Quote:
Here's the thing --
In order to get the gamer what he or she wants, a company would need to test all of their hardware to make sure things simply work and regression test all games from the past couple years with that hardware. This is never going to happen.
Both AMD and NVIDIA cut major corners.
AMD has multiple driver trunks, and a rotating schedule of only about 2 dozen titles for regression testing that do not repeat within something like 6 months to a year. If one driver breaks something, it will be at least two months until it gets "fixed" for real (as the next month might not exhibit the same problem, but it also won't necessarily be based on the same code). If one game is dropped from the regression testing schedule, it could break in one driver and not be caught for a very long time (or until reviewers start to complain about it).
We run into the problem a lot, especially with crossfire scaling, of things that used to work not working and then when we point it out, all of a sudden we've got a beta driver with a fix. It's just because AMD dropped that title from their regression testing. Over time they hit a lot of titles, but this is an optimization that does cause issues.
Because you can't have monthly WHQL drivers with the latest features on the latest hardware all regression tested on everything everyone could want to play. There isn't enough time there to do all the necessary QA. All the driver has to do is pass Microsoft's WHQL testing ... which is easier than actually working in all relevant games.
NVIDIA does something else --
They don't always test all their hardware every cycle. We'll see beta drivers tested completely first on high end hardware or newly released hardware. Older stuff is left out of testing, so we have divergent driver versions necessary for different classes of hardware. Since NVIDIA uses a unified model (at the moment), all drivers should work on all hardware, but if it hasn't been released to support a specific card then that means it hasn't been QA'd on that card.
NVIDIA regression tests with many more titles per WHQL release, but at the same time, there is much more time between WHQL releases. This gives them a longer period to look at more things, but at the same time stuff can stay broken for longer.
If all this stuff is really interesting, I could take some time and talk to AMD and NVIDIA again (we've had this discussion with them before) and I could do a write up about it explaining the pros and cons of both approaches.
Frankly, from my perspective, monthly WHQL is just a marketing tool ... it makes people feel better. But NVIDIA's approach isn't necessarily better -- it's just different.