Will AMD Carrizo offer a considerable performance increase over Kaveri?

[AdderDee]

So I looked up more about Carrizo and realized that "Excavator is expected to support new instructions such as AVX2, BMI2 and RDRAND." ( http://en.wikipedia.org/wiki/Excavator_(...hitecture) )

The main thing that caught my eye here was "AVX2." I've seen a lot of people talking about AVX2 on Dolphin's GitHub page and about how useful it is and I'm wondering - could this combined with HSA offer a considerable increase in performance in Dolphin over Kaveri? I know that AVX2 is quite important to Dolphin (although I'm not savvy enough about it to know exactly why) and HSA allows the GPU and CPU cores to work together on the same tasks (which would naturally result in a significant performance increase in single-threaded and dual-threaded applications.) So what do you all think?

EDIT: Reposted this because I thought the poll would cause a bit of a row

[KHg8m3r]

While AVX2 is a fast way for code to go through a CPU, the true power in Intel is the very fast single core speeds. Unless AMD pulls a rabbit out of their hat, their single-core speeds won't see some miraculous increase. And HSA probably won't be implemented in Dolphin due to how dolphin uses CPU cores. Maybe if a dev reaaaaally wants it they may try to implement it, but I doubt it, because a CPU is faster at the type of calculations that dolphin needs for the CPU core than a GPU can give (at least that's the way it was explained to me)

[admin89]

It's still fast enough for most games . Don't tempt me ! I might buy it for real

Spoiler: (Show Spoiler)

Just Kidding , I don't care about desktop anymore

[KHg8m3r]

But probably not the demanding games.
We'll have to wait and see the results of benchmarking.

[Fiora]

It doesn't feel to me like AVX2 will help much; the main thing that comes with AVX2 is 256-bit integer SIMD, but the only SIMD in Dolphin (at least on the CPU thread) is for paired singles, which isn't wide enough for that.

I guess FMA could help a little bit; I have a patch in for that but I doubt it's going to be huge. I don't think that much time overall is spent in paired float in typical games?

[DatKid20]

It probably will actually be good due to the CEO from back when AMD was beating Intel came back plus AMD has publicly said that Bulldozer wasn't the game changer they expected.

[MayImilae]

Moved to hardware. Otherwise, how would NV notice it?

[tueidj]

So disappointed Excavator will be the last of the line, was looking forward to seeing a line-up called Wheelbarrow (seriously these were the stupidest codenames ever).

[NaturalViolence]

MaJoR Wrote:Moved to hardware. Otherwise, how would NV notice it?

But I don't normally read hardware.

Spoiler: (Show Spoiler)

I happened to see the thread title on the front page of the forums and it peaked my interest. That doesn't count.

Everyone knows the standard method to summon me is to repeat my name 5 times into a mirror while drawing a pentagram on the floor around you with acid. Or you know, PM me. That works too.

@AdderDee

Our graphics devs have stated that HSA and mantle are next to useless (at least in their current form) and probably not worth implementing. And our cpu devs have stated AVX2 is next to useless and probably not worth implementing. So I don't know where you're getting your information from. Not that this should be particularly surprising. I don't know of any ISA extensions that have ever yielded more than a few percent improvement at the absolute most. Generally new ISA extensions just add more instructions that are useless to 99.99% of applications. All except for very very specific circumstances. Of course that doesn't stop them from always marketing them as some great new revolutionary feature. I really doubt that any x86 ISA extension will ever exist that will "supercharge" dolphin.

And HSA is at the moment more marketing than substance. Other than AMD constantly telling us about how it's going to change the world of programming and performance as we know it (and we all know how it turned out the last half dozen times they told us that) we don't really have many details about its future implementation.

Generally these "features" are so pointless that they are rarely factored into discussions about upcoming microarchitectures at all.

The actual details that matter won't be revealed until weeks before launch (several months away at this point). And even those aren't useful enough to make any accurate performance predictions with. It is impossible to tell how fast a chip is going to be until it is released and you can actually test it yourself. But despite this every year we always get threads with people asking us to tell them how the next Intel/AMD chips are going to perform even when they're months away from release. As if we have some secret prototypes hidden away in our basements to test with. Or that we can predict performance from vague marketing statements. Sometimes Intel/AMD release benchmarking data early. But it's usually BS or irrelevant to real applications (especially when it's AMD, they've grossly exaggerated their future products performance every year for the past 7 years now).

The best you can do is make predictions based on recent trends. And if we do that the future does not look bright for AMD and dolphin. Excavator is an improved variant of the bulldozer microarchitecture family. Next in line after steamroller. Since bulldozer based microarchitectures have never performed well with dolphin compared to Intels offerings that right there is a bad sign. The only thing AMD can really do to boost dolphins performance is boost core IPC significantly without a drop in clock rate. AMD has not significantly boosted IPC from one arch to the next in 5 years now. I don't really see any evidence that this pattern will change with excavator considering their increasing focus in GPU performance and core counts both in investor presentations and product releases. This is just my 2 cents though and is not an accurate means of predicting the performance of future chips. They could always surprise us.

DatKid20 Wrote:It probably will actually be good due to the CEO from back when AMD was beating Intel came back

AMDs current CEO (Rory Read, appointed in 2011) has never worked at AMD before. So I don't know where you got that from. Besides a good CEO wouldn't necessarily guarantee higher quality chips. In fact they usually have little relevance to that. Unless AMD can magically pull 150 billion dollars out of their hat I doubt they're be producing chips on par with Intel anytime soon. The shifts Rory Read is trying to make to the direction of the company are pretty much the exact opposite of what you would want for a company developing CPUs intended to be used with emulators. He has pushed the company to focus more on mobile and ultra-mobile platforms and shift away from desktops. And he has continued Dirk Meyers trend of moving product designs towards more parallel forms of processing (cpus with more cores and bigger IGPs, instead of bigger higher IPC cores). These things may turn out to be good decisions for the company in the long run, but they're certainly not good for those of us posting here looking for a computer to run dolphin with.

DatKid20 Wrote:plus AMD has publicly said that Bulldozer wasn't the game changer they expected.

That doesn't really matter either. Just because AMD finally admitted that bulldozer dramatically underperformed doesn't mean future architectures will suddenly start being more competitive.

tueidj Wrote:So disappointed Excavator will be the last of the line, was looking forward to seeing a line-up called Wheelbarrow (seriously these were the stupidest codenames ever).

Why would you want another bulldozer based uarch instead of a new one?

Now look what you've done MaJoR!

[tueidj]

(09-11-2014, 03:33 PM)NaturalViolence Wrote: The actual details that matter won't be revealed until weeks before launch (several months away at this point). And even those aren't useful enough to make any accurate performance predictions with. It is impossible to tell how fast a chip is going to be until it is released and you can actually test it yourself. But despite this every year we always get threads with people asking us to tell them how the next Intel/AMD chips are going to perform even when they're months away from release. As if we have some secret prototypes hidden away in our basements to test with.

You might be surprised to know that both Intel and AMD do often ship pre-release machines to both commercial and open-source developers (free of charge) in order for them to optimize their software before the chips reach market. It's simply a matter of having the right contacts which normally occurs by either filing bugs or asking for help/information regarding new features, combined with a solid reputation for the project (such as being used as a benchmark on well known-websites...).