NVIDIA Corporation
香港商輝達香港控股有限公司台灣分公司
NVIDIA Grace Blackwell 200 NVL72
Winning Reason
1. NVIDIA's latest DGX server, GB200 NV72, scales generative AI up to the trillion-parameter level for a 30x increase in the inference speed of large language models (LLM) over the previous generation, 4x faster training speeds, and a 25x reduction in the total cost of ownership (TCO) and energy consumption, proving so popular that demand is outstripping supply.
2. The new model incorporates a 5th-generation NVLink with liquid cooling into a single large 72-GPU machine, saving AI data center space and reducing water usage.
3. The NVLink-C2C increases the speed of critical database queries to 18x that of a CPU, reducing total ownership cost by 5x.
Product Feature
GB200 NVL72 is unlocking real-time trillion-parameter models in an energy efficient way by connecting 36 Grace CPUs and 72 Blackwell GPUs in a rack-scale design. The GB200 NVL72 rack scale solution is liquid-cooled and boasts a 72-GPU NVLink domain that acts as a single massive GPU that delivers 4X faster LLM Training, 30X faster real-time trillion-parameter LLM inference, with 25X lower energy.
The GB200 Grace Blackwell Superchip is a key component of the NVIDIA GB200 NVL72, connecting two high-performance NVIDIA Blackwell Tensor Core GPUs and an NVIDIA Grace CPU using the NVIDIA® NVLink® C2C interconnect to the two Blackwell GPUs. GB200 NVL72 utilizes the NVLink Switch system to provide 130 terabytes per second (TB/s) of low-latency GPU communications for AI and HPC workloads.