Jukka Larja (roskakori2006.delete@this.gmail.com) on April 28, 2022 9:05 am wrote:
> John H (john.delete@this.not.com) on April 27, 2022 5:20 pm wrote:
[...]
> > I'm curious what kind of software behavior is causing the simulators to see such a large
> > improvement relative to other games and applications. Is it the main code game loop
> > is likely a bit too large for the existing caches to handle effectively? Is the X3D
> > cache not yet big enough or maybe already big enough for other game types?
[...]
>
> [...] Most games are so GPU limited that getting significant differencies requires running a high-end GPU with low-end resolution.
[...]

I'd always assumed two things:

- Cache thrashing due to lots of random memory accesses, possibly from pointer chasing. Zen 1/+/2 core complex L3 cache divisions vs. Intel or Zen 3 could explain some of this.

- Long chains of if/then conditionals, leading to branch throughput dependencies and/or branch predictor aliasing.

...although those are just speculation. I've never tried to dig up actual research on it, although I have seen bits of anecdotal evidence for the benefits of larger caches in games, e.g.:

https://www.youtube.com/watch?v=RQ6k-cQ94Rc

Note that he only tests games in this video (with a giant GPU at low resolution, of course), rather than a variety of applications. But then, you've already been looking at 5800X3D benchmarks.

Peter E. Fry