Fine tune models like LLaMA 2
Optimize Transformers Models and LLMs through efficient processes, and accelerate the training of larger models with the cutting-edge Tensor Cores 4th generation technology and the latest 8-bit data format.
H100 PCIe GPU instance
Accelerate your model training and inference with the most high-end AI chip of the market!
Accelerate inference workloads up to 30 times
Accelerate your model serving workloads thanks to Transformer Engine 30x faster for AI inference and new data formats.
Maximize GPU utility up to your needs
With 2nd generation of Secure MIG (multi-instance GPU), partition the GPU into isolated, right-size instances to maximize utilization for the smallest to biggest multi-GPU jobs.
H100 PCIe GPU technical specifications
- gpu
GPU
NVIDIA H100 PCIe Tensor Core
- memory
GPU Memory
80GB HBM2e
- processor
Processor
24 vCPUs AMD Epyc Zen 4
- processor_frequency
Processor frequency
2.7 Ghz
- memory
Memory
240 GB of RAM
- memory_type
Memory type
DDR5
- bandwidth
Bandwidth
10 Gbps
- storage
Storage
Block Storage for the boot and 3TB of Scratch Storage NVMe
Secure your H100 PCIe GPU Instance resource for months or years
Estimate the GPU cost
Choose your plan
Estimated cost
| Option | Value | Price |
|---|---|---|
| Zone | Paris 2 | |
| Instance | 1x | 0€ |
| Volume | 10GB | 0€ |
| Flexible IPv4 | Yes | 0.004€ |
| Total | |
|---|---|
| Daily | 0€ |
| Weekly | 0€ |
| Monthly | 0€ |
Boost Innovation Sustainably: 100% Renewable Energy, 50% Less Power
DC5 PAR2 Paris
DC5 is one of Europe's greenest data centers, powered entirely by renewable wind and hydro energy (GO-certified) and cooled with ultra-efficient free and adiabatic cooling. With a PUE of 1.16 (vs. the 1.55 industry average), it slashes energy use by 30-50% compared to traditional data centers.
WAW2 Warsaw
WAW2 runs on 100% wind power (GO-certified) and uses a combination of direct free cooling, free chilling, immersion systems, and air conditioning to optimize system cooling. With a PUE of 1.32—better than the industry average—it minimizes energy consumption for maximum efficiency.
Customer success stories
"Execution difference is 40% in favor of using the H100 PCIe GPUs"
Sovereign AI specialists Golem.ai took a deep technical dive into the topic and shared their findings on our blog. “After running a hundred tests in total between Replicate.com and the H100 de Nvidia hosted by Scaleway, we conclude that the execution difference is 40% in favor of using the H100s,” says Golem.ai’s Kevin Baude.
Numerous AI applications and use cases
Natural Language Processing
Understands, interprets, and generates human language in a way that is both meaningful and contextually relevant.
Thanks to models and algorithms specialized in:
- Text classification
- Machine translation
- Entailment prediction
- Named entity recognition
- Sequence-to-sequence, like BERT for text extraction
- Text similarity search, like BERT to find semantic similarities
- Language modeling
Automatic Speech Recognition
Converts spoken language into written text, facilitating the translation of verbal communication into machine-readable data.
Thanks to models and algorithms specialized in:
- Speaker recognition (using Fast Fourier Transform (FFT) and a 1D Convnet., as explained by Fabi Badine)
- Accent recognition
- KWS, keywords spotting used by many HCI (Human-computer interfaces) like Alexa or Siri
(example: RNN-T on the Librispeech dataset.)
Generative AI
Generates new content, such as images, text, audio, code. It autonomously produces novel and coherent outputs, expanding the realm of AI-generated content beyond replication or prediction.
With models and algorithms specialized in:
- Image generation like Deep Dream that produces hallucination-like visuals as fchollet explains it
- Text generation with Transformers Models also called LLMs (Large Language Models) such as GPT2
- Code generation
Computer vision
Enables machines to interpret and understand visual information from the world, much like human vision.
Thanks to models and algorithms specialized in:
- Image classification (using BigTransfer(BiT) trained on public datasets, along with code in TF2, Jax and Pytorch as explained by Sayan Nat)
- Image segmentation
- 3D
- OCR
- Image enhancement
- Image and text
- vision models interpretability
- Image similarity search
Recommender
Predicts and suggests items of interest to users based on their preferences and behaviors, enhancing personalized recommendations and decision-making in various applications.
