RTX PRO 4500 Blackwell: what 32 GB at 896 GB/s buys over the RTX 4500 Ada, and where it sits between the 4000 and the 5000

A 32 GB card between the 4000 and the 5000

NVIDIA announced the RTX PRO 4500 Blackwell on 18 March 2025, together with the RTX PRO 5000 and 4000, and said the three cards would be available in the summer of 2025 from workstation makers and through distribution partners. It replaces the RTX 4500 Ada of 2023 in the same 4.4 by 10.5 inch dual-slot format, at 10 W less, and sits between the single-slot 24 GB RTX PRO 4000 and the 48 or 72 GB RTX PRO 5000. NVIDIA’s MIG guide places it on the GB203 chip; the 5000 and the 6000 use the larger GB202.

NVIDIA datasheets: RTX 4500 Ada (October 2023), RTX PRO 4500 (April 2026), RTX PRO 4000 (June 2026) and RTX PRO 5000 (July 2026). Two RTX PRO 5000 datasheets on NVIDIA’s site in October 2026 give 384-bit and 512-bit, so the table leaves its bus width open.

What changed against the RTX 4500 Ada

Bandwidth changed most. GDDR7 on a 256-bit bus gives 896 GB/s against 432, 2.07 times as much, and memory grows from 24 to 32 GB. CUDA cores rise 37 per cent and FP32 29 per cent, to 51 TFLOPS. NVIDIA quotes the Blackwell card’s tensor rate in FP4 and the Ada card’s in FP8, both with sparsity. The RTX PRO 4500 Server Edition, with the same 10,496 CUDA cores and 51 TFLOPS of FP32, is specified at 1.6 PFLOPS in FP4 and 811 TFLOPS in FP8, a 2:1 ratio. Halved on that basis, the workstation card’s 1,617 becomes about 809 TFLOPS of FP8, 28 per cent above the Ada card’s 634 by our arithmetic.

The other changes are in the platform: the host link moves to PCIe 5.0 x16, the outputs to DisplayPort 2.1b, and the video engines add 4:2:2 H.264 and HEVC encoding and decoding; NVIDIA’s codec support matrix marks 4:2:2 encoding as unsupported on the RTX 4500 Ada. Neither card is on NVIDIA’s vGPU list as of October 2026.

Measured: rendering, video and CAD

Puget Systems published its tests of the Blackwell range against the Ada predecessors on 18 December 2025, run under Windows 11 on a Ryzen 9 9950X3D with NVIDIA driver 573.92.

Puget Systems, 2025 Professional GPU Content Creation Roundup, 18 December 2025.

Puget’s tests of Topaz Video’s Starlight models, published in September 2026, place the card between its neighbours: about 14 seconds per frame in Starlight Mini and about 9 in Starlight Precise 2.5, against under 11 and just over 6 seconds for the RTX PRO 5000 and more than 19 and nearly 20 for the RTX PRO 4000. For value, Puget recommends the 4500 or the 48 GB 5000.

CAD gains least. Puget’s engineering roundup of December 2025 found the midrange Blackwell cards, the 2000, 4000 and 4500, 12 to 21 per cent faster than their Ada predecessors in SOLIDWORKS, the graphics tests in Inventor “relatively insensitive to GPUs once a certain threshold is reached”, and Revit “primarily CPU-dependent”. For a CAD seat alone the step is small, as our RTX PRO 4000 comparison found for the card below.

Local AI: what fits in 32 GB

We size with the rule of our users-per-card guide: 90 per cent of the card’s memory, less about 3 GiB for activations and the runtime, holds the weights and the KV cache, and the rest is left for the CUDA context and fragmentation. NVIDIA’s memory sizes are binary, so that is 25.8 GiB of 32 GiB; the driver keeps a little back as well, which is why the counts below are rounded down. Weights are checkpoint sizes on Hugging Face; the cache per token follows from each model’s configuration.

Our arithmetic: 25.8 GiB minus the checkpoint (OpenAI’s gpt-oss-20b, NVIDIA’s FP8 and NVFP4 Qwen3 and NVFP4 Llama 3.3 70B, Qwen’s BF16 14B and FP8 32B), divided by the FP8 KV cache of one 8,192-token conversation: 0.625 GiB for Qwen3-14B (80 KiB per token), 1 GiB for Qwen3-32B (128 KiB) and under 0.1 GiB for gpt-oss-20b, whose twelve full-attention layers store 12 KiB per token. A BF16 cache halves the counts.

A 32B model in 4-bit is the practical upper limit for this card, and the reason to choose it over the 4000: on 24 GB the same NVFP4 checkpoint leaves no room for cache by this rule. A 32B model in FP8 needs the RTX PRO 5000. The checkpoint also has to match the engine. NVIDIA’s FP8 and NVFP4 Qwen3 checkpoints name TensorRT-LLM on Blackwell, and TensorRT-LLM’s support matrix in its pre-release documentation of 26 September 2026 (the stable release, as of October 2026, is 1.2.1) lists NVFP4, MXFP4, per-tensor FP8 and an FP8 KV cache for the RTX PRO chips (SM120), but not FP8 with block scaling, as in Qwen’s own FP8 releases, nor AWQ or GPTQ. Once a model fits, bandwidth sets the pace of generation: 896 GB/s is a third more than the 4000’s 672 and two thirds of the 5000’s 1,344. Puget Systems’ own MLPerf® Client v1.0 runs, published on 18 December 2025, put the Blackwell cards about 50 per cent ahead of their Ada predecessors on average, the 2000 excepted.

