NVIDIA RTX PRO 6000 Blackwell
Eight RTX PRO 6000 Blackwell GPUs with 96 GB of GDDR7 each, 768 GB per server, served from Montreal for LLM inference, generative video and image, rendering and media pipelines.
- GPUs per server
- 8
- Memory per GPU
- 96 GB GDDR7
- GPU memory per server
- 768 GB
- Served from
- Montreal
Specifications
| GPU | |
|---|---|
| Edition | NVIDIA RTX PRO 6000 Blackwell Server Edition |
| Architecture | NVIDIA Blackwell |
| Memory | 96 GB GDDR7 with ECC, 512-bit |
| Memory bandwidth | 1,597 GB/s |
| CUDA cores | 24,064 |
| Tensor Cores | 752, fifth generation |
| RT Cores | 188, fourth generation |
| Peak FP4 | 4 PFLOPS (NVIDIA does not state sparse or dense) |
| FP32 | 120 TFLOPS |
| RT Core performance | 355 TFLOPS |
| Media engines | 4 NVENC, 4 NVDEC, 4 JPEG |
| Multi-Instance GPU | Up to 4 instances of 24 GB, 2 of 48 GB or 1 of 96 GB |
| GPU interconnect | PCIe 5.0 x16, no NVLink |
| Power per GPU | Configurable, up to 600 W |
| Form factor | Dual slot, full height, full length, passive cooling |
| Server as delivered | |
|---|---|
| GPUs per server | 8 |
| GPU memory per server | 768 GB (8 x 96 GB) |
| CPU, system memory and storage | Configured to your requirements and confirmed in your proposal |
| Network | Configured to your requirements and confirmed in your proposal |
| Location | Montreal, Canada |
Typical workloads
- LLM inference for 70B-class models in FP8 or FP4
- Generative video and image pipelines such as ComfyUI
- Rendering, VFX and virtual production
- Video transcoding and streaming
Rent dedicated servers with eight NVIDIA RTX PRO 6000 Blackwell GPUs, served from Montreal. Allocation is open, on 36, 48 or 60 month contracts, with every price in a written proposal.
Who the RTX PRO 6000 Blackwell is for
The RTX PRO 6000 Blackwell is for teams that need large memory per GPU at up to 600 W rather than HGX-class power: inference teams serving 70B-class models in FP8 or FP4, generative video and image studios, and media pipelines that render, encode and stream.
Each GPU has 96 GB of GDDR7 with ECC, 768 GB per server. A 70B-parameter model's FP8 weights take about 70 GB, so the model fits on one GPU with room for its KV cache, and a 4-bit version of the same model leaves about 60 GB free.
What it does well
96 GB per GPU
96 GB of GDDR7 at 1,597 GB/s on a 512-bit bus, and up to four MIG instances of 24 GB per GPU when jobs are small.
Blackwell Tensor Cores
752 fifth-generation Tensor Cores with FP4 support. NVIDIA prints 4 PFLOPS of peak FP4 per GPU, without saying whether it is sparse or dense, and 120 TFLOPS of FP32.
Media engines
Four NVENC, four NVDEC and four JPEG engines per GPU, 32 hardware encoders per server, for transcoding and video pipelines.
Ray tracing
188 fourth-generation RT Cores and 355 TFLOPS of RT Core performance for rendering, VFX and simulation.
The GPUs connect over PCIe 5.0 x16, without NVLink. Models that fit on one GPU, or that split cleanly across a few, suit the RTX PRO 6000; models that need heavy GPU-to-GPU traffic belong on an HGX server.
Configuration as delivered
Servers are configured to your requirements and confirmed in your proposal: CPU, system memory, local NVMe and network. The figures on this page are for the Server Edition, NVIDIA's passively cooled data center version of the card, configurable up to 600 W per GPU.
Our Montreal site has onsite engineering support. Every server is burned in before handover, with the results in an acceptance report, and our burn-in and acceptance page explains the tests.
How renting works
RTX PRO 6000 Blackwell allocation is open in Montreal. You rent whole 8-GPU servers on a 36, 48 or 60 month contract, and longer terms carry lower rates.
The company that will run the servers completes three onboarding forms before signing. You pay a down payment at signing, then one fixed monthly rate for the full term, with no egress fees, which matters when you move large video files in and out. We do not offer trials; burn-in and the acceptance report take their place. Montreal is a Canadian site and 1Legion is a Canadian company: see GPU servers in Canada, and our how to buy page for each step.
Workloads and alternatives
- LLM inference: 70B-class models in FP8 or FP4, one model per GPU or split across a few. See AI inference.
- Generative video and image: ComfyUI and similar diffusion pipelines, where memory decides resolution, length and batch size. See generative video.
- Rendering, VFX and transcoding: see media and entertainment.
For sizing, read 8 x 96 GB: what fits on an RTX PRO 6000 Blackwell node and ComfyUI and AI video on dedicated bare metal. If 48 GB per GPU is enough, the L40S is available on request; for the largest models, the HGX B300 has 270 GB per GPU.
Questions buyers ask
What fits on one RTX PRO 6000 Blackwell GPU?
96 GB holds a 70B-parameter model in FP8, about 70 GB of weights, with room left for the KV cache. Our article 8 x 96 GB: what fits on an RTX PRO 6000 Blackwell node works through the sizing.
Which edition of the RTX PRO 6000 Blackwell do you use?
The figures on this page are for the Server Edition, NVIDIA's passively cooled data center version, configurable up to 600 W per GPU. The edition and power setting of your servers are confirmed in your proposal.
Are the GPUs connected with NVLink?
No. The RTX PRO 6000 Blackwell Server Edition connects over PCIe 5.0 x16. If your model must be split across GPUs that need NVLink bandwidth, look at the HGX B300.
Where are the servers hosted?
In Montreal, Canada, where we have onsite engineering support. When you choose a Canadian site, the servers and their storage are located in Canada.
Can we try a server before signing?
We do not offer trials. Every server is burned in before handover, and you receive the results in an acceptance report. See burn-in and acceptance.
What contract terms are available?
36, 48 or 60 months, with lower rates for longer terms. You pay a down payment at signing, then one fixed monthly rate for the full term, with no egress fees.
Rent RTX PRO 6000 servers on 36 to 60 month terms.
Share the hardware, the number of servers, the term and the start date. You receive a written proposal built on our delivery specification.