Until now, the NVIDIA A100 calculation accelerator with Ampere architecture posed in photos and sketches only in the form of SXM4 modules used in the server segment. There was not even the slightest hint that it would appear in the form of a traditional expansion card with a PCI Express interface. A new entry in the CUDA 11 software toolkit brings such hope.
Analyzing the contents of the CUDA 11 update that came out on June 5th, one famous Japanese blogger found that next to two very similar versions of the A100 in the SXM4 version with 40 GB of HBM2 memory, a new modification of the accelerator with a designation of the form “A100-PG509-200” is mentioned. A similar designation structure is typical for expansion cards with PCI Express interface, and in the case of Ampere we can talk about interface version 4.0.
Apparently, NVIDIA is preparing a modification of the A100 for workstations or entry-level servers, which will be installed in the expansion slot PCI Express x16 motherboard. As already noted, the lack of blocks for hardware acceleration of ray tracing makes the GA100 unsuitable for use in games, but such an expansion card is still of theoretical interest.
It is worth adding that in the comments on the CUDA 11 update, Ampere accelerators are mentioned under a double designation: both in the form of the A100 module, shown in the first illustration, and in the form of the GA100 GPU itself. This separation only strengthens the belief that the flagship product of the Ampere family will not only exist as SXM4 modules.
If you notice an error, select it with the mouse and press CTRL + ENTER.