With 288GB HBM3e — 50% more than B200 — the B300 enables training of trillion-parameter models with fewer GPUs. The expanded memory eliminates the need for aggressive model parallelism, reducing communication overhead and accelerating training time for frontier-scale models.
Purpose-built for the era of AI reasoning, the B300 excels at chain-of-thought inference workloads where massive KV-cache memory is critical. A single DGX B300 delivers 192 PFLOPS for inference, enabling real-time reasoning at scale for agentic AI systems.
The GB300 NVL72 rack achieves 1.1 ExaFLOPS — true exascale in a single node. This enables climate simulations, drug discovery pipelines, and physics research at unprecedented resolution without requiring multi-rack interconnects.
The B300's 288GB memory capacity supports next-generation video generation models and world simulators that require massive context windows. Train and serve models that generate minutes of coherent video or simulate complex 3D environments in real time.
| GPU Architecture | NVIDIA Blackwell Ultra |
| Transistor Count | 208 Billion (4NP Process) |
| Die Size | Dual-Die CoWoS-L (Reticle Limit x2) |
| Tensor Cores | 5th Gen (Enhanced) |
| Memory Capacity | 288 GB HBM3e |
| Memory Interface | 8192-bit (12-Hi Stacks) |
| Memory Bandwidth | 8.0 TB/s |
| NVLink Bandwidth | 1.8 TB/s |
| Form Factor | SXM6 |
| Thermal Design Power | 1400W (Liquid Cooled) |