The NVIDIA RTX PRO 6000 Blackwell Workstation Edition is the desk-side flagship of NVIDIA's professional Blackwell lineup. It drops the full GB202 GPU and an enormous 96 GB of GDDR7 ECC into a tower-friendly, actively cooled card with certified pro drivers. For creators and engineers, the headline isn't raw frame rate — it's the ability to hold a 70-billion-parameter model, a billion-polygon scene, or hours of 8K footage entirely in VRAM on one card.
Quick verdict: The single-card king for memory-hungry professional work — local LLMs, large-scene rendering, 8K video and simulation. 96GB + ECC has no consumer equal, and it cleanly outclasses the RTX 6000 Ada. The catch is price (≈$8,500, often $10k–13k) and a 600W / 12″ footprint. If your workload fits in 32GB, an RTX 5090 is far better value. Score: 92/100.
RTX PRO 6000 Workstation Edition — Full Specifications
Core specs are from NVIDIA's official datasheet; third-party figures (clocks, node, PSU guidance) are flagged.
| Specification | Detail |
|---|---|
| GPU & Architecture | |
| GPU chip | GB202 (Blackwell) — NVIDIA does not name the die |
| Process node | TSMC 4N-class (third-party; not on datasheet) |
| CUDA cores | 24,064 |
| Streaming Multiprocessors | 188 SMs (enabled) |
| RT cores | 188 (4th gen) — 380 RT TFLOPS |
| Tensor cores | 752 (5th gen, FP4) |
| Base / boost clock | ~1,590 MHz / ~2,617 MHz (third-party; NVIDIA doesn't publish base clock) |
| Memory | |
| Memory size | 96 GB GDDR7 with ECC |
| Memory bus | 512-bit |
| Memory bandwidth | 1,792 GB/s (~1.8 TB/s) |
| Compute & AI | |
| FP32 (single precision) | 125 TFLOPS |
| RT core performance | 380 TFLOPS |
| AI performance | 4,000 AI TOPS (FP4 with sparsity) |
| Power & Cooling | |
| Total board power | 600 W (fixed) |
| Power connector | 1× 16-pin (12V-2x6 / CEM5) |
| Recommended PSU | ~1,000 W system (third-party guidance) |
| Cooling | Active double flow-through (dual axial) |
| Form Factor & Display | |
| Form factor | Dual-slot, 5.4″ H × 12.0″ L (extended height) |
| Interface | PCIe 5.0 x16 |
| Display outputs | 4× DisplayPort 2.1b (active) |
| Max resolution | Up to 8K @ 240 Hz, 16K @ 60 Hz |
| Media engine | 4× NVENC (9th gen) + 4× NVDEC (6th gen), 4:2:2 support, AV1 |
| Enterprise Features | |
| MIG | Up to 4× 24 GB, 2× 48 GB, or 1× 96 GB |
| AI Management Processor | Yes |
| APIs | DX12 (SM 6.6), OpenGL 4.6, Vulkan 1.3, CUDA, OpenCL 3.0 |
| Availability & Price | |
| Launch | Announced GTC, March 18, 2025; shipping later 2025 |
| Price | MSRP ≈ $8,500 → NVIDIA later listed ≈ $13,250; street ≈ $8,000–$13,000+ |
Real-World Performance
Independent testing (StorageReview, Puget Systems) shows what the 96GB buys you:
- Local LLMs: ran a 120B-parameter model in LM Studio at ~163 tokens/sec — a model the 48GB RTX 6000 Ada simply cannot load. Roughly 2× the Ada in MLPerf-style Llama2 text generation (8,008 vs 3,957).
- 3D rendering: Blender "Monster" scene at 7,870 samples/min vs 5,632 on the RTX 6000 Ada (~40% faster), with strong multi-GPU scaling in OctaneRender.
- AI image: Stable Diffusion 1.5 FP16 at 0.705 s/image vs 1.477 s on Ada (~2×).
- Video / 8K: four 9th-gen NVENC engines with 4:2:2 hardware accelerate DaVinci Resolve and camera-native footage; ~23% faster than Ada in Topaz Video AI.
RTX PRO 6000 Workstation vs RTX 5090 — Which Should a Pro Buy?
