The NVIDIA RTX PRO 6000 Blackwell Server Edition is the data-center member of NVIDIA's RTX PRO 6000 Blackwell family (which also includes Workstation and Max-Q editions). It pairs the full-fat Blackwell GB202 GPU with a massive 96 GB of GDDR7 ECC memory in a passively cooled, server-optimized card designed to slot into validated OEM racks from Dell, HPE, Lenovo, Supermicro and Cisco. It is, in effect, the modern successor to NVIDIA's L40 / L40S / A40 line.
Quick verdict: An outstanding universal data-center GPU for AI inference, rendering, simulation and virtualization. Its 96GB pool, FP4 throughput and MIG/vGPU/Confidential Computing make it a clear, much faster L40S replacement. It is not a large-scale training replacement for HBM-based HGX parts (no NVLink, GDDR7 bandwidth is lower than HBM). Score: 88/100.
RTX PRO 6000 Server Edition — Full Specifications
Cross-checked against NVIDIA's data-center page, the Lenovo Press OEM product guide, ServeTheHome and TechPowerUp. Items NVIDIA does not officially publish for this SKU are flagged.
| Specification | Detail |
|---|---|
| GPU & Architecture | |
| GPU chip | NVIDIA GB202 (Blackwell) |
| Process node | TSMC 4N-class (4nm) — flag: NVIDIA doesn't print an exact node |
| CUDA cores | 24,064 |
| RT cores | 188 (4th gen) — ~355 RT TFLOPS |
| Tensor cores | 752 (5th gen) |
| Boost clock | ~2,617 MHz (third-party estimate — not officially published) |
| Memory | |
| Memory size | 96 GB GDDR7 with ECC |
| Memory bus | 512-bit |
| Memory bandwidth | up to ~1,597 GB/s (~1.6 TB/s) — lower than Workstation/Max-Q |
| AI & Compute (Tensor) | |
| FP4 (sparse) | ~4 PFLOPS |
| FP8 | ~2 PFLOPS |
| FP16 / BF16 | ~1 PFLOP |
| TF32 | 234 TFLOPS |
| FP32 (single precision) | 120 TFLOPS |
| Power & Cooling | |
| Total board power | Configurable, up to 600 W |
| Cooling | Passive (server front-to-back airflow); liquid SKU also offered |
| Power connector | 16-pin (12V-2x6 / CEM5), server cabling |
| Form Factor & Connectivity | |
| Form factor | Dual-slot, Full-Height Full-Length, ~4.4″ × 10.5″ |
| Interface | PCIe Gen 5 x16 |
| Display outputs | 4× DisplayPort 2.1b — disabled by default (effectively headless) |
| NVLink | Not supported (PCIe peer-to-peer only) |
| Media engine | 4× NVENC (9th gen), 4× NVDEC (6th gen), 4× JPEG decoders, AV1 |
| Enterprise Features | |
| MIG (Multi-Instance GPU) | Up to 4 isolated instances of 24 GB each |
| vGPU | Yes — NVIDIA vGPU software (multi-tenant virtual workstations) |
| Confidential Computing | Yes — hardware TEE for secure AI |
| Availability & Price | |
| Launch | Announced GTC, March 18, 2025; broad server availability 2H 2025 |
| Price | MSRP ≈ $8,565 → reportedly raised to ≈ $13,250; street ≈ $8,500–$15,000 (volatile) |
Server vs Workstation vs Max-Q — Same Silicon, Different Build
All three RTX PRO 6000 Blackwell variants share the same GB202 die, 24,064 CUDA cores, 752 Tensor cores, 188 RT cores and 96GB GDDR7 ECC. They differ in power, cooling and deployment:
| Attribute | Server Edition | Workstation Edition | Max-Q Edition |
|---|---|---|---|
| Cooling | Passive (chassis airflow) | Active dual flow-through | Active blower |
| Power | Configurable, up to 600 W | 600 W (fixed) | 300 W |
| Memory bandwidth | ~1,597 GB/s | 1,792 GB/s | 1,792 GB/s |
| Display outputs | Disabled by default (headless) | 4× DP 2.1b, active | 4× DP 2.1b, active |
| Target | Validated OEM servers / dense racks | Desk-side tower workstation | Density-limited multi-GPU |
The Server Edition is the only variant with no onboard fans — the host server is its cooling system. An HPE ProLiant DL380a Gen12 (4U) can hold up to eight of these cards. Note the one real raw-spec trade-off: its memory bandwidth (~1.6 TB/s) is slightly lower than the Workstation/Max-Q editions (1.79 TB/s) — a detail many spec aggregators get wrong.
