Define the AI infrastructure you need for the next 1 to 3 years – we handle everything else
Whether you need to develop Foundation Models or run multiple large-scale training jobs, our custom-built GPU clusters let you define the exact hardware and resources (GPUs, storage, network) tailored to your machine learning workload.
Keep control of your AI development under strict European data regulations. Your data stays secure and compliant with Scaleway's infrastructure, ensuring it's safe from extraterritorial access throughout the machine learning lifecycle.
Moshi –Kyutai’s revolutionary AI voice assistant–, as well as Mixtral –a highly efficient mixture-of-experts model built by Mistral AI– were both trained on Nabu 2023, the 1st Custom-built Cluster. At its release, Mixtral outperformed existing closed and open weight models across most benchmarks.
DC5, in PAR2 region, 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.
"We're currently working on Scaleway SuperPod, which is performing exceptionally well.", Arthur Mensch on the Master Stage at ai-PULSE 2023. Mistral AI, used Nabu to build its Mixtral model , a highly efficient mixture-of-experts model. At its release, Mixtral outperformed existing closed and open weight models across most benchmarks, offering superior performance with fewer active parameters, making it a major innovation in the field of AI. The collaboration with Scaleway enabled Mistral to scale its training efficiently, allowing Mixtral to achieve groundbreaking results in record time.
“We've benchmarked a bit of Nabu 2023, Scaleway AI Supercomputer and achieved very appropriate results compared to other CSPs. Results that could be greatly improved with a bit more tuning! We can't wait to benchmark the full 127 DGX nodes and see what performance we achieve” Guillaume Salou, ML Infra Lead at Hugging Face
“We've seamlessly benchmarked all existing Open Source LLM from 7B to 70B with Scaleway's supercomputer Jero 2023. From beta testing to product development of a POC using a version we fine-tuned to meet our needs, Scaleway's European sovereignty coupled with our private data center has allowed us to scale without constraints while maintaining control over our data.” Patrick Escande, Executive Vice-President, Research & Innovation at Free pro
Moshi, Kyutai’s revolutionary AI voice assistant brings unprecedented vocal capabilities. Trained using Scaleway’s high-performance Cluster and served with our L4 GPU instances, Moshi excels in conveying emotions and accents with 300x codec compression. This setup enabled Moshi to process 70 different emotions and accents with ultra-low latency, allowing for seamless, human-like conversations. Thanks to this high-performance environment, Kyutai was able to achieve this breakthrough.
Contact our AI experts to get a more personalized cluster
| Custom-built Cluster name | Number of GPUs | max PFLOPs FP8 Tensor Core |
|---|---|---|
| Nabuchodonosor 2023 | 1,016 H100 Tensor Core GPUs (SXM5) | Up to 4,021.3 PFLOPS |
| Jeroboam 2023 | 16 H100 Tensor Core GPUs (SXM5) | Up to 63.2 PFLOPS |
CPU
Dual Intel® Xeon® Platinum 8480C Processors 112 Cores total
Total CPU cores
14,224 cores
GPU
1,016 Nvidia H100 Tensor Core GPUs (SXM5)
Total GPU Memory
81,280GB
Processor frequency
Max of 3.80 GHz
Total RAM Memory
254 TB of RAM
Storage type
1.8 PB of a3i DDN low latency storage
Storage capacity per DGX
2.7 TB/s Read and 1.95 TB/s Write
Inter-GPU Bandwidth
Infiniband 400 Gb/s
CPU
Dual Intel® Xeon® Platinum 8480C Processors 112 Cores total
Total CPU Cores
224 cores
GPU
16 Nvidia H100 Tensor Core GPUs (SXM5)
Total GPU memory
1,280GB
Processor frequency
Max of 3.80 GHz
Total RAM memory
4 TB of RAM
Storage type
64TB of a3i DDN low latency storage
Inter-GPU Bandwidth
Infiniband 400 Gb/s
Our Custom-built Clusters, Nabu & Jero 2023 are made of NVIDIA DGX H100 systems with Nvidia H100 Tensor Core GPUs 80GB (SXM5). They can reach lightning fast multi-node scaling for AI, thanks to their latest generation GPUs:
Thanks to the InfiniBand NDR interconnection (400Gb/s), each 8 GPU compute node offers 3.2 Tb/s of bandwidth to all the other nodes on a totally non-blocking network architecture.
Its brand new GPUDirect RDMA technology accelerates direct communication across all nodes of the cluster using InfiniBand, enabling:
The Custom-built Clusters benefit from DDN a3i storage optimized for ultra-fast computing. With over:
Benefit from a comprehensive management for the cluster with SLURM. An Open-source cluster management and job scheduling system for Linux cluster.
Generates new content, such as images, text, audio or, 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:
Understands, interprets, and generates human language in a way that is both meaningful and contextually relevant.
Thanks to models and algorithms specialized in:
Converts spoken language into written text, facilitating the translation of verbal communication into machine-readable data.
Thanks to models and algorithms specialized in:
Predicts and suggests items of interest to users based on their preferences and behaviors, enhancing personalized recommendations and decision-making in various applications.
