Blackwell workstationLive Marketplace Asks Available

NVIDIA RTX PRO 6000

RTX PRO 6000 Blackwell is a 96 GB GDDR7 ECC card with about 1.79 TB/s of bandwidth, 24,064 CUDA cores, and FP4 Tensor Cores. It is a PCIe workstation part: no NVLink. The small host in the menu, 8 vCPUs and 24 GB RAM beside 96 GB of VRAM, matches a single-GPU inference box more than an HGX node.

Browse Live Marketplace →All Hardware Fleet

Flexible Compute Modes

Deploy NVIDIA RTX PRO 6000 Your Way: Docker Pods or Root KVM VPS

NovaPod lets you decide how you want to harness GPU power. Whether you need an instant pre-cached container stack, a full root virtual machine with dedicated public IP, or a multi-node NVLink cluster, pay only for what you run with second-by-second crypto billing.

Docker GPU PodCold Boot < 15s

Fast, Lightweight AI Containers

Pure GPU compute without hypervisor overhead. Boot straight into PyTorch, vLLM, ComfyUI, or JupyterLab.

1-second precision metering
37+ Pre-cached ML framework templates
Zero idle fee when instance is stopped

Deploy as Docker Pod

Root KVM VPSPersistent NVMe

Full Root Virtual Machine

Dedicated operating system (Ubuntu 24.04 LTS / Windows GPU Server) with complete sudo root SSH access from minute one.

Dedicated public IPv4 (all ports open)
Pause anytime; pay only $0.00014/GB/hr for disk
Run Docker-in-Docker, k3s, custom drivers

Deploy as Root VPS

Multi-GPU Cluster900 GB/s NVLink

Supercomputing & Distributed Training

Pool up to 8× NVIDIA RTX PRO 6000 on a unified node fabric with Quantum-2 InfiniBand RDMA for multi-node scale.

Zero-loss all-reduce gradient exchange
PyTorch FSDP, DeepSpeed Zero-3 ready
Native GPUDirect Storage & RDMA fabric

Configure Cluster

Hardware Specifications

Architecture: NVIDIA Blackwell, workstation
VRAM: 96 GB GDDR7 with ECC
Bandwidth: 1.79 TB/s
CUDA cores: 24,064
Tensor Cores: 5th generation, FP4
RT Cores: 4th generation
FP32: About 126 TFLOPS
FP4 dense: About 2,016 TFLOPS
Form factor: Dual-slot PCIe Gen5
NVLink: Not supported
Board power: Up to 600 W, lower Max-Q mode
Reference host: 8 vCPUs, 24 GB RAM
ECC: On
Encode: AV1, HEVC, H.264
Best fit: One GPU, 96 GB, no fabric

96 GB without HBM

RTX PRO 6000 Blackwell is a workstation GPU: 96 GB of GDDR7 with ECC, about 1.79 TB/s, 24,064 CUDA cores, fifth-generation Tensor Cores with FP4, and fourth-generation RT cores. Board power is up to 600 W, with a lower Max-Q mode. It is PCIe Gen5. There is no NVLink.

The menu host is small on purpose: 8 vCPUs and 24 GB of RAM next to 96 GB of VRAM. That is a single-GPU inference or render box. It is not an HGX node and it will not tensor-parallel a 405B model.

The 70B case

Llama-class 70B weights at FP8 are on the order of 70 GB, so they fit with room for a moderate KV cache. A 32B model fits in FP16. AWQ and other 4-bit exports leave even more room. FP16 70B does not fit. That job needs H200 or B200.

Bandwidth is the tax. 1.79 TB/s is far under H100's 3.35 TB/s. Large-batch decode and training steps that stream a lot of data will feel it. Single-stream or modest-batch inference of a model that needed 96 GB is where this card is rational, because you avoided an H100 you would not have filled.

Rendering and media

RT cores and a large framebuffer are why DCC and CAD stacks care about this part. Blender, Maya, 3ds Max, Cinema 4D, V-Ray, and Omniverse-class viewports are the usual targets. ECC matters if a frame that glitches is a customer deliverable.

Encode blocks on this generation include AV1 and high-throughput H.264 and HEVC. Use it for a transcode farm only if the GPUs are actually busy. An idle professional card is still an hourly ask.

What it is not

It is not a multi-GPU training fabric. No NVLink means scale-out is data parallel at best, and PCIe will dominate if you pretend it is an NVSwitch box. Pre-training and 100B+ tensor parallel belong on H100, H200, or B200.

Against the older RTX 6000 Ada, memory goes from 48 GB GDDR6 to 96 GB GDDR7 and you gain FP4. Against a 32 GB 5090, you pay more for ECC and three times the framebuffer. Against L40S, you are choosing Blackwell numerics and 96 GB over a 48 GB datacenter Ada card.

