96 GB
GDDR7 memory with ECC
Twice the 48 GB of the RTX 6000 Ada and more than the 80 GB A100 and H100 we fit. ECC corrects single-bit errors in GPU memory, which matters on runs that last days.
GPU dedicated servers · NVIDIA Blackwell · by the month
RTX PRO 6000 dedicated server96 GB on one GPU. The whole machine is yours.
An RTX PRO 6000 dedicated server from us is a bare-metal machine with an NVIDIA RTX PRO 6000 Blackwell passed straight through to your own operating system. The complete server starts at $1,382.67 a month. The GPU carries 96 GB of GDDR7 with ECC. No GPU in our order catalogue holds more.
- Full GPU, no vGPU
- Unmetered port
- 4 cities
- No contract
You get root on the machine, the GPU on the PCIe bus, and the driver and CUDA version you choose. The price is the processor, memory, disks, the port and an IPv4 address, plus the GPU as its own monthly line. Change any of it in the configurator.
The GPU
What the RTX PRO 6000 Blackwell gives a server
It is NVIDIA's professional Blackwell GPU, the successor to the RTX 6000 Ada. Four parts of it decide what the server is good for.
752
5th-generation Tensor cores
The Tensor cores run FP4, FP8 and FP16 matrix maths for inference and training. NVIDIA rates the GPU at 4,000 AI TOPS in FP4 with sparsity.
188
4th-generation RT cores
The RT cores do the ray and triangle tests in hardware, rated at 380 TFLOPS. OptiX renderers such as Blender Cycles, V-Ray and Arnold GPU use them.
4 + 4
NVENC and NVDEC engines
Four 9th-generation encoders and four 6th-generation decoders, all with AV1, beside the compute cores. The A100 and H100 have no video encoder at all.
Against the accelerators
RTX PRO 6000 vs H100, A100 and the 48 GB GPUs
These are the GPUs with 40 GB or more that we fit, each priced as a complete server. The RTX PRO 6000 holds more memory than the H100 and costs $300 a month less as a complete server.
What it gives up is NVLink. Two of them in one machine are two separate 96 GB pools that exchange data over PCIe. The A100 and H100 carry an NVLink bridge connector, and a build with a bridge is quoted by hand on A100 or H100 servers. The A100 80 GB also reads its memory a little faster, 1,935 GB/s against 1,792 GB/s, which shows when you serve many users at a large batch size.
- RTX PRO 6000 Blackwell 96GB96 GB GDDR7 ECC$1,382.67/mo complete
- A100 80GB80 GB HBM2e$972.47/mo complete
- H100 80GB80 GB HBM2e$1,682.67/mo complete
- Instinct MI21064 GB HBM2e$582.67/mo complete
- RTX 800048 GB GDDR6$472.47/mo complete
- L40S48 GB GDDR6 ECC$532.67/mo complete
- A4048 GB GDDR6$572.47/mo complete
- RTX A600048 GB GDDR6 ECC$622.47/mo complete
- RTX 6000 ADA48 GB GDDR6 ECC$698.47/mo complete
- A100 40GB40 GB HBM2$772.47/mo complete
Platforms
Which machine the GPU goes in, and the PCIe link it gets
The RTX PRO 6000 has a PCIe Gen 5 interface. Our Intel Xeon Silver and Gold machines run it at PCIe 3.0; the AMD EPYC machine runs it at PCIe 4.0. Once the weights or the scene sit in GPU memory the link is mostly idle. It matters for training loops that copy a fresh batch from system memory on every step, so take the EPYC machine for those.