Using for example Deep Learning Recommendation Model v2 (DLRMv2) that employs DCNv2 cross-layer and a multi-hot dataset synthesized from the Criteo dataset.
Choose your instance's format
| Instance Name | Number of GPU | TFLOPs FP16 Tensor Cores | VRAM | Prices until June, 30 | Prices from July, 1 |
|---|---|---|---|---|---|
| H100-1-80GB | 1 H100 PCIe Tensor Core | Up to 1,513 teraFLOPS | 80GB | €2.52/hour | €2.73/hour |
| H100-2-80G | 2 H100 PCIe Tensor Core | Up to 3,026 teraFLOPS | 2 x 80GB | €5.04/hour | €5.46/hour |
Enjoy the simplicity of a pre-configured AI environment
Optimized GPU OS Image
Benefit from a ready-to-use Ubuntu image to launch your favorite deep learning containers (pre-installed NVIDIA driver and Docker environment).
Enjoy your favorite Jupyter environment
Easily launch your favorite JupyterLab or Notebook thanks to the pre-installed Docker environment
Choose your AI containers among multiple registries
Access multiple container registries: your own build containers, Scaleway AI containers, NVIDIA NGC registry and any other registry
NVIDIA Enterprise AI software at your disposal
Access hundreds of AI softwares optimized by Nvidia to maximise the efficiency of your GPUs and boost your productivity. Among hundreds of softwares developed by NVIDIA and tested by leaders of their industry, harness the efficiency of
- NVIDIA Nemo for LLM fine-tuning,
- NVIDIA TAO for computer vision,
- NVIDIA TRITON for inference
Deploy and Scale your infrastructure with Kubernetes
Frequently asked questions
What is included in the instance price?
3TB of Scratch Storage are included in the instance price, but any Block Storage provisioned by you, is at your expense.
For redundancy and thus security reasons we strongly recommend that you provision extra Block Storage volume, as Scratch Storage is ephemeral storage that disappears when you switch off the machine. Scratch Storage purpose is to speed up the transfer of your data sets to the gpu.
How to use Scratch storage then? Follow the guide
Whats the difference between H100-1-80G and H100-2-80G?
These are 2 formats of the same instance embedding NVIDIA H100 PCIe Tensor Core.
- H100-1-80G embeds 1 GPU NVIDIA H100 PCIe Tensor Core, offering a GPU memory of 80GB
- H100-2-80G embeds 2 GPUs NVIDIA H100 PCIe Tensor Core, offering a GPU memory of 2 times 80GB. This instance enables faster time to train for bigger Transformers models that scale 2 GPUs at a time. Thanks to the PCIe Board Factor, the servers of the H100 PCIe GPU instance are made of 2 GPUs. By launching a H100-2-80G instance format, the user benefits from a fully dedicated server with 2 GPUs.
What is the environmental impact of the H100 PCIe instance?
NVIDIA announced the H100 to enable companies to slash costs for deploying AI, "delivering the same AI performance with 3.5x more energy efficiency and 3x lower total cost of ownership, while using 5x fewer server nodes over the previous generation."
What inside the product can confirm this announcement?
- The thinner engraving of the chip reduces the surface and thus the energy required to power the chip
- Thanks to innovations like the new data format FP8 (8bits) more calculations are done with the same amount of consumption resulting in time and energy optimization
In addition, at Scaleway we decided to localize our H100 PCIe instances in the adiabatic Data Center DC5. With a PUE (Power User Effectiveness) of 1.15 (average is usually 1.6) this datacenter saves between 30% and 50% electricity compared with a conventional data centre.
Stay tuned for our benchmarks on the topic!
How can I use MIG to get the most out of my GPU?
NVIDIA Multi-Instance GPU (MIG) is a technology introduced by NVIDIA to enhance the utilization and flexibility of their data center GPUs, specifically designed for virtualization and multi-tenant environments. It allows a single physical GPU to be partitioned into up to seven smaller Instances, each of which operates as an independent MIG partition with its own dedicated resources, such as memory, compute cores, and video outputs.
Read the dedicated documentation to use MIG technology on your GPU instance
How to choose the right GPU for my workload?
There are many criteria to take into account to choose the right GPU instance:
- Workload requirements
- Performance requirements
- GPU type
- GPU memory
- CPU and RAM
- GPU driver and software compatibility
- Scaling
For more guidance read the dedicated documentation on that topic