News
Industry trends: Nvidia releases GH200 super GPU, opening the star sea of AI computing industry chain
Release time:
2023-07-19 17:29
On May 5, NVIDIA founder and CEO Jensen Huang officially released the new GH29 Grace Hopper superchip and the DGXGH2023 supercomputer with 200 GH256 superchips at the NVIDIA Computex 200 pre-show press conference.
The super chip GH200 has been fully put into production, and Google, Microsoft, and Meta will be the first to get it. The GH200 super chip uses NVLink-c2c interconnect technology to integrate an ARM-based energy-saving 72-core Grace CPU and a high-performance NVIDIA H100 GPU, 96GB HBM3 and 512GB LPDDR5X into the same package through the Chiplet process to provide a consistent memory model of CPU+GPU, eliminating the need for traditional CPU-to-GPU PCIe connections. The GH200 superchip offers up to 900GB/s of total bandwidth, providing a huge advantage for certain memory-constrained workloads. Compared to PCIe5, the chip increases the bandwidth between the GPU and CPU by 7 times, reducing the power consumption of the interconnect by more than 5 times. Huang said that the GH200 Grace Hopper super chip is fully operational and will power complex AI and high-performance computing workloads. Google Cloud, Meta and Microsoft will be the first to get the GH200.
The new NVIDIA Helios supercomputing will be launched at the end of the year, and the GPU data center is expected to replace the existing CPU data center in the future. DGXGH200 is a supercomputing power composed of 256 GH200 super chips, with up to 1EFLOP of computing power, and 144TB of shared memory (nearly 100 times more than the previous generation DGXA500). In addition, Nvidia said it is building its own large-scale AI supercomputer, NVIDIA Helios. The supercomputer will be equipped with four DGXGH4s, each of which will be interconnected via the NVIDIA Quantum-200 InfiniBand 2 Gb/s network to increase data throughput for training large AI models. In the future, GPUs are expected to replace current CPUs. According to NVIDIA's configuration, a data center with 400 CPUs can be built at a cost of $1000 million, but it needs 1GWh of power to process 960X large model data. But at the same cost, you can build a data center with 11 GPUs, only need 1.1GWh of power consumption, and can handle the data volume of 48X large models.
Based on the GH200 Grace Hopper super chip, NVIDIA and SoftBank also announced a collaboration to develop a new platform to meet generative AI and 5G/6G applications. Meanwhile, SoftBank plans to launch a new distributed AI data center in Japan. It is reported that the platform uses the GH200 super chip, which is expected to improve the performance, scalability and resource utilization of application workloads.
Recommended News
Service hotline
UNITS 09,9/F,LUEN CHEONG CANCENTRE NO.8 YIP WonG ROAD TUEN MUN N.T HONG KONG