Much like every other public event this year, NVIDIA’s GTC 2020 has also been canceled due to COVID-19. Thankfully, the company will release a pre-recorded keynote which will have all the specifications of its upcoming Ampere GPU architecture. However, some details of a new DGX system, that’s powered by the Ampere architecture, have been spotted in a trademark database.
The name of the new system is DGX A100 and it was found by Komachi, a Twitter user who specializes in hardware info and leaks. Komachi found the DGX A100 listed at Justia Trademarks with all its relevant specifications.
Now, for those who aren’t familiar with NVIDIA’s naming convention, the DGX system has been made to solely cater to the Deep Learning and High Performance Computing (HPC) communities. It is an amazing system that has the ability to offer supercomputer levels of computing power in a small workstation form factor.
Before now, NVIDIA’s DGX systems were only based on its Pascal and Volta GPUs. With the introduction of the Ampere architecture, it was about time people anticipated a new DGX system and here it is. The final confirmation that the DGX A100 is based on the Amere architecture was found in the trademark listing where it’s mentioned that the new workstation is powered by a GA100 GPU.
That is great but will it be any good? In order to find out, let’s look at NVIDIA’s current “world’s largest GPU”. It featured DGX-2 based on the Volta architecture and had 16 Tesla V100 units inside that resulted in 81,920 CUDA cores. From that information that’s currently available, the Ampere based GA100 will have 8,192 CUDA cores, 256 RT cores, 1,024 tensor cores, and a boost clock of 1,750 MHz.
Those are definitely healthy figures and are a testament to the GA100’s power. The GA100 will also be featured on NVIDIA’s upcoming flagship RTX 3080 Ti. Other less powerful Ampere based GPUs such as the GA103 and GA104 will most likely power the RTX 3080 and RTX 3060, respectively.
Some leaked variants of the new GA100 GPU have been tested and are rated at around 30 TFLOPs (FP32). Regardless of the fact that these speculations are true or not, this has certainly hyped every GPU enthusiast out there, especially people from the HPC community. We’ll just have to wait till May 14th, 2020 when the keynote will be uploaded to find out.