Rumor | NVIDIA Ampere to offer 10-20% IPC increase over Turing, 4x RT performance with minimal FPS impact, up to 2 GHz OC clocks, and an overhauled software stack to take on AMD RDNA 2.0
Recently, we've seen references to the first Ampere machine — the NVIDIA DGX A100 pop-up on a trademark site. The DGX A100 is expected to be powered by the GA100 GPU, which is purported to be the complete Ampere die. While NVIDIA's upcoming GTC 2020 keynote will concentrate primarily on high-performance computing (HPC), a new leak has surfaced that may please prospective owners of GeForce Ampere cards.
Earlier this week, Tom from the Moore's Law is Dead YouTube channel said that he had access to some exclusive insider info pertaining to Ampere. According to Tom, Ampere will cater to both HPC and gaming segments and is a full evolution of Turing rather that just a die shrink with more RT cores. Ampere will bring support for DLSS 3.0 that should work with any game that uses TAA.
Tom also quotes his sources as saying that NVIDIA will introduce an NVCache that makes use of the system RAM and SSD similar to AMD's HBCC to speed up load times and optimize VRAM usage. VRAM increases are not expected to be very high but expect 10 GB for the RTX 3080 and 12 GB for the RTX 3080 Ti. There is a possibility that GeForce cards could be fabbed on Samsung's 8nm process while the HPC variants could go the TSMC 7nm route.
The more juicy information is that Ampere could offer anywhere between a 10 and 20% IPC increase over Turing and double the Tensor cores per SM. Ampere will also offer 4x the ray tracing performance per tier. Tom says that this need not necessarily imply 4x the number of physical RT cores per SKU. Nevertheless, whatever be the increase in number of RT cores, we can expect ray tracing to have much less of a performance impact in games than it is today.
According to his sources, an RTX 3060 should be able to provide ray tracing performance equivalent to the RTX 2080 Ti. The actual raster performance is still anybody's guess at this point, though. Core clocks around the 1,900 MHz mark should be common, but 2 GHz can be attained by manual overclocking. Power consumption per tier is expected to be lower than Turing, but actual numbers aren't yet available.
If the above information is indeed true, what this essentially means is that Turing will not age well in comparison to the upcoming Ampere cards. Those who have spent quite a lot for the high-end RTX 2080 Super and RTX 2080 Ti may find that a US$300-350 Ampere card might very well offer a similar performance incentive. Tom's sources say that "Turing doing RT will be like Kepler doing DX12".
Oh, and not to forget, apparently there will be no GTX cards in the entire Ampere lineup with even the lowest spec cards speculated to feature at least a few RT cores. A speculated potential overhaul of the software stack that will now integrate both the GeForce Experience and GeForce Control Panel is also on the anvil.
Tom's sources say NVIDIA is likely to announce Ampere HPC products during the GTC 2020 pre-recorded keynote with the consumer GeForce lineup expected to be announced sometime in September. Needless to say, we are super stoked.