Jouni Osmala (a.delete@this.b.com) on July 24, 2020 9:22 pm wrote:
> > > Yep. Every second SIMD related conversation seems to bring out the games.
> >
> > It's probably a side-effect of SIMD having had a significant impact on game performance
> > when it was first introduced ages ago and GPUs didn't even have hardware T&L.

Maybe back then significant parts of the games were still programmed in assembler and optimization was more focused on small scale tricks. Personally, I don't remember ever rooting for MMX, 3DNow! or SSE capable processor, because some game made use of SIMD.

> And lecture made by Ubisoft Montreal employee about how to make faster code
> by organizing data in a way that allows use of Vector extensions.

Organizing data is useful even without vector extensions. In fact, it's probably more useful, since much of the code can't effectively make use of SIMD, but making your caches and memory controller happy will benefit any code. Especially for consoles, where CPUs have traditionally been pretty weak when it comes to branch prediction and caches (PS5 and new Xbox of course change that).

> The difference between experience of user of Game's and developer of Games is that developer
> has to think I want to run fast enough with these processors that users have what can I do with
> it. And after developer has cut features and tuned it, the user gets a game that run's fast
> enough with the common processors and it doesn't matter what processor they have since GPU is
> going to become the bottleneck. Developer on the other hand has to live with those limitations
> when designing the game. AVX-512 would be nice when it becomes old enough to be
> in processors that are the limiting factor of what game can do. But to get that
> point it has to be in the market in lowest end of Gaming systems for years.

There are many things that would be nice. Most games are now 64 bit, both PS4 and Xbox One are x64 with SSE and AVX. If SIMD was as useful (and as easy, though that's really part of it being useful) as people here seem to think, it would be used more. And even if AVX-512 makes programming easier (I'm not convinced that the difference is meaningful for a typical C++ or C# programmer), having to deal with 512 bits instead of just 128 means it's harder to get close to peak numbers.

Game optimization, for the large part, doesn't have anything to do with getting 2-4 times better speed in some small loop thanks to clever use of assembler or SIMD. It's mostly about having good tools for profiling and when a bottle neck is discovered, doing something different at that point. Often it's like: "Oh, this part has never before caused any problems. Do we actually need to run that code?" And then instead of doing the stupid thing every frame you cache the result and the code speeds by several orders of magnitude, or you ask the artists if they actually meant to use that really expensive effect that isn't even visible, or something like that.

It's very rare that optimizing the code at instruction level makes sense from cost/benefit perspective (it's very different on GPU side).

> Based on Sony:s PS5 presentation modern CPU:s are potentially becoming limiting factor for
> some games ported from PS5 to PC. Of course GPU manufacturers may find ways to eat that work
> also. And Game developers may need to cut more to be able to run on lower end PC:s.

Tl;dr: Sure, there's a great leap in CPU power, but I wouldn't worry too much as a PC gamer. It's not really that long a leap, and it's more like a brisk walk than a leap anyway.

Let's take a look at Steam hardware survey once again. About 20 % are dual-core. They can't really run most of the new AAA games. Unless we are making something especially casual, we don't need to worry about them. 47 % are quad-core, 23 % are hex and 7 % octa. Relative number of quads are going down and hex and octa cores going up. Even the quad-cores found in laptops will run current generation console ports fine, probably even if console version targeted 30 FPS and on PC you'd like to get 60.

The next generation, PS5 with it's 3.5 GHz and Xbox with 3.8 GHz Zen 2, will be much harder target to match. If studios keep making 30 FPS games on consoles, it will be hard to run them at 60 FPS on PC. It will be even harder to run them at 144 or 165 FPS, as some people like to do. But it's not that simple comparison.

First of all, the next generation is not yet out and PC market is a moving target. Getting a CPU that's significantly, if not overwhelmingly faster than those console Zen 2s, isn't all that expensive. Consoles will definitely reserve at least one core for OS, which will further level the playing field.

Second, Microsoft will keep selling current, lower performance Xbox and even if Sony hasn't said anything specific, it doesn't look like PS4 is going away all that fast. It's likely that the move to next generation will be slow. If target on current generation is 30 FPS, it will be trivial to run at 60 FPS (from CPU point of view) on next generation. At the same time though, PC version won't be presumed to run at 2-5 times higher FPS, so it won't be a problem.

(The above is actually pretty interesting from HDD vs. SSD point of view, and if it means that gaming PC will finally need an SSD for game storage. PS5 exclusives probably won't be able to run all that well from an HDD, but if PS4 and current generation Xbox One aren't going away, how many PS5 exclusives are there going to be? As for Xbox, MS is talking like every game in forseeable future will run on base system, so maybe there won't be Xbox One Series X exclusives at all.)

All in all, it's much easier to make a good port of a console game to PC these days than it used to be. Graphics APIs are a bit different (if you target DirectX 11 on PC) and may couse some issues with low-effort ports, but in general consoles are quite boring. There's no need to come up with something special for Cell SPU code. In fact, the most interesting thing about the next generation are probably the storage systems and support for hardware texture decompression. Everything else can be countered with a bit more CPU and a bit more memory on PC side (and a lot more GPU will go far beyond).

-JLarja