Target Workloads & Performance
- AI inference & serving: LLMs and multimodal/agentic AI — NVIDIA claims up to 5× the LLM inference throughput of the L40S. The 96GB pool fits large models without sharding.
- Rendering & video: over 2× faster rendering and ~3.3× faster text-to-video vs L40S; four NVENC/NVDEC engines for distributed render farms.
- Virtualization / VDI: vGPU + MIG (4× 24GB) for multi-tenant virtual workstations and isolated workloads.
- Scientific computing: up to ~6.8× faster Smith-Waterman (genomics) vs L40S; drug discovery, CFD, data analytics.
- Fine-tuning / smaller training: large VRAM helps, but with no NVLink, multi-GPU scaling relies on PCIe Gen 5 — better for inference-dense and single/few-GPU jobs than massive distributed pretraining (where H100/H200/B200 dominate).
What IT Buyers Ask
- "Server vs Workstation Edition — can I just buy the cheaper one?" No — the passively cooled Server Edition only works inside a validated server with sufficient airflow; the Workstation Edition has its own fans for a tower.
- "How does it compare to H100 / L40S?" It's a strong, cheaper-per-GB L40S successor for inference, graphics and virtualization — positioned below HBM-based HGX parts for large-scale training.
- "How many MIG instances?" Up to four isolated 24GB instances, ideal for Kubernetes/GPU-sharing and multi-tenant setups.
- "Does it have display outputs?" Physically yes (4× DP 2.1b) but disabled by default — treat it as headless for remote/data-center use.
- "What about power and density?" up to 600W configurable lets operators tune density vs performance (e.g., 8× per 4U chassis).
- "vGPU licensing?" vGPU software (and its licensing cost) is required for VDI/virtual-workstation deployments.
Pros and Cons
✅ Pros
- 96GB GDDR7 ECC — fits large models and scenes on a single card
- Strong FP4/FP8 inference (~4/2 PFLOPS) — big gains over L40S
- Passive + configurable up to 600W enables high rack density (up to 8 per 4U)
- MIG (4×24GB), vGPU and Confidential Computing for secure multi-tenant use
- Broad OEM validation (Dell, HPE, Lenovo, Supermicro, Cisco) and PCIe Gen 5
- One universal SKU for AI, rendering, simulation and VDI
❌ Cons
- No NVLink — limits high-bandwidth multi-GPU scaling for large training
- ~1.6 TB/s GDDR7 bandwidth is below HBM rivals (H100 ~3.35 TB/s)
- Requires a validated server with strong airflow — not a generic chassis
- Display outputs disabled by default (headless)
- High and volatile pricing; vGPU adds licensing cost
The Bottom Line
Judged as a data-center product, the RTX PRO 6000 Blackwell Server Edition is excellent: category-leading VRAM-per-dollar, superb inference and graphics-plus-AI versatility, dense passive deployment, and the full enterprise feature stack (MIG, vGPU, Confidential Computing). It loses a few points for the missing NVLink, GDDR7 bandwidth that trails HBM parts on the most bandwidth-bound workloads, and volatile pricing. For inference, rendering and virtualization racks it's a standout. Our score: 88/100.
Related reading
Frequently Asked Questions
What is the difference between the RTX PRO 6000 Server and Workstation editions?
They use the same GB202 GPU and 96GB GDDR7 ECC memory, but the Server Edition is passively cooled (relies on server chassis airflow), has configurable up to 600W power, slightly lower ~1.6 TB/s bandwidth, and disabled display outputs. The Workstation Edition has active fans, a fixed 600W, full 1.79 TB/s bandwidth, and active display outputs for a desk-side tower.
How much memory does the RTX PRO 6000 Server Edition have?
96 GB of GDDR7 with ECC on a 512-bit bus, delivering up to about 1,597 GB/s (1.6 TB/s) of bandwidth. It can be partitioned into up to four isolated 24GB MIG instances.
Does the RTX PRO 6000 Server Edition support NVLink?
No. This SKU does not support NVLink — multi-GPU communication is over PCIe Gen 5 peer-to-peer only. For large-scale distributed training, NVIDIA's HBM-based HGX H100/H200/B200 platforms remain the better fit.
How much does the RTX PRO 6000 Blackwell Server Edition cost?
It launched around an $8,565 MSRP in March 2025. Pricing has since been volatile — Tom's Hardware reported NVIDIA raised it toward $13,250, and some retailers list it as high as ~$15,000. Buy through validated OEM channels for accurate quotes.
Can I use the RTX PRO 6000 Server Edition in a normal PC?
Not practically. It has no fans and relies on server front-to-back airflow, and its display outputs are disabled by default. It is designed for validated OEM servers from Dell, HPE, Lenovo, Supermicro and Cisco — for a tower workstation, choose the Workstation Edition instead.
