NVIDIA RTX PRO 6000 Blackwell Server Edition
The RTX PRO 6000 Blackwell Server Edition puts 96 GB of GDDR7 and fifth-generation Tensor Cores into a standard PCIe card. We size, build, install and support RTX PRO servers in the UK, from a four-GPU 2U node to a 32-GPU rack.
The RTX PRO 6000 Blackwell Server Edition is NVIDIA's data-centre version of its top professional RTX GPU. It runs large language models and AI agents, and it also has RT Cores, video engines and display outputs for rendering, digital twins and video work. One card type can therefore serve several teams.
It replaces the L40S as NVIDIA's general-purpose PCIe data-centre GPU. Compared with the L40S it doubles the memory (96 GB against 48 GB), nearly doubles the bandwidth, adds FP4 precision and can be split into four isolated GPUs with Multi-Instance GPU (MIG).
| GPU architecture | NVIDIA Blackwell |
|---|---|
| أنوية CUDA | 24,064 |
| Tensor Cores | 752, fifth generation (FP4, FP6, FP8, FP16, BF16, TF32) |
| RT Cores | 188, fourth generation |
| FP4 Tensor Core | 4 PFLOPS |
| FP8 Tensor Core | 2 PFLOPS |
| FP16 / BF16 Tensor Core | 1 PFLOPS |
| TF32 Tensor Core | 234 TFLOPS |
| FP32 | 120 TFLOPS |
| Peak RT Core performance | 355 TFLOPS |
| ذاكرة GPU | 96 GB GDDR7 with ECC |
| Memory interface and bandwidth | 512-bit, 1,597 GB/s |
| Multi-Instance GPU | Up to 4 isolated instances |
| Media engines | 4 × NVENC (9th gen), 4 × NVDEC (6th gen) |
| Display outputs | 4 × DisplayPort 2.1 |
| Host interface | PCIe Gen 5 x16 |
| Maximum power | Up to 600 W, configurable |
| عامل الشكل | Air cooled: dual slot, full height, full length, passive. Liquid cooled: single slot |
| الأمان | Secure boot with hardware root of trust |
Figures from NVIDIA's product page and datasheet. Tensor Core figures are NVIDIA's peak numbers; real throughput depends on the model, precision and software.
96 GB holds a 70-billion-parameter model in FP8 on one card, or a much larger one across four or eight. FP4 roughly halves the memory again for models that tolerate it.
Omniverse and OpenUSD scenes, synthetic data for robotics and vision, and simulation, on the same GPUs as the models that use them.
RT Cores, DLSS 4 and neural shaders for final-frame rendering, plus four encoders and four decoders for video pipelines and streaming.
120 TFLOPS of FP32 for simulation and engineering codes, and GPU-accelerated data science with RAPIDS.
Share GPUs between designers and engineers with MIG or NVIDIA virtual GPU software, so CAD and visualisation users don't each need a tower.
Run inference by day and rendering or fine-tuning by night on the same servers, instead of buying separate fleets.
RTX PRO 6000 cards go into standard PCIe servers. These are the three sizes we're asked for most.
384 GB of GPU memory in a 2U chassis that fits an existing rack. Good for a department's private assistant, retrieval-augmented generation (RAG) over company documents, or a first production inference service.
About 3 to 3.5 kW at full GPU power
768 GB of GPU memory. MGX servers such as the AMAX AceleMax AXG-428AG pair eight dual-slot GPUs with two AMD EPYC 9005 processors, up to five PCIe Gen 5 x16 slots for network cards and eight E1.S drives.
About 6 to 7 kW at full GPU power
Four eight-GPU servers with Spectrum-X Ethernet or Quantum InfiniBand between them: about 3 TB of GDDR7 and 128 PFLOPS of FP4 in one rack. AMAX sells this design as the RackScale 32. It suits inference at scale, multi-agent systems and large rendering farms.
Plan for 25 to 30 kW per rack
| RTX PRO 6000 Blackwell Server Edition | L40S | |
|---|---|---|
| المعمارية | Blackwell | Ada Lovelace |
| ذاكرة GPU | 96 GB GDDR7 ECC | 48 GB GDDR6 ECC |
| عرض نطاق الذاكرة | 1,597 GB/s | 864 GB/s |
| أنوية CUDA | 24,064 | 18,176 |
| Tensor Cores | 752 (5th gen) | 568 (4th gen) |
| RT Cores | 188 (4th gen) | 142 (3rd gen) |
| FP32 | 120 TFLOPS | 91.6 TFLOPS |
| FP4 | Yes | No |
| MIG | Up to 4 instances | No |
| PCIe | Gen 5 x16 | Gen 4 x16 |
| Maximum power | Up to 600 W | 350 W |
| Video engines | 4 NVENC / 4 NVDEC | 3 NVENC / 3 NVDEC |
The L40S is still a sensible choice where power is tight or models fit in 48 GB. Where they don't, the RTX PRO 6000 usually needs fewer GPUs for the same job.
We work out GPU memory, users and storage from your models, context lengths and traffic, and pick the smallest system that does the job.
A written quote with the exact configuration, the lead time and the power and cooling it needs.
We install the operating system, NVIDIA data-centre drivers, CUDA, the container toolkit and your inference or rendering stack, then run a burn-in under full load.
We deliver to your site or data centre, rack and cable it, and support it from the UK, with the manufacturer's warranty behind the hardware.
No. Both have 96 GB of GDDR7, but the Server Edition is passively cooled and relies on the server's fans, with a configurable power limit up to 600 W, and comes in a single-slot liquid-cooled version too. The Workstation Edition has its own fans for desktop towers, and the Max-Q Workstation Edition runs at 300 W.
Roughly 70 billion parameters in FP8, or about twice that in FP4, leaving some memory for the KV cache. Long contexts and many users need more memory, so we size from your real context lengths and concurrency.
No. GPUs talk to each other over PCIe Gen 5. That's fine for running many copies of a model, or one model split across a few GPUs. For large training jobs that need NVLink, look at HGX B200 or B300 systems.
Yes. MIG splits a card into up to four isolated GPUs, each with its own memory and compute. NVIDIA's virtual GPU software is another option for virtual desktops.
It depends on GPU supply and the server you choose. Each product in the shop shows our current lead time, and your quote confirms it before you order.
NVIDIA, the NVIDIA logo, RTX, RTX PRO, Blackwell, CUDA, NVLink, MGX, Omniverse, Spectrum-X, Quantum and DLSS are trademarks and/or registered trademarks of NVIDIA Corporation in the US and other countries. AMAX and AceleMax are trademarks of AMAX Engineering Corporation. AMD and EPYC are trademarks of Advanced Micro Devices, Inc. All other trademarks belong to their respective owners. Workstation describes these products as an independent supplier; no endorsement by their owners is implied. The product video is NVIDIA's.