Result not verified by MLCommons Association. The MLPerf name and logo are registered and unregistered trademarks of MLCommons Association in the United States and other countries. All rights reserved. Unauthorized use strictly prohibited. See www.mlcommons.org for more information.

We supply the RTX PRO 4500 workstation card and the cards either side of it, the 24 GB RTX PRO 4000 and the 48 or 72 GB RTX PRO 5000. Tell us the model and the context length, and we say which one fits.

MIG: two 16 GB halves, with a caveat

NVIDIA’s MIG user guide of 4 November 2026 lists the RTX PRO 4500 Blackwell with at most two instances: MIG 1g.16gb, two of them, each with half the memory and half the SMs, or 2g.32gb for the whole card, both also as +gfx profiles that support graphics. That makes it the smallest card in our range in the guide. By the same rule, each half holds an 8B model in FP8, such as NVIDIA’s 8.8 GiB Qwen3-8B-FP8, with cache for about four 8K conversations; gpt-oss-20b, at 12.8 GiB, leaves no room for cache in a 16 GB half.

The caveat is in the documentation. The Server Edition’s product page states “Up to 2 MIG @ 16 GB”, but the workstation card’s datasheet of April 2026 and its product page do not mention MIG, while NVIDIA’s RTX PRO 5000 and 6000 datasheets do. The guide’s table for the 4500 names no edition, and its full-card profile lists three NVENC and three NVDEC engines, the Server Edition’s count; the workstation card has two of each. The guide’s extra prerequisites for workstation cards, a minimum vBIOS and the switch from graphics to compute mode with DisplayModeSelector 1.72.0 or later, name only the 6000 and the 5000, and NVIDIA’s DisplayModeSelector page does not list the 4500. Lenovo’s product guide for the workstation card as an option for its servers, published on 26 May 2026, lists MIG as “No support”. MIG needs Linux; since its update of 4 November 2026 the guide gives a minimum driver for the RTX PRO 4500 as well, R575 (575.64 or later), again without naming an edition. If MIG is the reason to choose the workstation card, have the vBIOS version and MIG support confirmed in writing before ordering; our step-by-step MIG guide covers the commands.

The Server Edition is a different card

The RTX PRO 4500 Blackwell Server Edition keeps the 10,496 CUDA cores, the 32 GB and the 51 TFLOPS of FP32, and changes the rest for servers: passive cooling on a single-slot full-height, full-length board, 165 W, memory at 800 GB/s instead of 896, and three encoders and three decoders. It is on NVIDIA’s vGPU list from release 20.0 and splits into two 16 GB MIG instances. NVIDIA positions it against the L4, claiming “over 5x the performance” in small-model inference, without test conditions on the page. It needs the airflow of a rack server, not a tower; our article on the RTX PRO 4500 Server Edition covers it in detail.

Power, cooling and video

The workstation card has a total board power of 200 W and one 16-pin PCIe CEM5 connector, and NVIDIA’s quick start guide lists an adapter from two 8-pin PCIe cables for it, so the power supply needs two free 8-pin leads or a native 16-pin cable. It is a full-height dual-slot card of 4.4 by 10.5 inches. NVIDIA describes the cooler only as active; Leadtek, one of its board partners, lists a blower. NVIDIA’s datasheet and product page give two ninth-generation NVENC and two sixth-generation NVDEC engines; the RTX PRO 5000 and the Server Edition have three of each. The four DisplayPort 2.1b outputs drive four 4K displays at 165 Hz or two 8K displays at 100 Hz.

RTX PRO 4000, 4500 or 5000: which to choose

The RTX PRO 4000 (24 GB, 672 GB/s, 145 W, single slot) suits CAD seats, towers with room only for a single-slot card, and models up to 14B in FP8 or gpt-oss-20b.

The RTX PRO 4500 (32 GB, 896 GB/s, 200 W) is the card for a 32B model in 4-bit with room for several conversations, for GPU rendering and video, where Puget measured 32 to 73 per cent over its Ada predecessor, and for MIG on the smallest card in our range that NVIDIA’s MIG guide lists, with the caveat above.

The RTX PRO 5000 (48 or 72 GB, 1,344 GB/s, 300 W) is the step for 32B models in FP8, 70B in 4-bit, three encoders, and graphics vGPU on the 72 GB card, which NVIDIA supports from vGPU 20.2 on Red Hat Enterprise Linux with KVM 9.6 only. Our RTX PRO 5000 comparison sizes both versions, and the family comparison adds the 6000.

What we supply

We supply the RTX PRO 4500 Blackwell workstation card across the EU with manufacturer warranty, on one EU contract and invoice, as a single card or in a workstation built to order with the power supply and airflow it needs, alongside the rest of the professional GPU range. For rack servers with vGPU, we supply the RTX PRO 4500 and RTX PRO 6000 Server Editions.

FAQ