NVIDIA announced DGX GH200 in a release dated 28 May: 256 Grace Hopper superchips connected through NVLink, with 144 terabytes of shared memory. Availability remains expected by year end.
Photo: © 2020 Will Buckner, “Nvidia sign”, 31 October 2020. CC BY 2.0. Cropped; historical headquarters sign, not launch imagery. No endorsement implied.
NVIDIA’s technical account specifies 480 GB of CPU memory and 96 GB of HBM3 per superchip. GPU threads can address peer memory. That total combines CPU memory and GPU HBM.
NVIDIA targets recommendation models, graph neural networks and large analytics workloads. It calls DGX H100 the most performance-efficient training choice when workloads fit entirely within its aggregate GPU memory.
Helios would connect four systems through InfiniBand, beyond each system’s NVLink fabric. Operation is expected by year end.
Discover more from Sulat AI & Tech News
Subscribe to get the latest posts sent to your email.
