NVIDIA RTX PRO 5000 Blackwell benchmarks: 48 GB and 72 GB, and their limits

Where the card sits

In Puget Systems’ published benchmarks the RTX PRO 5000 Blackwell runs 35 to 58 per cent ahead of the RTX 5000 Ada in V-Ray GPU, Blender and Octane, and 19 per cent behind the RTX PRO 6000 Workstation Edition in Topaz Video. Its 48 GB and 72 GB versions scored the same in Puget’s tests, within its margin of error. The RTX PRO 5000 Blackwell is a cut-down GB202, the chip of the RTX PRO 6000, with 14,080 CUDA cores against the 6000’s 24,064. NVIDIA announced it in March 2025; a 72 GB version appeared on NVIDIA’s site in October 2025 and became generally available in December 2025. Both versions share every speed figure in NVIDIA’s datasheet; they differ in memory, MIG instance size and vGPU support.

NVIDIA datasheets and product pages (2025 to 2026), MIG user guide of 13 October 2026, vGPU support matrix of 29 September 2026. NVIDIA’s current RTX PRO 5000 datasheet (July 2026) gives 1,344 GB/s for both versions; NVIDIA’s December 2025 blog on the 72 GB card quotes 2,142 TOPS. The MIG guide lists an RTX PRO 4500 without naming its edition; the workstation datasheet omits MIG. The 4500 and 600 W 6000 columns are the workstation cards; the Server Editions run vGPU, the 4500 (800 GB/s) from vGPU 20.0 and the 6000 (1,597 GB/s) from 19.0.

Against the 600 W RTX PRO 6000, the 5000 has three quarters of the bandwidth and about half the tensor peak; against the RTX PRO 4500, 1.5 times the bandwidth. Bandwidth sets the pace of text generation; rendering and prompt processing depend mainly on compute. Our family comparison and the RTX PRO 4500 article cover the neighbours.

What has been measured, and by whom

The independent record is short. Puget Systems tested the 48 GB card in its content creation and engineering roundups of 18 December 2025, and both memory sizes in two Topaz Video tests in August and September 2026. Phoronix included the 48 GB card in its Linux review of 21 May 2026 (Ubuntu 26.04 LTS, NVIDIA driver 595.58.03) and publishes its results as charts. We found no review of the desktop card by StorageReview.

Two sources are easy to misread. NVIDIA’s laptop range also has an RTX PRO 5000 Blackwell, with 10,496 CUDA cores, 24 GB and 896 GB/s at 95 to 175 W; StorageReview’s RTX PRO 5000 reviews are of such laptops and say nothing about the desktop card. Several websites also publish tokens-per-second tables for the desktop card that leave out the engine version, the concurrency or the test system; we do not use them, because they cannot be compared fairly.

Rendering, video and CAD: Puget’s numbers

Puget’s December 2025 tests ran on a Ryzen 9 9950X3D with NVIDIA driver 573.92, its 2026 Topaz tests on a Threadripper PRO 9965WX with driver 596.72.

Puget Systems: 2025 Professional GPU Content Creation Roundup, 18 December 2025 (first five rows, 48 GB card); Topaz Video 1.6.1 analysis, 24 August 2026 (overall score); Starlight test, 17 September 2026.

For a buyer, memory size does not change speed when the work fits: Puget found a “negligible difference in performance between the two variants, small enough to fall within our margin of error” in Topaz Video 1.6.1, and less than 1 per cent in Starlight Mini. The card also lands closer to the RTX PRO 6000 Workstation Edition than its core count suggests: in Topaz Video it is 19 per cent behind with 59 per cent of the cores and half the board power. We supply both memory sizes, as a card or in a workstation built to order. Describe your renderer and your largest scene and our team will help.

CAD gains least. Puget’s engineering roundup found that “even the graphics tests in Inventor are relatively insensitive to GPUs once a certain threshold is reached”, that Revit “is primarily CPU-dependent”, and that every AMD card it tested beat the fastest NVIDIA card by at least 18 per cent in the SOLIDWORKS drawing test.

