NVIDIA RTX PRO 4000 Blackwell 24GB Specs, Benchmarks & Pricing

The NVIDIA RTX PRO 4000 Blackwell is a professional single-slot workstation GPU announced on March 18, 2025, built on the NVIDIA Blackwell architecture (GB203 die, TSMC 4nm process) with 8,960 CUDA cores across 70 streaming multiprocessors, 280 fifth-generation Tensor Cores, and 70 fourth-generation RT Cores. It is the successor to the RTX 4000 Ada Generation and is positioned as "the most advanced single-slot GPU for professionals," delivering 40 TFLOPS of single-precision FP32 compute and 1,290 AI TOPS (theoretical FP4 with sparsity).

The card features 24 GB of GDDR7 ECC memory on a 192-bit memory bus providing 672 GB/s bandwidth, a significant improvement over the predecessor's 20GB GDDR6 at 360 GB/s. The fifth-generation Tensor Cores introduce FP4 precision support alongside FP8, BF16, TF32, FP16, INT8, and INT4, enabling broader AI and machine learning workloads. The fourth-generation RT Cores double the ray-triangle intersection rate versus the prior generation.

With a TDP of 145W in a single-slot, full-height form factor (4.4" H x 9.5" L), the RTX PRO 4000 Blackwell requires one 16-pin power connector. It connects via PCIe 5.0 x16 and provides four DisplayPort 2.1b outputs supporting up to four independent displays at 4K 165 Hz or two displays at 8K 100 Hz. Video engine support includes one ninth-generation NVENC encoder and one sixth-generation NVDEC decoder.

The RTX PRO 4000 Blackwell uses the same GB203 die as the GeForce RTX 5080 (which has all 10,752 shaders enabled), but with 8,960 of the 10,752 shader processors active (70 of 84 SMs enabled). The GPU is distributed through PNY and available from workstation OEMs including BOXX, Dell, HP, and Lenovo.

Strengths

  • Excellent FP16 compute performance (top 25% of GPUs)
  • 24GB VRAM for model and data capacity

Specifications for NVIDIA RTX PRO 4000 Blackwell

SpecificationPerformance Ranking
FP32 TFLOPs
61st @ 40 TFLOPs (Mid Tier)(Mid)
FP16 TFLOPs
75th @ 322.5 TFLOPs (Top Tier)(Top)
Tensor Core Count
61st @ 280 Cores (Mid Tier)(Mid)
Memory Capacity (GB)
73rd @ 24 GB (Mid Tier)(Mid)
Memory Bandwidth (GB/s)
59th @ 672 GB/s (Mid Tier)(Mid)
Int8 TOPs
60th @ 322.5 TOPs (Mid Tier)(Mid)

Real-time NVIDIA RTX PRO 4000 Blackwell GPU Prices

We're tracking 0 of the NVIDIA RTX PRO 4000 Blackwell GPUs currently available for sale.
Buy Now

Compare Price/Performance to other GPUs

We track real-time prices of other GPUs too so that you can compare the price/performance of the NVIDIA RTX PRO 4000 Blackwell GPU to other GPUs.
Compare GPU Price/Performance

Compare NVIDIA RTX PRO 4000 Blackwell to Another GPU

Compare the NVIDIA RTX PRO 4000 Blackwell directly to another GPU to see specs, benchmarks, and prices side-by-side.
Compare GPUs Side-by-Side

Price History

NVIDIA RTX PRO 4000 Blackwell Price History

Insufficient historical data for price trends. More data will be available as we continue tracking prices.

Product Identifiers

Manufacturer Part Numbers (3)
NVIDIA Part Number
900-5G147-2270-000
NVIDIA Part Number
900-5G147-2570-000
NVIDIA Part Number
900-5G147-0070-000
Available from 4 Partners (4 products)
PNY
NVIDIA RTX PRO 4000 Blackwell 24GB GDDR7 Graphics Card
VCNRTXPRO4000B-PB(sku)
Dell
Dell NVIDIA RTX PRO 4000 Blackwell Workstation Edition, 24 GB GDDR7, full height, PCIe 5.0x16, 4 DP
490-BLNG(part number)
HP
HP NVIDIA RTX PRO 4000 Blackwell 24 GB 4DP Graphics Card
B11F3AA(part number)
Lenovo
NVIDIA RTX PRO 4000 Blackwell 24GB DP*4 GDDR7 Graphics Card
4X61T95636(part number)

