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#
| Item | Parameter |
|---|
| Architecture | NVIDIA Blackwell 2.0 (GB202) |
| Process | TSMC 4nm |
| Transistors | 92.2 billion |
| Die Size | 750 mm² |
| CUDA Cores | 24,064 |
| Tensor Cores | 752 (5th-generation) |
| RT Cores | 188 (4th-generation) |
| FP32 | 125 TFLOPS |
| FP4 Tensor Core | 4,000 TFLOPS (inference) |
| Memory | 96 GB GDDR7 ECC |
| Memory Bandwidth | 1.79 TB/s (512-bit) |
| TDP | 600W (server version) / 300W (Max-Q version) |
| PCIe | Gen 5 x16 |
| ECC | Supported |
| MIG | Supported (up to 4 instances) |
| Launch Date | March 2025 |
| Price | Approximately $8,000-$10,000 (estimated) |
Comparison with RTX 6000 Ada#
| Metric | RTX Pro 6000 Blackwell | RTX 6000 Ada | Improvement |
|---|
| Architecture | Blackwell 2.0 | Ada Lovelace | New generation |
| Process | TSMC 4nm | TSMC 4N | Similar |
| Memory | 96GB GDDR7 | 48GB GDDR6 | +100% |
| Bandwidth | 1.79 TB/s | 960 GB/s | +86% |
| FP32 | 125 TFLOPS | 91.1 TFLOPS | +37% |
| FP8 | ~2,000 TFLOPS (estimated) | 1,458 TFLOPS | +37% |
| TDP | 600W | 300W | +100% |
| PCIe | Gen 5 | Gen 4 | Upgraded |
Version Differences#
| Version | TDP | Cooling | Suitable Scenarios |
|---|
| Server Version | 600W | Passive cooling | Data center, supports 8-card parallel |
| Workstation Version | 600W | Dual-fan active cooling | Small studios, supports 1-4 card config |
| Max-Q Version | 300W | Single turbo fan | Mobile 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
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)
External Links#