Skip to main content

NVIDIA RTX Pro 6000 Blackwell#

Product Overview#

NVIDIA RTX Pro 6000 Blackwell is a professional workstation GPU based on the Blackwell architecture, released by NVIDIA in March 2025. It is the successor to the RTX 6000 Ada. It features 96GB GDDR7 ECC memory (first adoption of GDDR7 in professional cards), 125 TFLOPS FP32 compute, and 752 5th-generation Tensor Cores, delivering up to 4,000 TFLOPS FP4 inference compute. It supports PCIe 5.0 and Multi-Instance GPU (MIG), making it a flagships workstation card for AI development, large model training, and professional graphics.

Strategic Position: In NVIDIA's professional GPU product line, the RTX Pro 6000 Blackwell is the flagship model for 2025-2026, replacing the RTX 6000 Ada. Compared to the consumer-grade RTX 5090, it offers double the memory capacity (96GB vs 32GB), supports ECC memory and professional drivers, making it the preferred choice for enterprise AI workstations.

Core Specifications#

ItemParameter
ArchitectureNVIDIA Blackwell 2.0 (GB202)
ProcessTSMC 4nm
Transistors92.2 billion
Die Size750 mm²
CUDA Cores24,064
Tensor Cores752 (5th-generation)
RT Cores188 (4th-generation)
FP32125 TFLOPS
FP4 Tensor Core4,000 TFLOPS (inference)
Memory96 GB GDDR7 ECC
Memory Bandwidth1.79 TB/s (512-bit)
TDP600W (server version) / 300W (Max-Q version)
PCIeGen 5 x16
ECCSupported
MIGSupported (up to 4 instances)
Launch DateMarch 2025
PriceApproximately $8,000-$10,000 (estimated)

Comparison with RTX 6000 Ada#

MetricRTX Pro 6000 BlackwellRTX 6000 AdaImprovement
ArchitectureBlackwell 2.0Ada LovelaceNew generation
ProcessTSMC 4nmTSMC 4NSimilar
Memory96GB GDDR748GB GDDR6+100%
Bandwidth1.79 TB/s960 GB/s+86%
FP32125 TFLOPS91.1 TFLOPS+37%
FP8~2,000 TFLOPS (estimated)1,458 TFLOPS+37%
TDP600W300W+100%
PCIeGen 5Gen 4Upgraded

Version Differences#

VersionTDPCoolingSuitable Scenarios
Server Version600WPassive coolingData center, supports 8-card parallel
Workstation Version600WDual-fan active coolingSmall studios, supports 1-4 card config
Max-Q Version300WSingle turbo fanMobile workstation, ~15-20% performance loss

Key Features#

  • 96GB GDDR7 ECC Memory: First adoption of GDDR7 in professional cards, capacity is 2× that of RTX 6000 Ada
  • 5th-Generation Tensor Cores: Supports FP4 inference, delivering up to 4,000 TFLOPS
  • PCIe 5.0: Bandwidth doubled (vs PCIe 4.0)
  • Multi-Instance GPU (MIG): Can partition single GPU into 4 independent instances, each with up to 24GB memory
  • NVIDIA Confidential Computing: First professional GPU to support confidential computing
  • 8th-Gen NVENC + 6th-Gen NVDEC: H.264 decode throughput improved 2×, adds 4:2:2 H.264 and HEVC decode support

Manufacturer Information#

ItemContent
CompanyNVIDIA Corporation
Official Websitehttps://www.nvidia.com
Product Pagehttps://www.nvidia.com/zh-tw/products/workstations/professional-desktop-gpus/rtx-pro-6000-family/
LaunchMarch 2025
ArchitectureBlackwell 2.0

Suitable Scenarios#

  • AI development workstations (model training, fine-tuning, inference)
  • Large model local inference (70B+ usable)
  • Professional graphics (CAD, 3D rendering)
  • Data science (with ECC memory)
  • Video editing and encoding (8th-gen NVENC)