Both are GB202. For gaming-rate FP32 the RTX 5090 is in the same ballpark and far cheaper — but the pro card wins on everything production-critical:
| Attribute | RTX PRO 6000 WS | GeForce RTX 5090 |
|---|---|---|
| VRAM | 96 GB GDDR7 | 32 GB GDDR7 |
| ECC memory | Yes | No |
| AI performance | ~4,000 AI TOPS | ~3,352 AI TOPS |
| Cooling exhaust | Double flow-through (vents out of case) | Open-air |
| Drivers | Certified enterprise / ISV | GeForce Game Ready |
| Data-center use | Permitted | Restricted by EULA |
| Price | ≈ $8,500+ | ≈ $1,999 MSRP |
The decision is simple: if your workload exceeds 32GB of VRAM (large models, huge scenes, long-context inference) the PRO 6000 is the only single-card answer. If it fits in 32GB, the RTX 5090 delivers most of the throughput for a quarter of the price.
What Creators & AI Devs Ask
- "Is 96GB worth 4–6× the price of a 5090?" Only if you actually exceed 32GB — then it's not a luxury, it's the requirement.
- "How big a model fits in 96GB?" A 70B model in FP8, or very large quantized models with long context, entirely in VRAM.
- "vs the RTX 6000 Ada?" Roughly 2× the AI performance, ~40% faster in Blender, and double the VRAM (96 vs 48GB).
- "Will it fit my workstation?" It's 12″ long, dual-slot, extended-height, draws 600W and uses a 16-pin connector — you need a full-size tower and a ~1,000W PSU.
- "Certified drivers?" Yes — ISV-certified enterprise driver branch for CAD/DCC apps, unlike the GeForce line.
Pros and Cons
✅ Pros
- 96GB GDDR7 with ECC — class-leading single-card capacity
- ~1.8 TB/s bandwidth and full GB202 (24,064 CUDA / 752 Tensor / 188 RT)
- 4,000 AI TOPS — ~2× the RTX 6000 Ada in AI workloads
- Active double flow-through cooling vents heat out of the case
- Certified enterprise drivers, ISV certifications, MIG, AI Management Processor
- 4× DP 2.1b (8K@240 / 16K@60) and 4× NVENC with 4:2:2 for pro video
- Drops into a standard tower (unlike the headless Server Edition)
❌ Cons
- Very expensive — ~$8,500 MSRP, frequently $10k–13k+ at retail
- 600W draw demands a high-wattage PSU and good airflow
- 12″ length / extended height won't fit smaller cases
- Single 16-pin connector carrying 600W — seat the cable carefully
- Overkill (and poor value) if your work fits in 32GB — get an RTX 5090
The Bottom Line
For professionals who genuinely need 96GB on a single card, the RTX PRO 6000 Blackwell Workstation Edition is essentially the only choice — and a very good one. It's close to category-defining: unmatched capacity, ECC reliability, production-grade drivers and cooling, and a clean generational leap over the RTX 6000 Ada. The only real knocks are steep, volatile pricing and 600W power/cooling demands. Our score: 92/100.
Related reading
Frequently Asked Questions
How much VRAM does the RTX PRO 6000 Blackwell Workstation Edition have?
96 GB of GDDR7 with ECC on a 512-bit bus, delivering 1,792 GB/s (about 1.8 TB/s) of bandwidth. That's enough to hold a 70-billion-parameter model in FP8, billion-polygon scenes, or large quantized models with long context entirely in VRAM.
RTX PRO 6000 Workstation Edition vs RTX 5090 — what's the difference?
Both use the GB202 chip, but the PRO 6000 has 96GB of ECC memory (vs 32GB non-ECC), ~4,000 AI TOPS, certified enterprise drivers, double flow-through cooling, and is permitted for data-center use. The RTX 5090 is far cheaper and great if your work fits in 32GB; the PRO 6000 is for workloads that don't.
What power supply does the RTX PRO 6000 Workstation need?
The card has a fixed 600W board power and a single 16-pin (12V-2x6) connector, so a roughly 1,000W system PSU is recommended, along with a full-size tower that can fit the 12-inch, dual-slot, extended-height card.
Is the RTX PRO 6000 Workstation Edition good for local LLMs?
Yes — it's one of the best single-card options for local AI. In testing it ran a 120B-parameter model in LM Studio at around 163 tokens/sec and delivered roughly 2× the AI text-generation performance of the RTX 6000 Ada.
How much does the RTX PRO 6000 Workstation Edition cost?
It launched around $8,500 MSRP in 2025. Pricing has been volatile since — NVIDIA later listed it near $13,250, and street prices commonly range from about $8,000 to $13,000+ depending on supply and reseller.