| Machine | PCIe link to the GPU | Cheapest build | With the GPU | Cities | Order |
|---|---|---|---|---|---|
| Intel Xeon Silver / GoldIntel Xeon Silver 4110, 8 cores, 2.10 GHz · 16 GB DDR4 · 2 × SATA-SSD 240 GB (RAID 1) · 1 Gbps | PCIe 3.0 | $83.67/mo | $1,382.67/mo | New York, Miami, San Francisco and Bucharest | Configure Intel Xeon Silver / Gold with the RTX PRO 6000 |
| AMD EPYCAMD EPYC 7413, 24 cores, 2.65 GHz · 32 GB DDR4 · 1 × SATA-SSD 240 GB · 1 Gbps | PCIe 4.0 | $217.59/mo | $1,516.59/mo | New York and Bucharest | Configure AMD EPYC with the RTX PRO 6000 |
Each price is the cheapest build of that machine as the configurator opens it, and the button opens it with the GPU already selected. More processor, memory, disks and a faster port are priced there. Bucharest builds are billed in euros. Amsterdam machines are not offered with this GPU. More than one RTX PRO 6000 in one machine is quoted by hand: ask on chat or see multi-GPU servers.
Sizing
What fits in 96 GB of GPU memory
A model's weights take its parameter count times the bytes per parameter. The GPU needs a quarter to a half more on top for the KV cache and the runtime, and the table below applies that rule. A 70B model at 4-bit fits with room for context; at 8-bit it loads, but a long context or a second user runs it out.
| Model size (weights) | 4-bit | 8-bit | 16-bit |
|---|---|---|---|
| 7-8 billion parameters | Fits5 GB | Fits8 GB | Fits16 GB |
| 13-14 billion parameters | Fits8 GB | Fits14 GB | Fits28 GB |
| 30-34 billion parameters | Fits19 GB | Fits34 GB | Tight68 GB |
| 70 billion parameters | Fits38 GB | Tight70 GB | Does not fit140 GB |
Weights only, rounded up. For a plan model by model, with the machine it needs, see LLM servers.
Workloads
What an RTX PRO 6000 server runs, and where it stops
LLM inference on one GPU
Serve a 70B model at 4-bit, or a 30B model at 8-bit, behind your own API with vLLM, llama.cpp or Ollama. Prompts and outputs stay on your machine.
Fine-tuning
QLoRA on a 70B model, or LoRA on a 13B model at 16-bit, fits in 96 GB. ECC memory guards a run of several days against silent bit flips.
Rendering and visualisation
Scenes that do not fit in 24 or 48 GB render on one GPU instead of being split. The RT cores handle the ray tracing.
Video at volume
Four AV1 encoders and four decoders transcode many live channels in hardware while the CUDA cores run something else. For a machine set up with Flussonic or Wowza, see GPU transcoding servers.
- No NVLink. GPUs in one machine share data over PCIe only. When two GPUs must trade tensors faster than that, an A100 or H100 build with an NVLink bridge is quoted by hand.
- A PCIe 3.0 or 4.0 slot. Neither machine that takes it has a Gen 5 slot, so the GPU runs at a Gen 3 or Gen 4 link. Inference and rendering barely notice; heavy host-to-GPU traffic does.
- Not for training a large model from scratch. That takes many GPUs joined by a fast fabric. One RTX PRO 6000 is for inference, fine-tuning and single-GPU training.
Delivery and terms
How soon it is online, and what the monthly price covers
- About 30 minutes
Already in a racked machine
Nothing to fit: the server is installed and handed over. Instant servers lists what is racked right now.
- 4 to 24 hours
A GPU we hold
A technician fits it to the machine you configured, builds the disks and installs the operating system.
- 5 to 10 working days
A GPU we buy in
We order it from a distributor, then fit it.
- 10 to 15 working days
A GPU on allocation
When distributors ration a GPU, you get the date before you commit.
- Included
- Processor, memory, disks, IPv4 and IPv6, and an unmetered port. Nothing counts traffic in either direction.
- Access
- Root or administrator on your own operating system, reinstalled from the control panel at no charge, and the server's out-of-band console on a separate management network.
- Setup
- $29 once for a GPU fitted to order. A racked machine ordered as it stands has no setup charge.
- Term
- One month at a time, no contract. 3, 6 and 12-month cycles save up to 15%.
- Support
- Chat answered in seconds by an AI assistant that can check your account and your server, and a person the moment you ask for one.
