RTX PRO 4000 Blackwell SFF
The RTX PRO 4000 Blackwell SFF is Nvidia’s 24GB graphics card with a 70W board power rating, built on the Blackwell GB203 architecture. Below you’ll find what it can run locally, its benchmark scores and its full spec sheet. No board versions are in stock right now, but the specs and benchmarks below still apply whenever it returns to retailers.
Prices update regularly and may differ at the retailer. This page contains affiliate links.
Benchmarks
Standardized benchmark scores for this chip, each compared to the fastest card we track and linked to its source. They apply to every board version.
Render
Compute & AI
3D graphics
Local AI performance
What the RTX PRO 4000 Blackwell SFF can run on your own machine, and how fast.
Largest model it can run (8K context)
- Q4 (the usual choice, near-original quality)
- ~32B
- Q8 (closer to original, half the model size)
- ~19B
- FP16 (unquantised, for training and research)
- ~10B
Parameter counts the card holds with room for an 8K conversation, counting the model weights plus its attention cache. A longer conversation needs more memory, so it lowers these figures.
Speed & value
- Speed, 8B model at Q4
- 66-76 tok/s
- Memory bandwidth
- 432 GB/s
- Price per GB of VRAM
- —
Generation speed tracks memory bandwidth, because every word produced reads the whole model from memory. Shown as a range.
Models that fit, at Q4
| Model | Size | Memory needed | Est. speed |
|---|---|---|---|
| Llava 34B | 34.8B | 21.5GB | 16-19 tok/s |
| Yi 34B | 34.4B | 21.5GB | 16-19 tok/s |
| Nous Hermes2 34B | 34.0B | 21.6GB | 16-19 tok/s |
Popular models small enough for this card at Q4, one per family. Speed is left blank for mixture-of-experts models, where a bandwidth estimate is not reliable enough to publish.
Choosing a card for AI? Compare every GPU for local AI, including setups that pool the memory of two or more cards.
RTX PRO 4000 Blackwell SFF availability
Live retailer stock for this model.
Currently out of stock
None of the board versions of the RTX PRO 4000 Blackwell SFF are in stock at the retailers we track right now. Prices refresh daily, so check back — or compare in-stock alternatives below.
Browse in-stock graphics cards for AISpecifications
The reference RTX PRO 4000 Blackwell SFF measures 168 mm long.
Performance
- Boost clock
- 1342 MHz
- Shader cores
- 8,960
- SM count
- 70
- Texture units (TMUs)
- 280
- Render outputs (ROPs)
- 96
- RT cores
- 70 · 4th gen
- Tensor / AI cores
- 280 · 5th gen
Memory
- VRAM
- 24GB GDDR7
- Memory bus
- 192-bit
- Bandwidth
- 432 GB/s
- Memory speed
- 18 Gbps
GPU die
- Architecture
- Blackwell GB203
- Process node
- TSMC 4N
- Foundry
- TSMC
- Die size
- 378 mm²
- Transistors
- 45.6 billion
- Generation
- RTX PRO Blackwell
Power
- TDP
- 70W
Display & outputs
- Max resolution
- 15360 × 8640
- Max displays
- 4
- Outputs
- 4 Mini DisplayPort 2.1b
- PCIe interface
- PCIe 5.0 ×16
APIs & software
- DirectX
- 12
- Vulkan
- 1.4
- OpenGL
- 4.6
- OpenCL
- 3.0
- CUDA
- Compute 12.0
- Shader model
- 6.7
Video encoding & decoding
| Codec | Encode | Decode |
|---|---|---|
| H.264 / AVC | Supported | Supported |
| H.265 / HEVC | Supported | Supported |
| AV1 | Supported | Supported |
| VP9 | Not supported | Supported |
Encoder: NVENC 9th gen · Decoder: NVDEC 6th gen
Features
- Upscaling
- DLSS Super Resolution, Ray Reconstruction, Frame Generation, Multi Frame Generation, DLSS 5 Neural Rendering
- Also runs
- FSR 3, FSR 3 Frame Generation, XeSS 2, XeSS 2 Frame Generation
- ECC memory
- Yes
Physical
- Reference length
- 168 mm
- Slot width
- 2 slots
- Released
- 2025-08-11
8,960 shader cores feed a 192-bit memory interface delivering 432 GB/s of peak bandwidth. 70 RT cores handle real-time ray tracing; 280 tensor cores accelerate AI workloads and DLSS Super Resolution. Compare the RTX PRO 4000 Blackwell SFF against consumer and workstation cards on the ranking of graphics cards for AI.
How the RTX PRO 4000 Blackwell SFF compares
The RTX PRO 4000 Blackwell SFF has no offer in stock right now, so these are the closest cards you can buy today, picked on memory, estimated generation speed and live prices. Gaming cards are included wherever they hold as much for less. Open a card for its full spec sheet.
Affiliate disclosure: Some links on this page are affiliate links — MaxMyBuild may earn a commission at no extra cost to you if you buy through them. As an Amazon Associate, MaxMyBuild earns from qualifying purchases. For Amazon, current price and availability are shown on the Amazon product page; product prices and availability are accurate as of the date and time of purchase and are subject to change. Prices shown for other retailers are approximate and may differ at checkout. “Search” links are non-affiliate and provided for convenience only.