The Sensitive Data Services (TSD) that we operate in partnership with the University of Oslo is about to take major step forward with the introduction of Colossus, a new high-performance cluster purpose-built for sensitive data workloads.
While all TSD users stand to gain from faster processing and shorter queues, some groups will see particularly strong benefits. AI-heavy workloads, natural language processing, and genomic/transcription pipelines (such as alternatives to Dragen on the new H200s) are prime candidates to take advantage of Colossus.
What’s inside Colossus
- 2,880 AMD Turin cores (Zen5) across 15 nodes
- 192 cores and 1.5 TiB memory per node, with 7.68 TB local scratch
- 4× NVIDIA H200 GPUs (141 GiB, NVLink) and 8× NVIDIA RTX Pro 6000 GPUs (96 GiB)
- 200 Gbps NDR InfiniBand interconnect
This is a major generational jump from earlier Zen1/Zen2 CPUs and V100/A100 GPUs, raising the bar for sensitive data analysis and AI workloads in Norway.
New access model
Alongside the hardware upgrade, TSD is rolling out a new contribution model. Every project will receive a default quota of 100k CPU hours, up from the previous 10k. Only projects requiring more than 100k hours (or needing priority access) will need to go through the RFK process. All resources, including GPUs and big-memory nodes, are now broadly accessible across the TSD community, reducing bottlenecks and improving fairness.
What’s next
As Colossus comes online, the team is also preparing new features and documentation:
- Open OnDemand for easier HPC access via the browser
- New HPC documentation tailored for TSD
- Slurm REST API for seamless integration with services like AutoTekst transcription and other AI models
- API available for other TREs (Trusted Research Environments). Other secure environments can send HPC jobs into Colossus to process sensitive data.
Together, these changes mean that more than 3,000 TSD projects can now access world-class compute and GPU resources, helping Norway’s research community tackle sensitive data challenges at scale.