The NVIDIA DGX A100 has been the premier GPU compute solution for several months now. In this generation, NVIDIA’s 8x GPU solution upgraded from 256GB of aggregate GPU memory (8x 32GB) to 320GB (8x 40GB.) Now, NVIDIA is doubling the per-GPU memory figure which in-turn doubles the capacity of the DGX A100.
NVIDIA DGX A100 640GB Edition
The new NVIDIA A100 80GB edition utilizes the SXM4 form factor. In the NVIDIA DGX A100 eight NVIDIA A100 80GB GPUs are placed on a HGX board with NVSwtiches to provide a communication fabric between all of the GPUs. Compared to the 4x GPU Redstone platform, the addition of switches allows for higher bandwidth on the fabric and much higher bandwidth than PCIe Gen4.
As we would expect, the NVIDIA DGX SuperPOD has been upgraded with the new systems.
NVIDIA says that Cambridge-1 will feature the DGX A100 640GB edition along with a few other recently announced supercomputer projects.
From the perspective of today’s announcements, this is probably the least exciting of NVIDIA’s bunch. This is simply building upon the NVIDIA A100 80GB announcement doubling the system’s HBM capacity. Realistically, this announcement is a factor of NVIDIA having the DGX product line instead of simply selling HGX platforms to partners to configure and sell. While some may see this as competitive with partner systems, the configuration that NVIDIA offers is more rigid. Oftentimes customers will want to deploy large customers with different types of configurations and support. That is why we see systems like the Inspur NF5488M5 we reviewed. Pricing is generally lower from partners and one has a broader opportunity to make configuration changes. Also, NVIDIA’s partners have 16x GPU systems while NVIDIA still only has 8x GPU DGX A100’s. We fully expect to see HGX-2 class machines with the A100 80GB achieving 1.28TB of memory.
For customers of the 40GB systems, NVIDIA says it is offering upgrade kits, but we did not get pricing information on those. Current customers will need to ask their reps for specific pricing and timing.