- Pronouns
- She/Her
As far as we know it's an Ampere GPU but was tested and taped out alongside Ada Lovelace GPUs. Three things that I would consider relevant here:By the by, I'm probably a bit out of the loop, but I was wondering what the thing is with the Drake chip and whether it is solidly Ampere or if there are next-gen (Lovelace) features that we have been able to divine from the leak. Can someone help me understand this?
edit: Considering Ampere is on 8nm Samsung/7nm TSMC and we are looking at a potential move to a different node for Drake, that could also be some reason to suspect they might have Ada-fied the chip in the process, right (what do you think)?
One: Ada Lovelace is not a major version number change on the GPU side, in a literal sense, and a metaphorical sense in that it's more of an Ampere Double Plus rather than an incompatible replacement. As such I think the noise around whether T239 is "Truly Ada" or "Truly Ampere" is unhelpful. We already know it's both. It's based on Ampere, just like Ada, but has a different CUDA version, of which, Ada Lovelace is a superset if I understand CUDA version numbers correctly. The bottom line is that the difference in the GPU side is... Not all that relevant to a console. It can evidently do some thing Ampere cannot, and Ada Lovelace can do some things Drake cannot. However, those could be things like console specific features, backwards compatibility with Maxwell, etc., with Ada's benefits on top of that being minimal or PC specific. It doesn't mean much. Drake is neither purely Ampere nor purely Lovelace, and instead has a CUDA version BETWEEN the two. Bottom line, based on Ampere, not Ada.
Two: Node. Ada Lovelace is usually made on 4N, Ampere on 8nm. But, unlike 8nm Ampere products, Drake was taped out tested and produced alongside Ada Lovelace, and launches at the earliest in 2023. It would make sense for it to "inherit" the 4N process node from Ada Lovelace just because of when it was made. The fact that the size of the GPU only really makes sense if it's smaller than 8nm, and that Nvidia has plenty of 4N production on-hand, seem to all point to this being the most likely outcome; Drake is probably on 4N.
Three: Optical Flow Acceleration, much the same as CUDA core. We know it is a version above Ampere GPUs, but inherited instead from Ampere Orin SOCs. We can't know for certain whether this is truly "Ada" in nature, and thus DLSS 3.0 capable, but given the timeline of production, I would hazard a guess that it is somewhere in between, as with the rest. It is better than Ampere's, but it is not Ada's, and that isn't necessarily a bad thing. DLSS 3.0 does not confer significant benefit at 60FPS or below. Meanwhile, highly efficient DLSS 2.X would be a gigantic boon for the system, its secret sauce. Having an OFA good enough, fast enough, reliable enough to be used in automotive self driving is extremely encouraging. As an aside, Drake, unless I am mistaken, also includes Ada Lovelace video encoding for AV1, which should mean better video captures. This is something inherited from Ada Lovelace, rather than Ampere.
I think sometimes people miss the forest for the trees. This isn't an Ampere GPU, strictly. Nor is it really Ada Lovelace. It is a custom SOC for a very valuable, very performance sensitive, very price sensitive customer. It will have been made from day 1 with all that in mind. This is a custom chip, which what appears to be cherry picked elements from multiple generations of Nvidia GPU, as you would expect from a custom SOC for a video game console. There are things it needs that a PC does not, and things PCs need that it does not, and can discard for better efficiency.
While I know none of this is a definitive answer, my best advice would be to not worry about it.