NVIDIA Blackwell B200 hardware
Spike 1 provides NVIDIA’s flagship DGX B200 systems for TU/e academic research. Engineered for demanding scientific machine learning, computer vision, and foundation model workloads, each DGX node combines eight interconnected NVIDIA Blackwell B200 GPUs, high-bandwidth NVLink switching, and high-performance 250 TiB DDN project storage.
DGX B200 architecture highlights
Section titled “DGX B200 architecture highlights”The DGX B200 introduces significant generational speedups in tensor compute, memory capacity, and inter-GPU communication:
| Capability | Specification | Workload benefit |
|---|---|---|
| GPU accelerator | 8x NVIDIA Blackwell B200 GPUs | Next-generation streaming multiprocessors with high-density Tensor Cores |
| GPU memory (vRAM) | 192 GB HBM3e per GPU (1.5 TB total per node) | Accommodates large models, extended context windows, and large batch sizes |
| Memory bandwidth | 8.0 TB/s per GPU | Dramatically reduces memory-bandwidth bottlenecks during attention operations |
| Interconnect | 5th-Gen NVLink (1.8 TB/s bi-directional per GPU) | NVSwitch mesh connecting all 8 GPUs for ultra-fast all-reduce and tensor parallelism |
| Precision support | FP4 (NVFP4), FP8, BF16, FP16, TF32, FP64 | Second-generation Transformer Engine with native micro-scaling acceleration |
| Project storage | 250 TiB high-performing DDN space | High-throughput parallel storage mounted at /data with S3 connection options |
| Software suite | NVIDIA AI Enterprise | Optimized containers, validated runtimes, and profiling tools |
Second-generation Transformer Engine & precision
Section titled “Second-generation Transformer Engine & precision”The Blackwell B200 features hardware-accelerated precision formats tailored for modern deep learning:
- NVFP4 (4-bit floating point): Hardware micro-scaling format providing up to 2× the inference throughput of FP8 while preserving accuracy for supported model architectures.
- FP8 (8-bit floating point): Accelerated training and fine-tuning with automatic
scaling via PyTorch
torch.aoor NVIDIA Transformer Engine. - BF16 / FP16: Native 16-bit mixed precision widely supported across standard PyTorch, JAX, and Hugging Face pipelines.
Run:AI resource presets on Spike 1
Section titled “Run:AI resource presets on Spike 1”Workloads request Blackwell B200 GPUs through the Run:AI scheduler. Match your request to your workload requirements:
- 1x B200 (192 GB vRAM): Prototyping, exploratory notebooks, single-GPU training, and medium-scale model fine-tuning.
- 2x B200 / 4x B200: Distributed data parallel (DDP) or Fully Sharded Data Parallel (FSDP) training across multiple accelerators.
- 8x B200 (Full DGX Node): Full-node distributed runs utilizing the full 1.8 TB/s NVLink fabric for maximum tensor parallel and pipeline parallel efficiency.
High-throughput DDN storage integration
Section titled “High-throughput DDN storage integration”Because B200 Tensor Cores process matrix operations rapidly, keeping GPUs saturated requires an efficient data pipeline:
- Mount your project’s persistent volume claim (PVC) at
/dataas outlined in Storage and data. - For large datasets, leverage the S3 connection option or package data into chunked sequential formats (e.g. WebDataset or Sharded Parquet) to avoid file-system metadata overhead.
- Consult Monitoring and profiling to track GPU kernel utilization and verify I/O wait times remain minimal.