B300 reference build.
16 nodes, 128 GPUs: our standard bill of materials. CPU, storage and fabric change to fit your workload.
Node specification
The standard configuration, per node. Where a row offers an alternative, choose it here and it carries into your BOM request.
- GPU
- 8× NVIDIA B300HGX B300, NVLink. 288 GB HBM3e each
- CPU
- 2× Xeon 6767P128 cores, 256 threads
- Memory
- 3,072 GBDDR5-6400 ECC, 32 × 96 GB
- Local storage
- 15.36 TB NVMe4 × 3.84 TB U.2, plus 2 × 960 GB boot
- Compute fabric
- 8 × 800 Gb/sInfiniBand XDR, one link per GPU
- Front end
- 1 × 400 GbENVIDIA BlueField-3 DPU
- Out-of-band
- 1 GbEDedicated BMC network
- Power
- 5+5 × 3,000 WRedundant supplies
- Cooling
- Direct liquidIn-rack coolant distribution
- Operating system
- Ubuntu 24.04NVIDIA drivers, provisioned by Helios. Bare-metal SSH access
Cluster
The 16-node reference cluster in four racks.
- Compute
- 16 nodes, 128 GPUs2 racks of 8 nodes, each with 6 switched PDUs and an in-rack CDU
- Compute fabric
- 1 × Quantum-X800InfiniBand XDR, 1:1 non-blocking, 128 links
- Front end
- Spectrum-4400 GbE to every node, Spectrum-3 aggregation and uplinks
- Out-of-band
- 1 GbESeparate management switching
- Network racks
- 2Switching, power and patching
What the BOM covers
In the BOM
- 16 GPU nodes as specified
- Compute racks, switched PDUs and in-rack CDUs
- Compute fabric with optics and cabling
- Front-end, aggregation and out-of-band networks
- System build, rack integration and burn-in
- Ubuntu and NVIDIA drivers, provisioned by Helios
Priced separately
- Storage clusters: WEKA, DDN or VAST
- Management and control-plane nodes
- Freight and installation
If a drive, memory module, optic or cable is supply-constrained, we may use an equivalent part. The GPU model and count, memory and storage capacity, and fabric bandwidth in your order form do not change.
Need a different build?
Tell us the workload, cluster size and timing. We'll scope CPU, storage and fabric with you. Preparing a request? Use the GPU cluster reservation checklist.