Reply With QuoteThere's not really any direct connection (although I'll talk about some indirect connections below). I think the devs would tell you that their plan is to keep the level of robustness *higher* than the proprietary driver anywaysOriginally Posted by allquixotic
The most obvious advantage of starting sooner is that we can finish sooner relative to HW launch.
What it also means is that there will be less time spent "learning the same lessons about hardware quirks from scratch" (since we'll all be sharing information as the new hardware is brought up for the first time) but that is traded off against the fact that developing earlier is harder because we are not able to rely on other teams having worked through HW issues and worked around them in VBIOS and microcode, and because testing needs to be done on simulators rather than on (faster) real hardware. On balance I expect we will come out a bit ahead.
The big difference I expect is that we will be less likely to get "stuck" on hardware issues the way we have been on SI at a couple of points, since everyone else will be working on the same HW at the same time as the open source devs. That's one of those "the worst case won't be as bad" advantages though, sort of like a higher minimum frame rate
I guess the most obvious point is that starting earlier doesn't actually give us more calendar time since we'll need to start work earlier on the *next* generation as well. We will have a bit more time since we won't be "catching up" any more, but that's only going to be maybe a 15% increase in per-generation development time.
I don't think we know the answer either, but I do expect that the smaller architectural changes from SI to subsequent parts should definitely make the next generation easier than SI.
Definitely not a pipe dream, but don't treat it as a given either.
There are a number of open questions right now :
First is how expensive the next round of performance improvements are going to be in the current open source driver stack once things like tiling and hyper-Z are enabled by default. My guess is that there should still be some low hanging fruit related to performance of slower games like Warsow.
The second is whether the combination of new shader architecture in SI and the use of LLVM in the open source driver's shader compiler will raise the general level of performance on shader-intensive apps -- we think it will but we don't have any testing to confirm or deny at this point.
The third is how close we can get power & thermal management to the proprietary driver.
Last is whether we will be able to usefully leverage the performance work done on the proprietary driver, since the two drivers have fairly different internal architectures (proprietary shares code across OSes, open shares code across HW vendors). If we are able to leverage some of the performance work then things get better than what I'm saying here, but since it's a "we don't know yet" I don't want to set any expectations.
We should know more about the first three over the next couple of months.
