NVIDIA B200 Specs B200 SXM6 HMB3E 192GB Blackwell GPU for AI (2026) Video Card
NVIDIA B200 Specs B200 SXM6 HMB3E 192GB Blackwell GPU for AI (2026) Video Card
NVIDIA B200 Specs B200 SXM6 HMB3E 192GB Blackwell GPU for AI (2026) Video Card
B200 Key Specifications
| Specification | B200 | H200 (for reference) | H100 (for reference) |
|---|---|---|---|
| Architecture | Blackwell (GB200) | Hopper | Hopper |
| Transistors | 208 billion | 80 billion | 80 billion |
| Memory | 192GB HBM3e | 141GB HBM3e | 80GB HBM3 |
| Memory Bandwidth | Up to 8 TB/s | 4.8 TB/s | 3.35 TB/s |
| FP4 Tensor Core | Yes (native) | No | No |
| FP8 Tensor Core | Yes (2nd gen) | Yes | Yes |
| Transformer Engine | 2nd generation | 1st generation | 1st generation |
| NVLink | 5th gen (1.8 TB/s) | 4th gen (900 GB/s) | 4th gen (900 GB/s) |
| TDP | Up to 1000W | Up to 700W | Up to 700W |
| Manufacturing | TSMC 4NP | TSMC 4N | TSMC 4N |
Architecture: What Blackwell Changes
Second-Generation Transformer Engine
The biggest improvement for AI workloads is the 2nd-gen Transformer Engine with native FP4 support:
- FP4 precision — 4-bit floating point for inference. Halves memory usage vs FP8, enabling larger models or higher batch sizes on a single GPU
- Dynamic precision management — automatically switches between FP4, FP8, and FP16 based on what each layer needs
- Higher throughput — combined with architectural improvements, NVIDIA claims up to 4x inference performance vs H100
FP4 is particularly impactful for LLM inference. A model that needs 80GB in FP8 on H100 would need only ~40GB in FP4 on B200, leaving 150GB+ free for KV cache and batching.
192GB HBM3e Memory
The memory jump is massive:
| GPU | Memory | Memory Bandwidth |
|---|---|---|
| H100 | 80GB HBM3 | 3.35 TB/s |
| H200 | 141GB HBM3e | 4.8 TB/s |
| B200 | 192GB HBM3e | Up to 8 TB/s |
192GB means:
- Llama 70B in FP16 fits on a single GPU (140GB) with 52GB to spare for KV cache
- Llama 70B in FP8 needs only ~70GB, leaving 122GB for massive batch sizes
- Llama 405B in FP4 could potentially fit on 2 B200s
- Multiple models served simultaneously from a single GPU
NVLink 5th Generation
NVLink bandwidth doubles from 900 GB/s (H100/H200) to 1.8 TB/s per GPU. For multi-GPU training, this means:
- Faster gradient synchronization during distributed training
- More efficient tensor parallelism for large model inference
- Better scaling efficiency when using 4-8 GPUs per node
GB200 and NVL72
NVIDIA is also shipping the B200 in pre-configured rack-scale systems:
- GB200 — a compute module with 2 B200 GPUs + 1 Grace CPU, connected via NVLink
- GB200 NVL72 — a full rack with 36 Grace CPUs and 72 B200 GPUs interconnected via NVLink, delivering 720 petaFLOPS of FP4 compute
These are designed for large-scale training and inference at the datacenter level.
B200 vs H100 vs H200
For LLM Inference
| Metric | B200 | H200 | H100 |
|---|---|---|---|
| Llama 70B (FP8) tokens/sec | ~4x H100* | ~1.9x H100 | 1x (baseline) |
| Memory for Llama 70B FP8 | 70GB (122GB free) | 70GB (71GB free) | 70GB (10GB free) |
| Memory for Llama 70B FP4 | ~35GB (157GB free) | N/A (no FP4) | N/A (no FP4) |
| Max batch size (70B FP8) | Very large | Moderate | Small |
*NVIDIA published claims. Real-world performance will vary by implementation and workload.
The B200's combination of more memory, higher bandwidth, and FP4 support could make single-GPU serving of 70B models practical at scale — something that's tight on H100 and comfortable but not optimal on H200.
For Training
NVIDIA claims up to 4x training performance on GPT-class models compared to H100, primarily from:
- Higher Tensor Core throughput
- FP8 training improvements (2nd-gen Transformer Engine)
- 2x NVLink bandwidth for better multi-GPU scaling
- More memory reducing the need for memory optimization techniques
For large model training, the B200 could reduce training time (and cost) by 3-4x compared to H100, assuming the software stack fully utilizes the new hardware features.