Server-supplied cooling
The heatsink has no onboard fan. Cooling comes from the host server, so validated airflow is essential even when the GPU is idle. Passive describes the card, not a fanless server.
PERCEPTA / PROFESSIONAL GRAPHICS
NVIDIA RTX PRO 6000 Blackwell
Server Edition 96GB
96 GB of ECC graphics memory. NVIDIA Blackwell architecture. Purpose-built for professional workloads.
Product preview
Discuss this GPU
Explore the specification
Pricing and availability to be confirmed. Not currently available to order or pre-order.
Manufacturer reference imagery. GPU card only; not a complete system.
Server Edition
Built for a qualified server environment.
The passive air-cooled Server Edition brings 96 GB of ECC GDDR7 memory to a dual-slot PCIe card for AI, rendering and supported virtualised GPU deployments. This listing covers the air-cooled board, not the liquid-cooled variant. Its host system must provide the power delivery, ducted airflow and software environment the card needs.
96 GBGDDR7 memory with ECC
1,597 GB/sPublished memory bandwidth
Up to 600 WConfigurable board power
PCIe 5.0x16 host interface
Partitioning and virtualisation
Up to four hardware-isolated MIG instances are supported. NVIDIA vGPU is also supported on the Server Edition, subject to compatible hosts, drivers, profiles and the appropriate software licensing.
A 96 GB compute resource
A single GPU with a 512-bit memory interface and 1,597 GB/s published memory bandwidth. Application fit, model size and concurrent workloads should be assessed against the actual deployment rather than capacity alone.
Technical specification
Manufacturer reference / checked 11 Sep 2026| Edition | Server Edition - passive air-cooled |
|---|---|
| GPU architecture | NVIDIA Blackwell / GB202 |
| CUDA compute capability | 12.0 |
| CUDA cores | 24,064 |
| Tensor cores | 752 / fifth generation |
| RT cores | 188 / fourth generation |
| Graphics memory | 96 GB GDDR7 with ECC |
| Memory interface | 512-bit |
| Memory bandwidth | 1,597 GB/s |
| Host interface | PCIe 5.0 x16 |
| Total board power | Up to 600 W; configurable, system and cable requirements apply |
|---|---|
| Cooling | Passive air-cooled; requires qualified ducted server airflow |
| Form factor | Full-height full-length, dual-slot |
| Nominal dimensions (height x length) | 4.4 x 10.5 inches; additional installation clearance required |
| Auxiliary power | 1 x PCIe CEM5 16-pin |
| Display connectors | 4 x DisplayPort 2.1b |
| Display operation | Disabled in factory-default display-off mode; display-on requires supported configuration |
| NVLink bridge | Not supported |
| Frame synchronisation | Supports separate NVIDIA RTX PRO Sync hardware; not included by this listing |
| FP32 compute | 120 TFLOPS |
|---|---|
| RT core throughput | 355 TFLOPS |
These are manufacturer-published peak figures, not measured application benchmarks or guarantees. Throughput varies with software, workload, power configuration and operating conditions. Different precision and RT metrics are not interchangeable.
| Video encoding | 4 x ninth-generation NVENC engines |
|---|---|
| Video decoding | 4 x sixth-generation NVDEC engines |
| JPEG engines | 4 |
| Video processing | AV1 support; 4:2:2 encode/decode support with compatible software |
| Multi-Instance GPU (MIG) | Up to 4 instances; supported profiles include 4 x 24 GB, 2 x 48 GB or 1 x 96 GB |
| NVIDIA vGPU technology | Supported, with compatible deployment and appropriate licensing |
| Confidential computing | Supported; compatible system and software required |
|---|---|
| Secure boot / root of trust | Supported |
Nominal card dimensions are not shipping dimensions or a guaranteed maximum installation envelope. MIG profile, driver and system requirements apply; memory partitioning is not automatic multi-GPU memory pooling.
Will it fit my workstation or server?
Provide the server make and model, GPU support matrix, riser layout, power-cable specification, cooling configuration and intended OS or hypervisor. Follow NVIDIA and the system vendor's engineering guidance. The configurable power limit is up to 600 W; the supported setting depends on the board, firmware and power cabling.
Which RTX PRO 6000 edition should I choose?
Choose the passive Server Edition for a validated server platform and, where required, licensed NVIDIA vGPU. Workstation and Max-Q cards have their own active cooling and do not support NVIDIA vGPU. This page does not describe the single-slot liquid-cooled Server Edition.
What should I know about MIG and software licences?
NVIDIA vGPU on Server Edition needs a supported platform and the appropriate licence for the selected use: for example RTX Virtual Workstation for Q-series profiles or NVIDIA AI Enterprise for vGPU for Compute. A GPU purchase should not be assumed to include either subscription. MIG and vGPU setup must be checked against the intended host, guest, driver and profile.
What is included, and can I pre-order?
This is a product preview for one GPU card, not a complete computer. We are not taking orders or pre-orders for this item yet. Price, exact supplied board and part number, condition, accessories, warranty terms and delivery timing will be confirmed before it is offered for sale.
Manufacturer imagery illustrates the edition, not stock currently held by Percepta. No software subscription, Sync board, power adapter or other accessory is promised by this listing.
Manufacturer documentation and specification notes
Hardware details are checked against the references below. Specifications describe the named edition; supplier-specific commercial details remain subject to confirmation. Reference images are credited to NVIDIA/PNY.
- NVIDIA edition-specific datasheet (PDF)
- NVIDIA passive-air Server engineering brief (PDF)
- PNY edition-specific hardware reference
- NVIDIA supported MIG GPUs
- NVIDIA MIG setup and display-mode requirements
- NVIDIA vGPU support and limitations
- NVIDIA vGPU profiles and licensing
- NVIDIA AI Enterprise compute profiles
Planning a workstation or server build? Send us your system details and intended workload so we can discuss the appropriate edition.
Discuss your requirement ↗