NVIDIA Ampere, the next gen graphics architecture from NVIDIA will supposedly have PCIe 4.0 support. While the Ampere GPUs are yet to be announced, NVIDIA did recently tease the HGX A100 platform on their YouTube channel. The news about the PCIe 4.0 support stems from a leaked FudZilla post regarding Supermicro’s upcoming server platform that went up ahead of embargo (autoscheduler error). The post has since then been taken down, which in the world of rumours is practically equivalent of a confirmation.
UPDATE: Confirmed. Official NVIDIA A100 specs have been released and they do have PCIe 4.0 support. Specs added below.
The NVIDIA HGX A100 is the latest iteration of the NVIDIA HGX platform that will be announced by Jensen Huang at the NVIDIA GTX 2020 keynote later today. The “A100” at the end signified that the platform is based on NVIDIA Ampere GPUs, in particular the A100 GPUs. The HGX A100 is a data centre platform that’s geared for HPC and AI applications. The Supermicro press release quotes two new AI machines based on the NVIDIA HGX platform – a 2U system (formerly codenamed Redstone) with 4 NVIDIA A100 GPUs and a 4U system (formerly codenamed Delta) with 8 NVIDIA A100 GPUs. The latter is supposedly capable of 5 Peta FLOPS of AI performance.
As per the screengrab of the now deleted post, the “Supermicro 2U system uses the NVIDIA HGX A100 4 GPU board with four direct-attached NVIDIA A100 Tensor Core GPUs using PCI-E 4.0 for maximum performance and NVIDIA NVLink for high-speed GPU-to-GPU interconnects.” The system has one PCI-E 4.0 x8 slot and up to four PCI-E 4.0 x16 expansion slots for more interconnects. So the NVIDIA Ampere architecture is capable of supporting PCIe 4.0. However, this does not mean that the NVIDIA Ampere consumer gaming graphics cards will ship with PCIe 4.0 interface.
This could go both ways. The Supermicro platform is built for AI and HPC applications. These are bandwidth heavy tasks that require very fast interconnects between the system memory and the GPUs to move large AI datasets around. This very much necessitates a faster interface such as PCIe 4.0. Gaming graphics cards, however, do not have the same bandwidth requirement.
Whatever is the successor to the Turing based NVIDIA RTX 2000 series GPUs, be it the RTX 3000 series of GPUs or something entirely different, gaming graphics cards are hardly capable of making the most of the existing PCIe 3.0 interface even when it comes to 4K gaming with all the graphics settings cranked up to the maximum. So providing a faster interface such as PCIe 4.0 is purely a gimmick. The only reason that the RTX 3000 graphics cards would have PCIe 4.0 support is if the entire PC industry decides that now’s the time to shift to PCIe 4.0. AMD did it to gain an advantage since their InfinityFabric architecture is PCIe based but Intel and NVIDIA don’t need to make the same move. So we’ll leave it to Jensen to confirm whenever he hits the stage to make the announcement for the Ampere based gaming graphics cards.
Here's the official specification for the NVIDIA Ampere A100 that was released at NVIDIA GTC 2020.
| NVIDIA A100 Specifications | |
|---|---|
| Transistor Count | 54 billion |
| Die Size | 826 mm2 |
| FP64 CUDA Cores | 3456 |
| FP32 CUDA Cores | 6912 |
| Tensor Cores | 432 |
| SMs | 108 |