H100 NVL vs H100 SXM: which is faster?

For LLM decoding the H100 NVL is listed with more memory (94 GB vs 80 GB) and more bandwidth (3381 GB/s vs 2869 GB/s). The H100 SXM has higher FP32 at 53.5 TFLOPS and a 700 W limit, so it suits compute-heavy training.

Can I run a 70B model in 16-bit on the H100 NVL?

No. 16-bit 70B weights need about 140 GB. On 94 GB, run 70B quantized to 4 or 8 bits; for 16-bit 70B, the H200 NVL with 141 GB is the closer fit.

Does Hopper need special software?

Hopper has compute capability 9.0 and runs current CUDA, PyTorch, vLLM and SGLang builds. Install the tools you need over SSH, or bring them in your own public Docker image.

How does billing work for the H100 NVL?

Per minute from your account balance while it runs, at the price locked when you deploy, plus bandwidth per GB. A stopped instance is charged for storage only. There are no monthly GPU plans.

GPU Cloud · NVIDIA · Hopper

Rent an NVIDIA H100 NVL GPU with 94 GB VRAM

Rent H100 NVL cloud GPU servers by the hour on Cloudj: 94 GB VRAM, NVIDIA Hopper and about 3381 GB/s of memory bandwidth for large-model inference. Built for LLM teams that need more memory than an 80 GB H100 and much faster decoding, it delivers 48.3 TFLOPS FP32 at a 400 W power limit. Pick a live configuration, start from the Ubuntu 22.04 + CUDA or PyTorch + JupyterLab template and connect over SSH with your key.

Rent H100 NVL All GPUs

H100 NVL specifications

VRAM per GPU

94 GB

Architecture

NVIDIA Hopper

Compute capability

FP32 compute

≈ 48.3 TFLOPS

Memory bandwidth

≈ 3381 GB/s

Power limit

400 W

Values reported by the infrastructure provider for machines currently listed; real performance varies by machine.

Live H100 NVL configurations and prices

Out of stock right now

Browse available GPUs

What you can run on the H100 NVL cloud GPU

High-throughput LLM serving

3381 GB/s of bandwidth speeds up token generation, and 94 GB keeps a quantized 70B model plus a large KV cache on one GPU.

Long-context and RAG pipelines

The extra 14 GB over an 80 GB card goes straight into context length and concurrent requests.

Fine-tuning 30B–70B models

Run LoRA on 30B-class models in BF16 or QLoRA on 70B-class models with Unsloth on a single Hopper GPU.

Model evaluation at scale

Run benchmark suites and A/B evaluations across checkpoints quickly, then delete the server when the scores are in.

Is the H100 NVL the right GPU for your workload?

Choose the H100 NVL when your AI or compute workloads require more memory than standard datacenter cards provide, up to 94 GB of VRAM. A smaller workstation or consumer GPU is a better choice for lightweight inference and development, while a multi-GPU setup or a larger memory tier is necessary if your models exceed 94 GB.

H100 NVL vs similar cloud GPUs

GPU	VRAM per GPU	FP32 compute	Power limit
NVIDIA H100 NVL	94 GB	≈ 48.3 TFLOPS	400 W
NVIDIA H100 SXM	80 GB	≈ 53.5 TFLOPS	700 W	Details →
NVIDIA H100 PCIE	80 GB	≈ 41 TFLOPS	350 W	Details →
NVIDIA A800 PCIE	80 GB	≈ 15.6 TFLOPS	300 W	Details →
NVIDIA CMP 170HX	64 GB	≈ 12.1 TFLOPS	250 W	Details →

Guides for H100 NVL

H100 cloud Asia: what to try before settling elsewhereAn H100 cloud Asia playbook: search JP, KR, SG, TW and IN codes, accept that datacenter GPUs can be scarce, and train elsewhere while serving from Asia.H100 cloud Europe guide: renting H100 GPUs in EU countriesAn H100 cloud Europe checklist: SXM, PCIe and NVL variants, searching EU country codes, the account level datacenter GPUs need and the Datacenter only switch.GPU cloud Asia guide: Japan, Singapore, Korea and nearbyA practical GPU cloud Asia guide: which country codes to search for East and Southeast Asian users, a fallback order, and how to pair a GPU with a VPS.GPU cloud by location: how to pick a country for your GPUHow GPU cloud by location works on Cloudj: filter a live list of machines by country code, decide between near users, lowest price or data rules.

All articles →

H100 NVL cloud GPU: frequently asked questions

Rent the H100 NVL cloud GPU by the hour

Pick a live configuration, choose a template and connect over SSH with your key. Billing stops when you destroy the instance.

Rent H100 NVL All GPUs