NVIDIA Blackwell B300 hardware
Spike A provides dedicated NVIDIA DGX Blackwell B300 nodes provided through Supercomputing Brabant and hosted at Eindhoven University of Technology (TU/e) as a founding major partner. Reserved exclusively for commercial partners, industry R&D teams, and enterprise collaborative research, the platform combines Blackwell’s next-generation accelerator architecture with all-flash VAST Data parallel storage to support large-scale foundation models, multi-modal training, and high-throughput inference with enterprise SLAs.
DGX B300 compute node specifications
Section titled “DGX B300 compute node specifications”Each Spike A DGX B300 node integrates eight interconnected NVIDIA Blackwell B300 GPUs backed by high-density CPU cores and host memory:
| Metric / Capability | Cluster Specification | Workload Benefit |
|---|---|---|
| Node system | NVIDIA DGX B300 | Flagship enterprise AI supercomputing node architecture |
| Node category | dgx-b300 |
Bright Cluster Manager label: brightcomputing.com/node-category: dgx-b300 |
| Node status | Ready |
Fully provisioned and operational in the Spike A compute pool |
| GPU accelerators | 8x NVIDIA Blackwell B300 GPUs (8.00) |
Extreme tensor compute density for large foundation models and distributed training |
| GPU interconnect | 5th-Generation NVLink & NVSwitch | High bi-directional bandwidth mesh across all 8 GPUs for ultra-fast all-reduce |
| Host CPU cores | 256 CPU cores (256.00) |
High core count for parallel data preprocessing, tokenization, and multi-worker I/O pipelines |
| Host system memory | 2.3 TB system RAM (2.16 TB allocatable) | Massive memory capacity for large in-memory dataset caching, sharded state, and buffer headroom |
| Precision support | FP4 (NVFP4), FP8, BF16, FP16, TF32, FP64 | Native micro-scaling formats for accelerated training and high-throughput inference |
| Storage integration | All-Flash VAST Data parallel storage | Direct high-throughput I/O feeding GPU memory at multi-GB/s rates |
Second-generation Transformer Engine & precision
Section titled “Second-generation Transformer Engine & precision”The second-generation Transformer Engine on Blackwell introduces hardware-accelerated micro-scaling formats (including 4-bit floating point, NVFP4). When supported by your framework, FP4 compute delivers up to twice the throughput of FP8 while maintaining model convergence for supported architectures:
- NVFP4 (4-bit floating point): Optimized for ultra-fast model inference and quantized fine-tuning workloads.
- FP8 (8-bit floating point): Recommended for pre-training and large-scale fine-tuning
with standard PyTorch
torch.aoor NVIDIA Transformer Engine. - BF16 / FP16: Standard fallback formats with broad framework support across all deep learning libraries.
Container and environment recommendations
Section titled “Container and environment recommendations”To leverage Blackwell B300 capabilities, ensure your container images include:
- CUDA Toolkit 12.8 or newer: Blackwell architectures require modern CUDA compiler and runtime support.
- NVIDIA driver compatibility: Spike A compute nodes run modern enterprise drivers certified for Blackwell B300 silicon.
- Framework versions: Use PyTorch 2.5 or newer, or specialized runtimes such as
TensorRT-LLM, vLLM (custom commercial deployments), or Megatron-LM built with
Blackwell architecture flags (
compute_100/sm_100). - Build workflow: Follow the approved Podman and Harbor workflow to build and tag rootless images for the Spike A container registry.
Run:AI resource presets on Spike A
Section titled “Run:AI resource presets on Spike A”Workloads on Spike A request Blackwell B300 resources through the Run:AI scheduler. Select the preset that fits your workload scale:
- 1x B300: Single-GPU prototyping, script verification, interactive notebooks, and small-scale evaluation.
- 2x B300 / 4x B300: Multi-GPU training for medium-to-large models via DistributedDataParallel (DDP) or Fully Sharded Data Parallel (FSDP).
- 8x B300 (Full DGX B300 node): Accesses all 8 GPUs, up to 256 CPU cores, 2.16 TB allocatable host memory, and full intra-node NVLink mesh for maximum tensor parallelism, pipeline parallelism, and large foundation model fine-tuning.
In addition, workloads targeting the Spike A pool match the node category label:
brightcomputing.com/node-category: dgx-b300.
Feeding the GPUs with VAST Data
Section titled “Feeding the GPUs with VAST Data”With the extreme compute throughput of the B300, input data loading is frequently the primary bottleneck. Spike A couples Blackwell nodes with VAST Data parallel storage to ensure continuous I/O saturation:
- Store training data in sequential chunked formats (e.g. WebDataset, Arrow/Parquet, or TFRecords) rather than millions of tiny loose files.
- Mount your project’s VAST Data persistent volume at
/dataor/vastas described in the Storage and data guide. - Review Monitor and profile a first workload to correlate GPU compute utilization with I/O wait times.