Every timeline starts once payment and fraud review clear. You are told which of the four applies before you pay.
RTX PRO 6000 machines on our GPU marketplace
Owners also list RTX PRO 6000 machines on our marketplace, each at a price the owner sets. Listings marked as coming through our sister company, Primcast, are rented on primcast.com.
Specifications
RTX PRO 6000 Blackwell specifications
| Specification | RTX PRO 6000 Blackwell |
|---|---|
| Architecture | NVIDIA Blackwell |
| GPU memory | 96 GB GDDR7 with ECC |
| Memory bandwidth | 1,792 GB/s |
| CUDA cores | 24,064 |
| Tensor cores | 752, 5th generation (FP4) |
| RT cores | 188, 4th generation |
| FP32 | 125 TFLOPS |
| RT performance | 380 TFLOPS |
| AI performance | 4,000 TOPS (FP4, sparse) |
| Video encode | 4 NVENC, 9th generation, AV1 |
| Video decode | 4 NVDEC, 6th generation, AV1 |
| Host interface | PCIe Gen 5 |
| Board power | up to 600 W |
| NVLink | none |
Answers
RTX PRO 6000 server questions
How much does an RTX PRO 6000 dedicated server cost?
A complete RTX PRO 6000 server starts at $1,382.67 a month. That is the GPU at $1,299 plus the cheapest machine that takes it, at $83.67: Intel Xeon Silver 4110, 8 cores, 2.10 GHz · 16 GB DDR4 · 2 × SATA-SSD 240 GB (RAID 1) · 1 Gbps. A GPU fitted to order adds a one-time setup of $29, and paying 12 months ahead takes 15% off.
How many RT cores and Tensor cores does the RTX PRO 6000 have?
188 fourth-generation RT cores and 752 fifth-generation Tensor cores, beside 24,064 CUDA cores. NVIDIA rates the RT cores at 380 TFLOPS and the Tensor cores at 4,000 AI TOPS in FP4 with sparsity. The counts are the same on the Workstation, Max-Q and Server Editions.
Is the RTX PRO 6000 the same GPU as the RTX 6000 Ada?
No. The RTX 6000 Ada is the previous generation, with 48 GB of GDDR6. The RTX PRO 6000 is Blackwell, with 96 GB of GDDR7, fifth-generation Tensor cores and AV1 encoders. We fit both: if 48 GB covers your work, the RTX 6000 Ada is the cheaper server, and GPU servers prices every GPU we fit.
Does the RTX PRO 6000 support NVLink?
No. The RTX PRO 6000 has no NVLink connector, so two in one server are two 96 GB pools that exchange data over PCIe. The A100 and H100 carry an NVLink bridge connector, and a build with a bridge is quoted by hand. So is more than one RTX PRO 6000 in one machine.
Can a 70B model run on one RTX PRO 6000?
Yes. At 4-bit a 70B model's weights take about 38 GB, which leaves room in 96 GB for a long context and several users. At 8-bit they take about 70 GB: it loads, with little room to spare. At 16-bit it needs two GPUs or more.
Is the GPU shared or virtualised?
No. The machine is rented to one account and the full GPU is passed straight through to your operating system. There is no hypervisor and no vGPU profile, and you install the NVIDIA driver and CUDA version you want.
How fast is an RTX PRO 6000 server online?
About 30 minutes after your order is approved when a machine with the GPU is already racked. When we fit a GPU we hold, 4 to 24 hours; one we buy in takes 5 to 10 working days. You are told which before you pay.
Where are your RTX PRO 6000 servers located?
In New York, Miami, San Francisco and Bucharest, in our own racks on our own network. Bucharest builds are billed in euros. The GPU is not offered on Amsterdam machines.
Configure your RTX PRO 6000 server
Pick the machine, the memory, the disks and the city. The configurator prices every change as you make it, with the GPU already selected.
Every GPU we fit · A100 and H100 servers · L40S servers · RTX 5090 servers · Racked right now