Language model benchmarks: what is published

Phoronix ran llama.cpp on its RTX PRO cards, the 48 GB 5000 among them, measuring generation of 128 tokens and prompt processing of 512 and 2,048 tokens. The models were Q8_0 files of Llama-3.1-Tulu-3-8B, Mistral-7B-Instruct-v0.3, DeepSeek-R1-Distill-Llama-8B and gpt-oss-20b on llama.cpp’s GPU and Vulkan back ends, plus GLM-4.7-Flash in IQ4_XS on the GPU back end and granite-3.0-3b-a800m-instruct in Q8_0 on Vulkan. The values are in chart images and the concurrency is not stated.

We found no system with the card in the MLPerf® Inference results up to round v6.0. 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. NVIDIA’s blog for the 72 GB card claims twice the text generation and 3.5 times the image generation performance of “prior-generation NVIDIA hardware” in “industry-standard benchmarks for generative AI”; the text of the post names no baseline card, model or precision. NVIDIA’s TensorRT-LLM performance overview (TensorRT-LLM 1.4.0rc1, updated 19 October 2026) covers the RTX PRO 6000 Server Edition, not the 5000. So we found no tokens-per-second figure for this card that states the engine and its version, the precision, the concurrency and the test system together. That leaves the ceiling below and a load test with your own model. GPU selection and a pilot plan with metrics are part of our Private AI/ML service: the first call is free of charge, and the assessment’s price is fixed before work begins. Send us the model and the number of concurrent users.

Single-user speed limit at 1,344 GB/s

Generating a token for one user reads every active weight once, so plain single-stream decoding cannot beat the bandwidth divided by the bytes read per token. That ceiling is an upper bound, not a forecast: real engines stay below it, and the KV cache adds reads as the context grows. Speculative decoding, in which one read of the weights checks several drafted tokens, is the exception that can pass it.

Our arithmetic: 1,344 GB/s divided by the bytes read per token (linear layers and output layer; the embedding table is only looked up). Weights in GiB, from OpenAI’s and NVIDIA’s checkpoints; bytes per token in GB, as bandwidth is quoted in GB/s. Fits: by our users-per-card rule, 90 per cent of the memory the driver reports, less about 3 GiB for the runtime, must hold the weights and the cache. The 1,792 GB/s of the RTX PRO 6000 workstation cards raises each ceiling by a third.

The mixture-of-experts models read only the experts each token uses, about 5 GB per token in all for gpt-oss-120b; but the fewer bytes a token reads, the more the fixed cost of each step weighs, and the further real engines fall below the line.

With several users, a dense model’s weights are read once per step for the whole batch, so aggregate throughput grows with concurrency until compute or cache space runs out. The 72 GB card has the advantage on cache: beside Llama 3.3 70B in NVFP4 it leaves about 22 GiB, enough for about seventeen 8,192-token conversations with an FP8 cache, while the 48 GB card has almost nothing left, by our users-per-card rule; our 48 GB or 72 GB comparison has the full table. Prompt processing is compute-bound, and there the gap to the 600 W RTX PRO 6000 widens from a quarter to about a half: 2,064 AI TOPS against 4,000, or by our arithmetic about 516 TFLOPS of dense FP8 against 1,000. Long RAG and coding-agent prompts are affected most.

Power, cooling and several cards

The card has a total board power of 300 W and one 16-pin PCIe CEM5 power connector; NVIDIA’s quick start guide lists an adapter from two 8-pin PCIe cables in the box. It is a full-height, dual-slot board of 4.4 × 10.5 inches; NVIDIA calls the cooler active, and its board partner Leadtek specifies a blower, which exhausts through the slot bracket. We found no published power, temperature or noise figures for the desktop card in text; Phoronix logged GPU power but shows it only in charts.

