NVIDIA B300 Infrastructure & Deployment_
The B300 — Blackwell Ultra — is NVIDIA's current flagship GPU, with 288 GB of HBM3e and roughly 1.5x the FP4 throughput of the B200. Leviathan is actively deploying B300 infrastructure in GB300 NVL72 racks for hyperscale AI facilities in the US.
What Is the NVIDIA B300?_
The B300 is the Blackwell Ultra refresh of the B200. It keeps the dual-die Blackwell design but moves to 12-high HBM3e stacks for 288 GB per GPU (up from 192 GB), and raises FP4 throughput by ~50% — targeting long-context and reasoning-model inference where memory capacity is the bottleneck.
It ships as the 8-GPU HGX B300 and, at rack scale, as the GB300 NVL72 — four B300 GPUs per compute tray with two Grace CPUs, 72 GPUs per liquid-cooled rack. Higher per-GPU power makes direct-to-chip liquid cooling non-optional at density.
B300 Specifications_
| Specification | NVIDIA B300 |
|---|---|
| Architecture | Blackwell Ultra (dual-die, refreshed Blackwell) |
| Memory | 288 GB HBM3e (12-high stacks) — ~8 TB/s |
| FP4 Performance | ~1.5x B200 (dense), tuned for reasoning/inference |
| GPU Interconnect | NVLink 5 — 1.8 TB/s per GPU |
| TDP | Up to ~1,400 W (liquid cooling required at rack scale) |
| Common Form Factors | HGX B300 (8-GPU), GB300 NVL72 compute tray (4 GPUs/tray) |
| Power Architecture | Supports 800V DC distribution (GB300 NVL72) |
| Networking | ConnectX-8 800G / Quantum-X800 InfiniBand |
B200 → B300: What Changes for Deployment_
288 GB per GPU (20.7 TB per NVL72 rack) in the same mechanical envelope as GB200 — cooling and manifold designs carry over.
Increased per-GPU TDP tightens liquid-cooling tolerances. Flow rates, CDU sizing, and leak-detection commissioning matter more than ever.
GB300 builds can adopt 800V DC distribution and 800G ConnectX-8 networking — plan the power and back-end fabric for it on new builds.
Ready to Deploy Your GPU Infrastructure?_
Tell us about your project. Book a call and we’ll discuss scope, timeline, and the best approach for your deployment.
Book a Call