NVIDIA RTX Pro 6000 Blackwell
96 GB of VRAM on a single card — run 70B-class models, fine-tune with QLoRA, and generate video without offloading. Launch in under a minute.
- VRAM
- 96 GB GDDR7 ECC
- Memory bandwidth
- 1,597 GB/s
- CUDA cores
- 24,064
- Tensor Cores
- 5th gen · FP4
- Interface
- PCIe Gen 5 x16
- Configurations
- 1 · 2 · 4 · 8 GPU
Powering teams that push boundaries
Trusted by companies including: Tesla, Hugging Face, Kaggle, Zoho, Weights & Biases, upGrad, Saama, Lossfunk
One card. 70B-class models.
96 GB of GDDR7 is the largest VRAM of any single PCIe card on the platform — here is what that buys you in practice.
Inference
Llama 3.3 70B, Qwen2.5 72B in FP8
Native MXFP4 weights (~63 GB) with room for KV cache
Qwen3 32B, DeepSeek-R1-Distill 32B in BF16
Fine-tuning
4-bit base weights + adapters, single card
CodeLlama 34B, Qwen 32B with long contexts
Llama 3.1 8B with 8-bit optimizers
Image & video generation
Headroom for ControlNets, LoRAs, and large batches
Wan 2.2 14B and similar — no CPU offload needed
8 × 96 GB = 768 GB VRAM for bigger models and parallel jobs
Works out of the box with vLLM, SGLang, PyTorch, Axolotl, and ComfyUI.
Pay for minutes, not months
On-demand and spot are billed by the minute — pause and you pay only for storage. Prices shown are US-region rates.
Billed per minute. Start and stop any time; pause and pay only for storage.
Interruptible capacity at roughly half the on-demand rate, billed per minute.
Reserved capacity, billed for the 30-day term rather than by the minute.
Longer commitments are quoted by sales — pricing varies with duration and cluster size.
Honest positioning against the rest of the fleet
The RTX Pro 6000 has the most VRAM of any single PCIe card here; Hopper cards win on interconnect and HBM bandwidth. Pick the card that matches the job.
| GPU | VRAM | Bandwidth | On-demand | When to pick it |
|---|---|---|---|---|
| NVIDIA RTX Pro 6000 Blackwell | 96 GB GDDR7 | 1.6 TB/s | $1.89/hr | The single-card sweet spot for 70B-class inference, fine-tuning, and video generation. |
| NVIDIA H100 SXM | 80 GB HBM3 | 3.35 TB/s | $2.69/hr | NVLink and HBM bandwidth for serious multi-GPU training. |
| NVIDIA H200 SXM | 141 GB HBM3e | 4.8 TB/s | $3.99/hr | Largest memory on the platform for frontier-scale models. |
| NVIDIA A100 80GB | 80 GB HBM2e | 2.0 TB/s | $1.49/hr | Proven Ampere workhorse at a lower price point. |
The single-card sweet spot for 70B-class inference, fine-tuning, and video generation.
NVLink and HBM bandwidth for serious multi-GPU training.
Largest memory on the platform for frontier-scale models.
Managed container or full VM
Every RTX Pro 6000 configuration — 1, 2, 4, or 8 GPUs — is available both ways.
Templates
Managed containers, ready in under a minute
- PyTorch template with CUDA, drivers, and Jupyter pre-configured
- SSH and JupyterLab access out of the box
- Pause and resume — pay only for storage while paused
On-Demand VMs
Full virtual machines with root access
- Bring your own stack — install anything, run any framework
- Same per-minute billing and 1–8 GPU configurations
- Suited for custom drivers, Docker-in-Docker, and long-running services
Technical specifications
Verified specifications for the NVIDIA RTX Pro 6000 Blackwell as deployed on JarvisLabs.
| Specification | Value | Notes |
|---|---|---|
| Architecture | NVIDIA Blackwell | Professional (RTX Pro) series, released March 2025 |
| Edition | Server Edition | Passively cooled dual-slot server card, as deployed on JarvisLabs |
| CUDA cores | 24,064 | GB202-based |
| Tensor Cores | 752 (5th gen) | FP4 / FP8 / FP16 / BF16 / INT8 |
| RT Cores | 188 (4th gen) | Ray tracing and 3D rendering |
| AI performance | Up to 4,000 AI TOPS | FP4 with sparsity |
| VRAM | 96 GB GDDR7 (ECC) | Largest single-card VRAM in the RTX line |
| Memory interface | 512-bit | |
| Memory bandwidth | 1,597 GB/s | Server Edition |
| Interconnect | PCIe Gen 5 x16 | No NVLink — multi-GPU scales over PCIe |
| ECC memory | Yes | Data integrity for long training runs |
| Multi-GPU on JarvisLabs | 1, 2, 4, or 8 GPUs | Templates and On-Demand VMs |
Professional (RTX Pro) series, released March 2025
Passively cooled dual-slot server card, as deployed on JarvisLabs
GB202-based
FP4 / FP8 / FP16 / BF16 / INT8
Ray tracing and 3D rendering
FP4 with sparsity
Largest single-card VRAM in the RTX line
Server Edition
No NVLink — multi-GPU scales over PCIe
Data integrity for long training runs
Templates and On-Demand VMs
Frequently asked questions
Everything you need to know about the NVIDIA RTX Pro 6000 Blackwell on JarvisLabs.
$1.89/hr on-demand and $0.99/hr on spot in US regions, billed per minute — you only pay while an instance is running. A 1-month reserved plan is $1.49/hr, billed for the term; longer commitments are quoted by sales.
70B-class models like Llama 3.3 70B in FP8 on a single card, gpt-oss-120b in its native MXFP4 format, and 32B models in full BF16 precision. For fine-tuning: QLoRA up to 70B, LoRA up to ~34B, and full fine-tunes of 7–8B models. Image and video models like FLUX.1 and Wan 2.2 run without CPU offloading.
The RTX Pro 6000 has more VRAM (96 GB vs 80 GB) at a lower price ($1.89/hr vs $2.69/hr), making it the better value for single-card inference and fine-tuning. The H100 has HBM3 with roughly twice the memory bandwidth plus NVLink, so it pulls ahead for multi-GPU training and bandwidth-bound inference.
It is a full generational leap: double the memory (96 GB GDDR7 vs 48 GB GDDR6), roughly double the bandwidth, 5th-gen Tensor Cores with native FP4 support, and ECC throughout.
No — the RTX Pro 6000 connects over PCIe Gen 5. On JarvisLabs you can launch 1, 2, 4, or 8 GPU configurations. Data-parallel training runs well over PCIe, and for multi-GPU inference vLLM supports both tensor and pipeline parallelism — pipeline parallelism often scales better on cards without NVLink.
Sign up, pick the PyTorch template (or an On-Demand VM for root access), choose RTX Pro 6000 and your GPU count, and launch. Instances are typically ready in under 60 seconds with SSH and JupyterLab access.
Yes. The full 96 GB of GDDR7 is ECC-protected, which prevents silent memory corruption during long training runs and production inference.
FP4 is a 4-bit floating-point format that the 5th-gen Tensor Cores accelerate natively. Models quantized to FP4/MXFP4 — like gpt-oss — take roughly a quarter of the memory of FP16 with strong quality, which is how a 120B-parameter model fits on this single card.
96 GB of Blackwell, one minute away
Launch an RTX Pro 6000 instance from $1.89/hr with per-minute billing. Pause any time and pay only for storage.