Product Overview
The NVIDIA B100, released in 2024, is the entry-level data center GPU of the Blackwell architecture. It uses a dual-die design interconnected via a 10 TB/s NV-HBI internal bridge. With a relatively modest 700W TDP, the B100 can drop directly into existing H100/H200 server baseboards, making it a top choice for cloud providers deploying Blackwell at scale.
Important: The B100 was largely surpassed by the B200 in actual 2024–2025 deployments. Many cloud providers (e.g., Modal, CoreWeave) skipped the B100 entirely.
Core Specifications
| Parameter | Value |
|---|
| Architecture | Blackwell GB100 |
| Process Node | TSMC 4NP |
| Transistor Count | 208 billion (dual-die) |
| Memory | 192 GB HBM3e |
| Memory Bandwidth | 8 TB/s |
| FP4 Tensor Core | 14 PFLOPS (sparse) |
| FP6 Tensor Core | ~9.3 PFLOPS (sparse, estimated) |
| FP8 Tensor Core | 7 PFLOPS (sparse) |
| FP16 Tensor Core | 3.5 PFLOPS (sparse) |
| FP64 Tensor Core | 30 TFLOPS |
| NVLink | 1.8 TB/s (5th Gen) |
| TDP | 700 W |
| PCIe | Gen 5 |
| Form Factor | SXM |
B100 vs B200 Key Differences
| Metric | B100 | B200 | Advantage |
|---|
| TDP | 700 W | 1,000 W | B100 lower |
| FP4 Compute | 14 PFLOPS | 18 PFLOPS | B200 +28% |
| FP8 Compute | 7 PFLOPS | 9 PFLOPS | B200 +28% |
| Memory | 192 GB HBM3e | 192 GB HBM3e | Same |
| Memory Bandwidth | 8 TB/s | 8 TB/s | Same |
| Server Compat. | Fits H100/H200 baseboard | Requires new server | B100 more flexible |
| Price (Reference) | N/A | $5.87/hr (cloud) | — |
Software & Drivers
Key Features
- 5th Gen Tensor Cores: Native FP4 / FP6 precision support
- 2nd Gen Transformer Engine: Automatic FP4 precision conversion
- NVLink 5.0: 1.8 TB/s GPU-to-GPU interconnect
- RAS Engine: Reliability, Availability, Serviceability
- Confidential Computing: Hardware-level TEE support
Use Cases
- LLM training and inference
- MoE (Mixture of Experts) models
- Recommendation systems
- Incremental upgrade of existing H100 clusters