NVIDIA has announced the fourth generation DGX system – the DGX H100. Powered by new H100 GPU technology, the DGX H100 is set to be the standard in AI and HPC workloads in the datacentre with 6x more processing power and 2x faster networking than the current DGX A100. Similar in architecture to the DGX A100, the new DGX H100 features 8 x H100 GPUs with 80GB RAM.
In the announcement, NVIDIA expects to start shipping these units in CQ4 (Oct-Dec 2022), and prices will be confirmed closer to the shipment date.
- 8 x NVIDIA H100 GPUs with total of 640 GB GPU memory
- 7.2 TB/s bidirectional GPU to GPU bandwidth, 1.5x more than the DGX A100
- 1 TB/s of peak bidirectional network bandwidth on 8x NVIDIA ConnectX-7 and 2x NVIDIA Bluefield DPU 400 gigabits/s network interfaces
- Dual x86 CPUs and 2 TB of system memory
- 30 TB of NVMe
- Security enhancements and confidential computing architecture for secure deployment in cloud/private cloud environments
The DGX H100 represents a powerful evolution of processing power, network speed and memory in a single form factor. The DGX H100 includes the innovations included in the DGX A100, including Multi-Instance GPU – allowing each H100 GPU to be divided into up to 7 logically independent units, or 56 in total in the DGX H100. In the DGX H100, these MIG units deliver approximately 3x more compute power and 2x more memory bandwidth. This provides a powerful level of agility, where the system can be configured for analytics, training and inference as required.
The DGX H100 is also supported by the NGC software stack including Cuda, so that it is easy to port applications onto the DGX H100 from any other NVIDIA GPU based AI solution.
While we have to wait until the end of 2022 for the DGX H100 to start shipping, NVIDIA has announced that the H100 PCIe GPUs will be shipping before then. While no firm date has been set, NVIDIA has indicated that this may be available in CQ3 (July to Sep time frame).
See the full specifications on the DGX H100 here in our products section.
Contact XENON today to learn more about the DGX H100, and the other infrastructure solutions XENON offers to support AI workloads such as high speed storage and networking.Talk to a Solutions Architect