takes the +me.all profile, and the second half then runs without media engines. NVIDIA’s vPC sizing guide states that the card supports MIG-backed vGPU, so each half can back virtual machines. In an eight-card server, MIG gives 16 isolated instances of 16 GB, each with its own memory and compute, which suits several small models or tenants that must not affect one another.

Servers that list the RTX PRO 4500 Server Edition

The chassis decides the density more than the card does. The table lists only what the makers’ pages state as we read them on 10 October 2026.

Supermicro “RTX PRO Blackwell Server Edition” systems page; Lenovo Press LP2391, updated 28 July 2026; NVIDIA vPC sizing guide, 19 August 2026. All read on 10 October 2026.

Lenovo’s guide also records in its update of 28 July 2026 that “The SR650a V4 now supports the GPU”; we could not read a card count for it. HPE has published a solution brief for the ProLiant Compute DL380 Gen12 with the RTX PRO 4500 and 6000 Server Editions, also without a count we could read, so neither is in the table. For any model, the maker’s configurator and GPU population rules decide the slots, risers, fans and cables.

We build AI servers to order with four to eight RTX PRO 4500 Server Edition cards and check the rack, power and airflow before we quote. Tell us the server model or rack position you have in mind and how many cards you plan.

RTX PRO 4500 Server Edition power, airflow and cables

At 165 W per card, eight cards draw 1,320 W, before processors, memory, drives and fans. NVIDIA’s product brief gives 165 W as default and maximum board power and 100 W as the minimum. A lower cap set with nvidia-smi has to be set again after each driver load, while a cap set out of band through SMBPBI stays in force across driver loads and reboots. Plan rack power at 165 W per card. The brief recommends a PCIe Gen5 x16, Gen5 x8 or Gen4 x16 link.

Each card takes one 16-pin PCIe CEM5 connector, and Lenovo’s guide states that the GPU option “does not ship with auxiliary power cables” and that “Cables are server-specific due to length requirements”, so the cables belong in the server order.

The card has no fan of its own. NVIDIA’s product brief gives an ambient operating temperature of 10 to 45 °C and a heat sink that takes airflow from left to right or from right to left. Lenovo describes it as a passively cooled card “requiring system airflow to operate” and lists 0 to 50 °C for its option. The inlet temperature that counts in the rack is the limit the server maker sets for the chassis with this GPU, which can be lower. A tower case or a general-purpose server without GPU airflow is not the place for it; the workstation card is the choice there.

What the RTX PRO 4500 Server Edition is for at mid-size scale

The card suits work that needs many separate GPUs of moderate size rather than one large memory pool. Three worked examples show the scale, using the vendors’ figures above and our arithmetic.

For virtual desktops for 1,000 knowledge workers at NVIDIA’s maximum density of 16 desktops per card on 2 GB profiles, the company needs 63 cards, which is eight servers of eight cards with almost no headroom. On 3 GB profiles, at 10 per card, the same desktops need 100 cards, or 13 such servers. A further server per cluster covers a host failure. NVIDIA presents these densities as maxima, so a pilot with your own desktops sets the final profile.

An inference platform for 500 to 2,000 users that runs several small models can start from one eight-card server, which holds 256 GB of GPU memory, either as eight cards of 32 GB or as 16 MIG instances of 16 GB. An embedding model, a reranker and classification models can each take a MIG half, and an assistant model of the 14B class in FP8 or a 32B model in 4-bit takes a whole card per copy, as the workstation guide sizes them. A 32B model in FP8 or a 70B model does not fit one card, and the comparison of eight RTX PRO 4500 or four RTX PRO 6000 Server Edition covers that decision.

For video analytics, each card has three NVDEC engines, so an eight-card server has 24, and a MIG half keeps one decoder of its own. The number of camera streams per engine depends on resolution, codec and model and has to come from the analytics vendor’s figures for your streams.

We size GPU servers from the models, user counts and desktop profiles you plan. Send us your model list and user numbers through the form below, and we propose the card count and layout.

NVIDIA compares the Server Edition with the L4 and says it “delivers over 5x the performance of the previous-generation L4 GPU”, without test conditions on the product page; our L4 vs RTX PRO 4500 Server Edition comparison covers the slot and power differences. Where a model needs more than 32 GB on one card, the L40S vs RTX PRO 4500 Server Edition comparison weighs 48 GB at 350 W against this card.

What we supply

We supply the RTX PRO 4500 Blackwell Server Edition across the EU with manufacturer warranty, on one EU contract and invoice, as cards or in AI servers built to order, assembled and burn-in tested. Alongside it we supply the RTX PRO 6000 Server Edition, the L40S, the L4 and the H200 NVL from our professional GPU range, with NVIDIA AI Enterprise and vGPU licences on the same invoice. Operating system, drivers, CUDA and a container runtime are installed on request. We check the rack, power and airflow before we quote, and the configuration and quote follow within one business day.

FAQ