Back to GPUs
GB200 NVL4
Data Center
NVIDIA - Blackwell Architecture
HPC
AI
data-center
enterprise
LLM
supercomputing
NVIDIA's revolutionary Grace Blackwell Superchip that unlocks the future of converged HPC and AI, delivering exceptional performance through four NVIDIA NVLink-connected Blackwell GPUs unified with two Grace CPUs over NVLink-C2C.
Launch Date
March 18, 2024
Launch MSRP
$40,000
Performance
Delivers 30X faster real-time LLM inference performance for trillion-parameter language models compared to H100, and 4X faster training for large language models at scale.
Technical Specifications
CUDA Cores:
18,500Tensor Cores:
720Ray Tracing Cores:
N/ABase Clock:
1.9 GHzBoost Clock:
2.5 GHzMemory:
192 GB HBM3eMemory Bus:
6144-bitMemory Bandwidth:
8,000 GB/sTransistors:
208 billionDie Size:
850 mm²TDP:
1,200 WKey Features
- HBM3e memory with unprecedented bandwidth
- 5th generation NVLink: 1.8TB/s
- PCIe Gen6: 256GB/s
- Second-generation Transformer Engine with FP4/FP8 precision
- Dedicated decompression engines
- Multi-Instance GPU (MIG) support