References

Notes

  1. FP32 TFLOPS of 40.0 sourced directly from the NVIDIA RTX PRO 4000 Blackwell official product page (nvidia.com), representing single-precision CUDA core FP32 performance at GPU-Boost clock. Consistent with 8,960 CUDA cores x 2,232 MHz boost clock x 2 ops/cycle / 1000 = 40.00 TFLOPS (derived boost clock, not explicitly published by NVIDIA).
  2. AI TOPS of 1,290 is the figure from the NVIDIA official product page, labeled as theoretical FP4 TOPS using the sparsity feature (per page footnote: "Theoretical FP4 TOPS using the sparsity feature"). Consistent with the RTX PRO 2000 Blackwell datasheet footnote which confirms Blackwell uses FP4 sparse for the AI TOPS figure.
  3. int8TOPS of 322.5 represents non-sparse (dense) INT8 performance, derived from FP4 sparse TOPS: 1,290 (FP4 sparse) / 4 = 322.5 (INT8 dense). Derivation: FP4 sparse = 2x INT8 sparse = 4x INT8 dense per Blackwell 5th-gen tensor core architecture. Dense value used for consistent cross-vendor comparison.
  4. fp16TFLOPS of 322.5 is FP16 Tensor Core throughput with 2:4 structured sparsity, derived from FP4 sparse TOPS using the Blackwell tensor core ratio (FP16 sparse = FP4 sparse / 4): 1,290 / 4 = 322.5 TFLOPS. Dense FP16 Tensor = 322.5 / 2 = 161.25 TFLOPS. The NVIDIA product page does not publish FP16 separately for this workstation card.
  5. Memory bandwidth of 672 GB/s per NVIDIA official product page and confirmed by Wiredzone, Provantage, and multiple third-party sources. The memory is 24 GB GDDR7 ECC on a 192-bit bus at 28.0 Gbps effective speed (672 / 24 bytes = 28 Gbps).
  6. CUDA core count of 8,960 confirmed by B&H Photo product listing snippet (8960 CUDA Cores), Wiredzone product page, Provantage, and CG Channel article. The GPU die is GB203 (same as GeForce RTX 5080 which has all 10,752 shaders enabled); the RTX PRO 4000 has 70 of 84 SMs active per TechPowerUp.
  7. Tensor core count of 280 confirmed by multiple third-party databases (gpus.net, HMC-tech, axiomgaming). 70 SMs x 4 tensor cores per SM = 280 tensor cores. RT core count of 70 confirmed by TechPowerUp search snippet and gpus.net (70 RT acceleration cores, one per SM).
  8. TDP of 145W per NVIDIA official product page. Some third-party sources (Provantage, CG Channel, gpus.net, axiomgaming) list 140W; NVIDIA official value of 145W is used as authoritative. Card requires one 16-pin power connector per Provantage listing.
  9. GPU die is GB203 on TSMC 4nm process with approximately 45.6 billion transistors and 378mm2 die area (per gpus.net, HMC-tech; consistent with GB203 known parameters). Architecture uses compute capability 12.0 (sm_120), confirmed by the official NVIDIA CUDA GPUs table at developer.nvidia.com/cuda-gpus.
  10. Release date of March 18, 2025 per TechPowerUp GPU database and PNY press release (pny.com). The GPU was announced March 18, 2025. Per the NVIDIA investor press release and CG Channel, the RTX PRO 5000/4500/4000 Blackwell were slated for distribution "in the summer" of 2025 through OEM workstation partners.
  11. FP4 hardware precision is supported by 5th-generation Tensor Cores. FP8 is also supported. Transformer Engine 2.15+ supports Blackwell (sm_120) for FP8, MXFP8, and NVFP4 (per official TE documentation). In practice, as of the March 2025 launch, many pre-built vLLM and framework images had incomplete sm_120 kernel support, requiring builds from source with CUDA 12.8+ and TORCH_CUDA_ARCH_LIST=12.0 to fully use FP8 on this card. Lower-level libraries (cuBLASLt, TensorRT) support FP8 natively on Blackwell.
  12. Estimated MSRP of $1,799 USD based on launch-time market pricing. NVIDIA does not publish official MSRP for professional workstation GPUs. Dell lists the card (part 490-BLNG) at $1,799.99 at dell.com. Wiredzone lists the NVIDIA OEM variant (NVPN 900-5G147-0070-000) at $1,824.00. The $1,799 price represents the consistent retail/OEM list price at launch.
  13. NVPN 900-5G147-2270-000 (bulk/OEM) sourced from B&H Photo product listing (product ID 1938761-REG). NVPN 900-5G147-2570-000 (retail) confirmed from gpus.net, serversupply.com, and centralcomputer.com. NVPN 900-5G147-0070-000 (Wiredzone OEM direct) sourced from Wiredzone product listing. The board_id prefix is 5G147 (analogous to other NVIDIA workstation cards using 5G-prefix board IDs).
  14. PNY SKU VCNRTXPRO4000B-PB confirmed from Newegg, Provantage, and B&H Photo product listings. Dell part number 490-BLNG confirmed from dell.com official product page. HP part number B11F3AA confirmed from gpus.net and ProVantage. Lenovo part number 4X61T95636 confirmed from lenovo.com product page URL (GB English locale).