What the Numbers Mean in Production

VRAM: 96 GB GDDR7 ECC
Bandwidth: 1.79 TB/s
70B FP8: One GPU
NVLink: None

Storage & Tooling Infrastructure

High-Speed Storage, Native Ports & Crypto Rails

Every NVIDIA RTX PRO 6000 instance on NovaPod connects directly to high-throughput infrastructure designed to eliminate I/O bottlenecks and onboarding delays.

25 Gb/s Shared Network Storage

Mount large model checkpoints and training sets across one or multiple instances simultaneously. Data outlives single pods and stays accessible when instances stop.

JupyterLab, VS Code & SSH

Connect from your local terminal with standard SSH keys or open web-based IDEs directly in browser tabs without managing complex reverse proxy configs.

Non-Custodial Multi-Chain Rails

Deposit USDT, USDC, SOL, ETH, or BNB to your dedicated personal deposit address. Confirmed on-chain in seconds with zero mandatory identity verification.

Target Workloads for NVIDIA RTX PRO 6000

Single-GPU 70B inference

FP8 weights for a Llama-class 70B are on the order of 70 GB, so they fit with room for a moderate KV cache. FP16 70B does not fit.

70B FP8 or AWQ32B in FP16vLLM on one cardModest batch

Fine-tuning

LoRA and QLoRA through 70B, and full fine-tunes in the 7B–32B range. This is not a pre-training node.

LoRA and QLoRADomain adaptation7B–32B full fine-tunesDebug here, train large models elsewhere

Rendering and DCC

RT cores and a 96 GB framebuffer are why CAD and DCC stacks ask for this part. ECC matters when a glitched frame is a delivery.

Blender, V-Ray, Maya, 3ds MaxCinema 4DOmniverse-class viewportsECC framebuffer

Media encode

AV1 and high-throughput H.264 and HEVC are on the card. Use it as a transcode box only when the GPU stays busy.

AV1HEVC and H.264Batch transcodeStop it when the queue is empty

When to Choose NVIDIA RTX PRO 6000

Pick it to avoid an H100 you would not fill

If the model fits in 96 GB and you do not need NVLink, this is the Blackwell card with a smaller host.

Pick a 32 GB 5090 for small models

Same generation family, a third of the framebuffer, no ECC requirement.

Pick the Ada 48 GB card only for smaller scenes

This part doubles memory to 96 GB GDDR7 and adds FP4.

Pick H100, H200, or B200 for multi-GPU training

No NVLink means scale-out stays data parallel, and PCIe will dominate if you pretend this is an NVSwitch box.

Versus the Ada 48 GB card

Memory doubles to 96 GB GDDR7 and Tensor Cores gain FP4. Rendering and 70B inference are the two jobs that notice.

Versus H100

More VRAM on paper, much less bandwidth, and no NVLink. H100 still wins multi-GPU training. This card wins single-GPU inference when 96 GB is enough and HBM would be idle.

Frequently Asked Questions

How is it priced here?

From the live ask, per second. The menu host is 8 vCPUs and 24 GB of RAM beside 96 GB of VRAM. That is a single-GPU box, not an HGX node.

Will 70B FP16 fit?

No. FP16 70B weights are about 140 GB. Use FP8 or a 4-bit export here, or move to H200 or B200.

How is it different from RTX 6000 Ada?

Memory goes from 48 GB GDDR6 to 96 GB GDDR7, and Tensor Cores gain FP4. Bandwidth is GDDR7, not HBM.

How is it different from H100?

More VRAM on the label, much less bandwidth (1.79 TB/s versus 3.35 TB/s), and no NVLink. H100 still wins multi-GPU training. This card wins single-GPU inference and rendering when 96 GB is enough.

Can it train?

Yes for fine-tunes. Full pre-training of a large model belongs on HBM parts with NVLink.

Is ECC on?

Yes. That is the professional distinction versus a consumer 5090.

Why is it slower than the FLOPS suggest?

Large-batch decode and training steps stream a lot of data. 1.79 TB/s is the tax versus HBM. Single-stream or modest-batch inference of a model that needed 96 GB is where the card is rational.

Video?

Hardware encode and decode, including AV1, are on this generation of professional Blackwell cards.

Can I tensor-parallel a 405B model?

No. There is no NVLink, and the host in the menu is one small machine. That job is an 8-GPU B200 node.

Alternative Hardware Options

Ready to Deploy NVIDIA RTX PRO 6000?

Launch an instance on the live marketplace in under 15 seconds. Fund your balance with cryptocurrency and pay per second with zero commitments.

Launch on Marketplace →