SPIKE-1 is now open to general research
As you may have read before, TU/e is one of the first universities to have access to the NVIDIA DGX B200 system. The cluster has been tested by different groups and is now open to all TU/e researcher’s proposal for access.
Different departments have invested different amounts proportionally, thus researchers can use the system based on their department quota. For funded research (e.g. NWO/EU), it is also possible to use more GPUs using a pay-per-hour model. If you are interested and want to know more, please fill the SPIKE-1 project proposal or send an email to supercomputing@tue.nl.
SPIKE-1 has five unique features compared to other HPC facilities:
- FP4 native hardware support
- NVLink & InfiniBand for large scale training
- Run:ai Container orchestration
- GPU Direct Storage
- Integrated with NVIDIA NGC ecosystem
For researchers, SPIKE-1 is ideal if you have the need for:
- FP4 research; or
- Large scale training needs (multi-GPU/multi-nodes); or
- Large VRAM (e.g. training foundation models)
What is SPIKE-1
SPIKE-1 is a modular AI supercomputer from Nvidia. TU/e Supercomputing Center installed the first batch of 4 (DGX B200) systems early 2025. The system is hosted in a datacenter in Kajaani (Finland). At the moment Proof of concepts are running on the system.
- 'State of the art': With the newest Nvidia ‘Blackwell’ Technology (DGX B200)
- The GPU’s will be renewed every 2-3 years, so It will STAY ‘state of the art’
- Placed in Finland, in a green datacenter
Technical details
Hardware
4x DGX B200 nodes available with each:
- 8x Nvidia ‘Blackwell’ B200 GPUs
- 1.53 TB VRAM (~196 GB per GPU)
- 2x 112 CPU core Intel Platinum 8570
- 2.16 TB RAM
Software
- Code runs in containers that can have a certain number of GPUs assigned
- NVIDIA AI Enterprise software
Please note that this is different from how most other HPC systems work (TU/e HPC Umbrella and Snellius both use Slurm for job scheduling). Some experience with building container images is recommended.
