NVIDIA DGX Spark
A Grace Blackwell AI supercomputer hosted in our Madrid datacenter.
Prototype, fine-tune and deploy AI models up to 200B parameters on the GB10 superchip with 128 GB of coherent unified memory. Hosted in our Tier III+ Madrid datacenter, accessible from anywhere.
Key Specifications
Grace Blackwell architecture, designed by NVIDIA for end-to-end generative AI
GB10 Superchip
Grace CPU + Blackwell GPU integrated in a single superchip with high-speed NVLink-C2C. Latest-generation Arm cores and 5th-gen Tensor Cores.
1 PFLOPS of FP4 AI
One petaflop of performance on FP4 operations with sparsity. Acceleration for LLMs, diffusion models, RAG and AI agents running locally.
128 GB Coherent Unified Memory
Memory shared between CPU and GPU without copies. Run models up to 200B parameters (quantised) without falling back to CPU offload.
ConnectX-7 SmartNIC
NVIDIA ConnectX-7 network adapter with RDMA support up to 200 Gbps. Lets you interconnect up to three DGX Sparks for 384 GB of combined coherent memory and 3 PFLOPS of combined AI performance.
4 TB Self-Encrypting NVMe M.2
High-performance NVMe M.2 SSD with hardware encryption (SED). Plenty of room for datasets, checkpoints and multiple local models.
DGX OS + NVIDIA AI Stack
DGX OS (Ubuntu-based) operating system with the full NVIDIA AI stack: CUDA, cuDNN, TensorRT, NGC containers, NIM microservices and NeMo Framework.
Configure your DGX Spark
Start with one node or scale up to three interconnected units via ConnectX-7 at 200 Gbps RDMA to combine coherent memory and compute.
DGX Spark Use Cases
Built for AI developers and researchers who need to iterate fast on real NVIDIA hardware
AI Prototyping
Iterate on LLMs, agents and diffusion models with the same NVIDIA stack you'll use in production.
Local Fine-tuning
LoRA, QLoRA and SFT for models up to 70B with your own data, never leaving your infrastructure.
Private LLM Inference
Deploy Llama, Mistral, Qwen or DeepSeek with vLLM/Ollama and keep your data in Spain.
Research and MLOps
A reproducible NVIDIA dev environment for small teams or individual researchers.
Why host your DGX Spark with EasyDataHost?
The DGX Spark draws power, generates heat and needs professional connectivity and maintenance. In our Tier III+ Madrid datacenter you keep it running 24/7 without taking up office space.
-
check_circle
Tier III+ datacenter with N+1 powerISO 27001, ENS High, PUE 1.3, 100% renewable energy
-
check_circle
Unmetered 25 Gbps connectivityOur own AS205081 network with peering at major European IXPs
-
check_circle
Remote access via private VPNDedicated VPN to the client, private VLAN available
-
check_circle
Spanish-speaking technical supportProactive monitoring and on-site intervention with no surcharges
-
check_circle
Data sovereignty in SpainGDPR and ENS compliance, no CLOUD Act extraterritoriality
Frequently Asked Questions
Everything you need to know about hosting the NVIDIA DGX Spark
What is the NVIDIA DGX Spark? expand_more
Why host it with EasyDataHost instead of keeping it at the office? expand_more
What does the monthly fee include? expand_more
What software ships preinstalled? expand_more
Can I train large models on a single DGX Spark? expand_more
Can multiple DGX Spark units be interconnected? expand_more
Can I rent it by the week or fortnight? expand_more
Related Services
Combine the DGX Spark with the rest of EasyDataHost's AI infrastructure
GPU Servers
RTX 6000 Server Edition and H200 for large-scale training and production.
Object Storage S3
Store datasets, checkpoints and artifacts with no egress fees and 25 Gbps access from the Spark.
Managed Services
End-to-end Spark management: updates, monitoring, deployments and MLOps.
Reserve your NVIDIA DGX Spark
1, 2 or 3-node configurations in Madrid from €450/month (+€75/month per additional node for the 200 Gbps interconnect). Ready for your team to start iterating on generative AI on real NVIDIA hardware.