NVIDIA Ampere GA100 officially launched: the world's largest 7nm GPU, 8192 CUDA graphics cores, 48 ​​GB HBM2 vRAM, 20 times more powerful than Volta

With Ampere GA100, NVIDIA will continue to build its position on the HPC high-performance computer market, targeting customer segments using specific applications such as scientific research, AI, and Deep Learning. This GPU will also be shipped in various forms such as modular Mezzanine cards or common graphics cards with PCIe 4.0 interface standard. However, this GPU will mainly be integrated into the dedicated Tesla A100 graphics card, used on the powerful workstation DGX A100 and HGX A100.

Thanks to manufacturing on 7nm process, Ampere GA100 will have up to 54 billion transistors, about 2.5 times more than the figure of 21.1 billion of its predecessor Volta GV100 despite the size of only 826mm2, no difference. So much compared to 815mm2. This means that GA100 is the largest GPU as well as having the highest transistor density on the market today.

Equipped with 128 SM multithreading processing clusters with 8192 CUDA cores, GA100 will be a monster of computing power. Not to mention, running machine learning algorithms will achieve a lot higher performance thanks to 512 3rd generation tensor core cores. Of course, GA100's power consumption is also not good, up to 400W. GA100 is also expected to have multiple configurations to fit a variety of integrated hardware, but the amount of RAM will only reach a maximum of 48 GB. This is still a big step forward compared to 32 GB of the previous generation.

One of the main technological highlights is the arrival of PCIe 4.0 and the latest NVLink 3.0 connection. This is understandable because with the tremendous power of GA100, the connections also need to be upgraded accordingly to ensure no bottlenecks when moving data to be processed in multi-GPU settings.

At the moment, the GA100 will only be integrated on the Tesla A100 card series for use on the DGX A100 and HGX A100 super workstations with many configurations. While the HGX A100 series uses a 4-GPU configuration for a more affordable price, serving the needs of cloud servers or data centers, the DGX-A100 series will have multiple configurations from 1 to 8 GPUs to serving specialized research works on AI. And of course, their price is not cheap when the DGX-A100 series will cost from 199,000 USD.