NVIDIA has released a large number of successful graphics cards in recent years, and the popular company is ready to impress the the market again. During a keynote at the digital GPU Technology Conference 2020, the company has made a series of announcements focused on a new CPU architecture and the first products that will make use of it.
As in the case of other GTC keynotes, the main focus was the announcement of new high-end products for enterprise-grade businesses. The Volta-powered GV100 GPU and V100 accelerators have been quite popular since their reveal a few years ago, and they are present in many data centers spread across the world.
NVIDIA managed to learn from the mistakes of the past, and many sensible details that have been leaked in the case of other unannounced architectures have remained a secret this time. Theories related to a new architecture called Ampere have been floating for some time, but there were no hints or clues until the official announcement was offered.
The star of the show was the the A100, the first GPU which sports the new architecture. It comes with an accelerator that has the same name for now, as it is the only accelerator that can harness the power of the GPU.
The A100 was designed to surpass the V100, and it succeeds without issues. With the use of advanced 7nm processes, NVIDIA has managed to cram 54 billion transistors, 2.5 times above the the ones present on the V100. As a result, the GPU die is even bigger than the one offered during the last generation, and it seems that it will get even bigger in the future.
It is worth noting that tensor performance was also boosted greatly, and new formats can be used. Memory bandwidth has been extended, and several stacks of HBM2 memory will convey a massive bandwidth of 1.6TB per second.
More general details about the Ampere architecture will be announced soon.