.avif)
Powering the next generation of AI & high-performance computing.
Engineered for large-scale AI training, deep learning, and high-performance workloads, delivering unprecedented compute power and efficiency.
Why rent the H100 NVL instead of buying?
Built for LLM inference at scale
The H100 NVL's dual-GPU PCIe design delivers 188 GB of combined HBM3 memory — more than any single GPU in the H100 lineup — making it uniquely suited for deploying the largest open-source LLMs in full precision without quantization or model sharding. Transformer Engine support delivers up to 30× higher inference performance compared to previous generations.
Pay only for what you use
An H100 NVL board costs upwards of $40,000. Runpod's on-demand pricing gives you instant access to the same hardware, converting capital expenditure to operational expense with no maintenance overhead or depreciation.
Deploy in seconds, scale or switch when you're ready
Provision an H100 NVL instance in seconds. Scale up for a large inference deployment, scale down during development, or switch configurations entirely without long-term hardware commitments. Runpod handles the infrastructure so your team stays focused on the model.
Use Cases
Popular use cases.
Designed for demanding workloads —learn if this GPU fits your needs.
Technical Specs
Ready for your most demanding workloads.
Essential technical specifications to help you choose the right GPU for your workload.
| Specification | Details | Great for... |
|---|---|---|
| Architecture | NVIDIA Hopper (GH100) | Large language model inference and deployment at scale, with Transformer Engine and FP8 support |
| Manufacturing Process | 4nm | — |
| CUDA Cores | 14,592 | Parallel compute tasks across LLM inference, scientific simulation, and data processing pipelines |
| GPU Memory | 188 GB HBM3 combined (94 GB per GPU) | Deploying the largest open-source LLMs in full precision without quantization or model sharding |
| Memory Bus | 5,120-bit | Sustaining peak bandwidth to HBM3 across both GPUs for large-context inference workloads |
| L2 Cache | 50 MB | — |
| Base / Boost Clock | 1,095 / 1,755 MHz | Sustained compute performance for long inference and training sessions |
| TDP | 350W per GPU | Predictable power budgeting in PCIe server infrastructure without SXM power requirements |
| FP64 Performance | 26 TFLOPS | High-precision scientific computing and simulation workloads |
| FP32 Performance | 51 TFLOPS | Standard-precision training, fine-tuning, and inference |
| TF32 Tensor Core | 756 TFLOPS | Accelerated training with near-FP32 accuracy for transformer and LLM workloads |
| BF16 Tensor Core | 1,513 TFLOPS | Stable large model training with FP32-range numerics at higher throughput |
| FP8 Tensor Core | 3,026 TFLOPS | Maximum inference throughput for quantized production models |
"The Runpod team has clearly prioritized the developer experience to create an elegant solution that enables individuals to rapidly develop custom AI apps or integrations while also paving the way for organizations to truly deliver on the promise of AI."
Amjad Masad
"Runpod is the only place I can deploy high-end GPU models instantly—no sales calls, no rate limits, no nonsense."
Daniel Chang
“The main value proposition for us was the flexibility Runpod offered. We were able to scale up effortlessly to meet the demand at launch.”
Josh Payne
“Runpod helped us scale the part of our platform that drives creation. That’s what fuels the rest—image generation, sharing, remixing. It starts with training.”
Matty Shimura
Comparison
Powerful GPUs. Globally available. Reliability you can trust.
30+ GPUs, 31 regions, instant scale. Fine-tune or go full Skynet—we’ve got you.
FAQs
Questions? Answers.
What are the current hourly rates for renting an H100 NVL on Runpod?
Rates vary by instance type. For the most current pricing, see the Runpod pricing page.
When does renting H100 NVL GPUs make more sense than purchasing?
Renting is particularly advantageous for short-term projects, proof-of-concept development, variable computing needs, and teams without capital for hardware investment. For ongoing production workloads with predictable, high utilization over 12+ months, ownership may become cost-competitive.
How does the H100 NVL perform compared to the H100 PCIe or H100 SXM?
The H100 NVL provides 188 GB combined memory (94 GB per GPU) vs 80 GB for the PCIe and SXM variants, making it the best choice for deploying the largest LLMs in full precision. The SXM offers higher raw compute throughput (3.35 TB/s memory bandwidth, 900 GB/s NVLink) for distributed training. See the GPU comparison pages for a side-by-side breakdown.
What security considerations apply when renting H100 NVL GPUs?
Runpod Secure Cloud instances provide enterprise-grade isolation. The H100 includes hardware-level confidential computing support. For GDPR, HIPAA, or SOC 2 requirements, review Runpod's compliance documentation and contact the team about compliant deployment options.
How scalable are H100 NVL rental solutions?
You can scale up for intensive training phases and scale back down during development, without long-term commitments. Most configurations support auto-scaling based on workload demands. For distributed training across multiple H100 NVL instances, confirm the networking configuration supports the required inter-GPU bandwidth to fully leverage NVLink capabilities.
10,100,100,100
Requests since launch & 400k developers worldwide