Four cards draw up to 1.2 kW before the processor. Without NVLink they talk over PCIe 5.0 x16, which suits one model or one team per card better than one model split across cards, whose tensor-parallel all-reduces then cross PCIe. The AI servers we build to order include workstations and short-depth servers on Threadripper PRO or Xeon W platforms with RTX PRO Blackwell Workstation cards, assembled and tested before delivery; for a server we also check the rack, power and airflow before we quote.

MIG, vGPU and software

NVIDIA’s MIG guide, updated on 13 October 2026, lists the 48 GB card with two isolated 1g.24gb instances or one 2g.48gb, each also as a +gfx variant for graphics; the 72 GB card is not in the guide yet and its datasheet gives two instances of 36 GB, so once MIG mode is enabled, read its profile names with nvidia-smi mig -lgip. MIG needs Linux, driver 575.57.08 or later in the guide’s update of 4 November 2026, vBIOS 98.02.73.00.00 or later on the 48 GB card, and the display mode switched to compute with DisplayModeSelector 1.72 or later; NVIDIA names no separate minimum for the 72 GB card. The card then drives no display, and NVIDIA says the system must be qualified for that mode.

vGPU covers only the 72 GB card, from vGPU 20.2 of August 2026, with graphics (Q- and B-series) vGPU types; we found no compute (C-series) types for it in NVIDIA AI Enterprise’s documentation. As of October 2026 NVIDIA’s support matrices list it only for Red Hat Enterprise Linux with KVM 9.6; for VMware vSphere they name, of the RTX PRO Blackwell cards, only the RTX PRO 6000 and 4500 Server Editions. The card has compute capability 12.0: NVIDIA’s datasheet lists CUDA 12.8, and PyTorch added Blackwell support in release 2.7 with CUDA 12.8 wheels, so check containers built on older CUDA releases first.

If you need vGPU on vSphere, we supply the RTX PRO 6000 Server Edition for rack servers, with the vGPU licences on the same quote. Tell us the vSphere version and the number of users, and we reply within one business day.

Which workloads suit the RTX PRO 5000, and which do not

The measured record supports the RTX PRO 5000 as a rendering and video card, well ahead of the RTX 5000 Ada and within 19 per cent of the RTX PRO 6000 Workstation Edition in Topaz Video at half the power. For language models its case rests on memory and arithmetic: 48 GB for models up to 32B in FP8, 72 GB for gpt-oss-120b or a 4-bit 70B model serving a small team; on 48 GB a 4-bit 70B model loads but leaves almost no room for conversations.

It is the wrong card for a 70B model in FP8: beside its 68 GiB of weights, 72 GB leaves too little room for the runtime and for the cache that useful concurrency needs. NVIDIA’s 84 GB RTX PRO 5500, whose datasheet appeared in October 2026, holds such a model with about 4.6 GiB of cache by the same rule, the 96 GB RTX PRO 6000 with about 15 GiB. It is also the wrong card for vGPU outside Red Hat KVM 9.6, for a CAD seat alone, and for a model too large for one card, which without NVLink must be split over PCIe. At the same 300 W, the RTX PRO 6000 Max-Q brings 96 GB, 1,792 GB/s and 3,511 AI TOPS, and the 72 GB card scored 7 per cent below it in Puget’s Topaz test; whether that step pays depends on the model and the budget.

What we supply

We supply the RTX PRO 5000 Blackwell in both memory sizes EU-wide, with manufacturer warranty on one EU contract and invoice, as a single card or in a workstation built to order; see the catalogue entry. Where the model needs more memory, we supply the 96 GB RTX PRO 6000 in all three editions and the 84 GB RTX PRO 5500, which NVIDIA lists as offered through its OEM and distribution partners as of 5 November 2026; quantities and lead time are confirmed in writing with the quote. For the software, our engineering partner Vixen.UNO delivers Private AI/ML: private LLMs on-premise with vLLM, Ollama or NVIDIA AI Enterprise, RAG assistants that respect each user’s access rights, and a pilot on one process with clear metrics, scaling only what has proved its value. The contract stays with Eurokommerz.

FAQ