NVIDIA B200 GPU
Coming Soon
NVIDIA's next-generation datacenter GPU. 192GB HBM3e at 8 TB/s bandwidth with 2nd-gen Transformer Engine and native FP4 support. Up to 4x faster LLM inference vs H100.
Powering teams that push boundaries
Trusted by companies including: Tesla, Hugging Face, Kaggle, Zoho, Weights & Biases, upGrad, Saama
Next-generation performance at every level
The B200 delivers generational leaps in memory, bandwidth, and compute for the largest AI workloads.
192GB HBM3e Memory
2.4x H100's 80GB. Run Llama 70B in full FP16 with 52GB to spare. Serve multiple large models simultaneously.
8 TB/s Memory Bandwidth
2.4x H100's 3.35 TB/s. Memory-bound workloads (most LLM inference) scale almost linearly with bandwidth.
2nd-Gen Transformer Engine
Native FP4 precision. Half the memory of FP8, enabling even larger models or bigger batch sizes.
1.8 TB/s NVLink
2x H100's NVLink bandwidth. Faster distributed training, more efficient tensor parallelism.
Key specs at a glance
Available today on JarvisLabs
While B200 is coming soon, these GPUs are ready for immediate use with per-minute billing.
What B200 enables
From next-gen training to ultra-scale inference and multi-modal AI.
Next-Gen LLM Training
Train 200B+ models with FP4/FP8 mixed precision. 4x speedup vs H100 reduces training costs proportionally.
Ultra-Scale Inference
Serve 70B models with massive batch sizes. 192GB handles multiple concurrent models per GPU.
Long-Context Applications
192GB accommodates enormous KV caches for 200K+ token contexts without memory pressure.
Multi-Modal AI
Image, video, and language models simultaneously. The memory headroom enables complex multi-modal pipelines.
B200 vs H100 vs H200
How Blackwell compares to the current Hopper generation.
| Specification | B200 | H200 | H100 |
|---|---|---|---|
| Memory | 192 GB HBM3e | 141 GB HBM3e | 80 GB HBM3 |
| Bandwidth | 8,000 GB/s | 4,800 GB/s | 3,350 GB/s |
| Tensor Perf (FP16) | ~2,000+ TFLOPS | 989 TFLOPS | 989 TFLOPS |
| NVLink | 1,800 GB/s | 900 GB/s | 900 GB/s |
| TDP | 1,000W | 700W | 700W |
| Transformer Engine | 2nd gen (FP4/FP8) | 1st gen (FP8) | 1st gen (FP8) |
| Architecture | Blackwell | Hopper | Hopper |
Technical specifications
Complete hardware specifications for the NVIDIA B200 data center GPU.
| Specification | Value | Note |
|---|---|---|
| Architecture | NVIDIA Blackwell (GB200) | Next-gen datacenter GPU |
| Manufacturing | TSMC 4NP | Advanced process node |
| Transistors | 208 billion | 2.6x H100 (80B) |
| VRAM | 192 GB HBM3e | 2.4x H100's 80GB |
| Memory Bandwidth | 8,000 GB/s (8 TB/s) | 2.4x H100 |
| Tensor Cores | 5th gen with FP4 native | FP4/FP8/FP16/BF16 |
| FP8 Tensor | 2nd generation | Higher throughput than Hopper |
| FP16 Tensor | ~2,000+ TFLOPS (est.) | Mixed-precision training |
| Transformer Engine | 2nd generation | FP4/FP8/FP16 dynamic |
| NVLink | 5th gen, 1,800 GB/s (1.8 TB/s) | 2x H100 NVLink |
| TDP | Up to 1,000W | 43% more than H100 |
Frequently asked questions
Everything you need to know about the NVIDIA B200 and when it's coming to JarvisLabs.
We'll add B200 instances as hardware becomes available. H100 and H200 are available today for immediate use.
Get notified when B200 is available
192GB HBM3e. 8 TB/s bandwidth. 2nd-gen Transformer Engine. Sign up to be the first to know.