Product Overview
The NVIDIA B200 is the flagship GPU of the Blackwell architecture, with 18 PFLOPS FP4 (sparse) per GPU — the most powerful AI training and inference GPU as of 2025. Like the B100, it uses a dual-die design, but TDP climbs to 1,000 W, requiring redesigned servers and liquid cooling. NVIDIA also offers the GB200 Grace Blackwell Superchip (2× B200 + Grace CPU) and the NVL72 rack (72 B200 GPUs + 36 Grace CPUs).
Core Specifications
| Parameter | Value |
|---|
| Architecture | Blackwell GB100 (dual-die) |
| Process Node | TSMC 4NP |
| Transistor Count | 208 billion |
| Memory | 192 GB HBM3e |
| Memory Bandwidth | 8 TB/s |
| FP4 Tensor Core | 18 PFLOPS (sparse) |
| FP8 Tensor Core | 9 PFLOPS (sparse) |
| FP16 Tensor Core | 4.5 PFLOPS (sparse) |
| FP64 | 40 TFLOPS |
| NVLink | 1.8 TB/s (5th Gen) |
| TDP | 1,000 W (up to 1,200 W) |
| PCIe | Gen 5 |
Key B200 Systems
| System | Description |
|---|
| HGX B200 | 8× B200 server baseboard |
| GB200 | Grace CPU + 2× B200 Superchip |
| NVL72 | 72× B200 + 36× Grace, 1.4 ExaFLOPS FP4 per rack |
Software & Drivers
Key Features
- 2nd Gen Transformer Engine: FP4 / FP6 automatic precision switching
- Mixture of Experts (MoE) Optimization: Hardware-accelerated MoE model routing
- Confidential Computing: Hardware-level encrypted computation
- RAS Engine: Server-grade reliability
Use Cases
- Trillion-parameter LLM training
- Large-scale AI inference (trillion-parameter single-GPU inference)
- MoE models (Mixtral, Llama 4, etc.)
- Multimodal large models
- Scientific computing (HPC)