RTX PRO 4000 Blackwell vs RTX 4000 Ada: what the upgrade buys, including both SFF cards

Two generations, the same formats

The RTX 4000 Ada and its small-form-factor sibling have been NVIDIA’s midrange workstation cards since 2023. Their Blackwell successors keep the formats, a full-height single-slot card and a low-profile 70 W card, and change almost everything inside.

NVIDIA datasheets: RTX 4000 Ada (2023) and RTX PRO 4000 Blackwell (June 2026). The two SFF cards are compared further down.

What actually changed

Memory moved from GDDR6 to GDDR7, which is where most of the bandwidth gain comes from, and grew from 20 to 24 GB. The Tensor Cores are a generation newer and add FP4; the RT cores are a generation newer too. The host link doubles its bandwidth per lane with PCIe 5.0, the display outputs move to DisplayPort 2.1b, and the video engines add 4:2:2 encoding and decoding for H.264 and HEVC, which the Ada cards lack. The full-size card now draws 145 W instead of 130 W and takes one 16-pin connector, with an adapter from one 8-pin cable in the box. Like the Ada card, it needs an auxiliary power cable; the SFF card still needs none.

One caution when reading specification sheets: NVIDIA publishes two sets of figures for the full-size card. Its product page gives 1,290 AI TOPS, 40 TFLOPS of FP32 and 122 RT TFLOPS; the newer datasheet of June 2026 gives 1,178, 37 and 112. We use the datasheet. Both say 145 W, while some partner pages still show the launch figure of 140 W.

Measured: rendering and content creation

Puget Systems tested the full-size cards against each other in its December 2025 roundup; the SFF cards were not part of it.

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

AEC Magazine’s February 2026 review found similar steps for the RTX PRO 4000 over its Ada counterpart: 1.71 times in V-Ray RTX, 2.6 times in Twinmotion path tracing and 1.42 to 1.46 times in Procyon’s Stable Diffusion test. For GPU rendering the upgrade is large and consistent.

Local AI: bandwidth and 4 GB more memory

For language models the bandwidth line matters most, because token generation reads the weights once per token: 672 against 360 GB/s puts the ceiling 87 per cent higher on the same model. Compute rises less. NVIDIA quotes the Blackwell card’s AI TOPS in FP4 and the Ada card’s in FP8, both with sparsity. Halved to FP8 at the 2:1 ratio NVIDIA’s Server Edition specifications show, the full-size card goes from 427.6 to about 589 by our arithmetic, 38 per cent more. The 427.6 is from the Ada card’s product page; its 2023 datasheet prints 327.6, which does not match the clock implied by the card’s FP32 rate. Puget Systems’ own MLPerf® Client runs, which measure the token rate of a local language model, found the Blackwell cards about 50 per cent ahead of their Ada predecessors on average, the RTX PRO 2000 excepted; these results are unverified, not reviewed by MLCommons Association.

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.

The extra 4 GB changes which models fit. A 32B model in 4-bit is about 18 to 20 GB of weights, depending on the format, which fits a 24 GB card with a short context and leaves almost nothing on a 20 GB card. gpt-oss-20b at 13.8 GB fits both, and 14B models in FP8 fit the 24 GB card but are tight on 20 GB. Only the Blackwell card runs NVFP4 natively, the 4-bit format NVIDIA’s TensorRT-LLM supports on Blackwell and not on Ada; the Ada card runs 4-bit weights as INT4 through AWQ or GPTQ instead. Our VRAM guide has the full arithmetic.

The SFF pair: 70 W, low profile

The two SFF cards share the format that makes them useful: 2.7 by 6.6 inches, low profile, dual slot, powered from the slot, actively cooled, and supplied with both a full-height and a low-profile bracket. They suit compact workstations and small servers that have no GPU power cable, provided the maker lists a 70 W low-profile card for that chassis.

NVIDIA datasheets: RTX 4000 SFF Ada (April 2023) and RTX PRO 4000 SFF Blackwell (August 2025); the Blackwell FP32 figure is from NVIDIA’s product page. A 2024 revision of the Ada datasheet gives 320 GB/s.

Inside, the Blackwell card has the same 8,960 cores as its full-size sibling, but at 70 W it runs them at a much lower clock and with slower memory. It reaches 60 to 65 per cent of the full-size card’s FP32 rate, depending on which of NVIDIA’s figures is used, and 64 per cent of its bandwidth at 48 per cent of the power. Against the RTX 4000 SFF Ada it brings 24 instead of 20 GB and 432 instead of 280 GB/s, or 320 GB/s by the 2024 revision of NVIDIA’s datasheet.

NVIDIA’s launch blog claims “up to 2.5x higher AI performance, 1.7x higher ray-tracing performance and 1.5x more bandwidth” than the previous generation. The last two match NVIDIA’s published figures: 73 against 44.3 RT TFLOPS is 1.65 times, and 432 against 280 GB/s is 1.54 times. The first matches an FP4 figure, 770 TOPS, divided by an FP8 figure, 306.8 TFLOPS; NVIDIA’s blog gives no test conditions. At the same precision the gain is about 25 per cent by our arithmetic. We have not found an independent benchmark of the two SFF cards against each other.

One detail matters for upgrades in older machines. The Blackwell SFF card connects at PCIe 5.0 x8 through a full-length connector, and the Ada card at PCIe 4.0 x16. In a PCIe 4.0 workstation the new card runs at 4.0 x8, half the link bandwidth of the card it replaces. For inference with the model resident in memory that rarely matters; for workloads that stream data to the GPU all the time, it can.

CAD and BIM: the smallest step

Puget’s engineering roundup of December 2025 is the reference for design departments. In SOLIDWORKS, the midrange Blackwell cards, the 2000, 4000 and 4500, gained 12 to 21 per cent over their Ada predecessors; in SOLIDWORKS drawings every AMD card tested beat the fastest NVIDIA model by at least 18 per cent; Autodesk Inventor ran well on every card tested; and Revit, in Puget’s words, “is primarily CPU-dependent”. A CAD seat that is fast enough today will not feel much faster with the new card. The reasons to upgrade a CAD workstation are the display outputs for high-resolution monitors, the extra memory for very large assemblies and the ISV certifications of the current generation.

When the upgrade pays, and when it does not

It pays for GPU rendering, video and visualisation, where Puget measured 34 to 83 per cent in V-Ray, Blender, Octane, Unreal Engine, After Effects and Topaz Video AI and about 25 per cent overall in DaVinci Resolve, and for local AI, where the bandwidth nearly doubles and the extra 4 GB moves 32B models into range.

It pays less for CAD and BIM seats, where the gain is 12 to 21 per cent in SOLIDWORKS and close to nothing in Revit.

For small-form-factor machines, the Blackwell SFF card is ahead in every published performance figure, but check the host’s PCIe generation first and expect a smaller step than the full-size card makes.

Neither generation of these cards supports MIG or vGPU. For MIG, the Blackwell workstation cards with documented requirements start at the RTX PRO 5000; NVIDIA’s MIG guide also lists an RTX PRO 4500 with two instances of 16 GB but names no edition, and the workstation 4500’s datasheet and product page do not mention MIG, so have MIG support on that card confirmed in writing. Virtual machines with vGPU need a card on NVIDIA’s vGPU list, such as the RTX PRO 4500 Server Edition, as our RTX PRO family comparison explains.

What we supply

Eurokommerz supplies the RTX PRO 4000 Blackwell and the RTX PRO 4000 SFF EU-wide with manufacturer warranty, as cards or in configured workstations, together with the rest of the professional GPU range. Tell us the applications and the chassis, and we will tell you which card fits.

FAQ