PhysX on a CPU likely to see very little benefit from SSE recompile

[Amorphous]

http://scalibq.spaces.live.com/Blog/...0661!387.entry

The results speak for themselves. Although there certainly is some performance left on the table by nVidia, it is nowhere near as dramatic as Kanter claimed. In synthetic tests, there is about 8% to be gained from recompiling. This is nowhere near the 2-4x figure that Kanter was using. In fact, 8% faster PhysX processing would mean even less than 8% higher framerates in games, since PhysX is not the only CPU-intensive task in a game. Perhaps the net gain in framerate would be closer to 3-4%, depending on the game. In other words, recompiling PhysX with SSE would not make CPUs threaten GPU physics. Not even close. The difference would be lost in the margin of error, most likely.

[Calmatory]

Great job Nvidia PR warrior! Now get lost.

To anyone interested, the paragraph does NOT talk about recompiling PhysX to take advantage of SSE! It talks about recompiling Bullet physics to take advantage of SSE. It has NOTHING to do with PhysX. Well, still there will be people thinking that somone actually tested PhysX with SSE.

Considering the pro-Nvidia source, broguth by an Nvidia warrior... Meh, I'll stick to Box2D.

[Iconyu]

Explain WHY nvidia chose to use x87 in the first place then I might care about this blog. Otherwise it looks like nvidia went way way out of their way to make physx run 8% slower.

[Manicdan]

so much hate in that article. basically a smear campaign against Kanter,

it also goes on about the history of who own what, and its making us realize that physics in games is trying so hard to get money, instead of building quality.

[trinibwoy]

Explain WHY nvidia chose to use x87 in the first place then I might care about this blog. Otherwise it looks like nvidia went way way out of their way to make physx run 8% slower.

Nvidia didn't choose to use x87, they bought Novodex that way. They didn't go out of their way to do anything, they just didn't go out of their way to improve it. At least try to do some research guys before jumping on the Nvidia hating bandwagon

[Hornet331]

Great job Nvidia PR warrior! Now get lost.

To anyone interested, the paragraph does NOT talk about recompiling PhysX to take advantage of SSE! It talks about recompiling Bullet physics to take advantage of SSE. It has NOTHING to do with PhysX. Well, still there will be people thinking that somone actually tested PhysX with SSE.

Considering the pro-Nvidia source, broguth by an Nvidia warrior... Meh, I'll stick to Box2D.

lol indeed.. read a few posts and realized they where talking about bullet...

[Particle]

Hey, if you don't like the taste of French dressing then you don't like ranch either. They're both dressings after all and can thus be treated as if they're the same entity.

As with my real life example of dressing, the statement may be true. Just because Fact A, though, it isn't necessarily proof that Fact B.

[Carfax]

Explain WHY nvidia chose to use x87 in the first place then I might care about this blog. Otherwise it looks like nvidia went way way out of their way to make physx run 8% slower.

Nvidia didn't "choose" to use x87. People often seem to forget that PhysX was created by Ageia, and was designed primarily to run on dedicated hardware.. Software driven PhysX was merely an afterthought on the PC, hence why it wasn't optimized thoroughly unlike other software Physics APIs.

On consoles however, PhysX runs on the CPU and is optimized for it very well from what I've heard.

So intent needs to be taken into account here..

[Calmatory]

Great, the blog post states that sometimes x87 comes out on top. Since I don't have any deeper understanding of inner workings of Bullet, nor any real practical experience of SSE.. Is that even possible?

Can there be cases when SSE comes out slower than x87?

For example single precision matrix multiplication can be what, 4x faster in SSE than in x87(One can do 4 MULs parallel, right?)? I have hard time believing that any single operation implemented in SSE would be slower than x87 equivialient. The results.. sound fishy at best. Oh, and AVX could potentially double the single precision speed, allowing 8 parallel MULs. ;

[Manicdan]

if Physx had 3 things:
unlimited multi core support for those with quads, hexi, or 8 threads, on all games
and
the ability to adjust intensity with finite detail
and
for a dedicated gpu for physx to work on any brands video output (ati for gaming, nvidia for physx)

Physx would gain ALOT of peoples respect. for those who think that wouldnt help nvidia at all, its not really true. if things look great, and people want more, they would probably go and buy a dedicated card to get the most out of it. but it wouldnt take away the experience completely for the majority of the world and there wouldnt be so much disrespect sent their way. if bullet ever takes off, and as havok does great things since cpus are gaining lots of cores, but mhz are not increasing much, there will be plenty of extra power left on the cpu for havoc. both of those things will threaten their business, and they need to setting the standard, instead of trying to milk the market with laziness.

[Carfax]

Great job Nvidia PR warrior! Now get lost.

To anyone interested, the paragraph does NOT talk about recompiling PhysX to take advantage of SSE! It talks about recompiling Bullet physics to take advantage of SSE. It has NOTHING to do with PhysX. Well, still there will be people thinking that somone actually tested PhysX with SSE.

Considering the pro-Nvidia source, broguth by an Nvidia warrior... Meh, I'll stick to Box2D.

The comparison is still relevent, even if it's not PhysX. The point the author was making, is that the huge gains Kanter talked about in his article from "simply recompiling" is very unrealistic..

[Manicdan]

The comparison is still relevent, even if it's not PhysX. The point the author was making, is that the huge gains Kanter talked about in his article from "simply recompiling" is very unrealistic..

to sum the article up:
one troll attacks another troll, and we get a news post out of it, lol

[trinibwoy]

if Physx had 3 things:
unlimited multi core support for those with quads, hexi, or 8 threads, on all games
and
the ability to adjust intensity with finite detail
and
for a dedicated gpu for physx to work on any brands video output (ati for gaming, nvidia for physx)

Which other API gives you any one of those things (rhetorical question)? It's funny how PhysX has to be so awesome in order to get respect yet no other API has any of those qualities you just described.

[Manicdan]

Which other API gives you any one of those things (rhetorical question)? It's funny how PhysX has to be so awesome in order to get respect yet no other API has any of those qualities you just described.

havok is built to work on CPUs and is not overly saturated. look at HL2, just enough to play the game, even on 2ghz. mafia2 has effects that push hardware to the limits just cause smoke is so dense, or everyone and their mother has to wear a cape or trench coat.

phsyx is known for trying to put too much of something that adds nothing except a higher cost on your electric bill. also their arrogance, like not supporting those who bought a nvidia gpu to work with an ati as their main output. and disabling the older ageia PPU "just because"

havok is not the same as phsyx at all. both have different histories and different methods for implementation, and should be judged independently

[Carfax]

if Physx had 3 things:
unlimited multi core support for those with quads, hexi, or 8 threads, on all games

PhysX already has multicore support, but it's up to the developer to take advantage of it. Metro 2033 already demonstrated this.

[Manicdan]

PhysX already has multicore support, but it's up to the developer to take advantage of it. Metro 2033 already demonstrated this.

which is why i said all games.
im not saying they didnt do it, we already went over that before.
just saying nvidia needs to take care of it (and the article says they are, which is good)

i would like to point out that i do not hate nvidia, nor am i actually mad. i am discussing the issues that the general public has run into and the things holding back a feature that should be much "cooler" than it has been.

[Carfax]

mafia2 has effects that push hardware to the limits just cause smoke is so dense, or everyone and their mother has to wear a cape or trench coat.

I can't believe you're comparing Mafia 2 to HL2

The former uses scalable HARDWARE driven physics, while the latter uses software physics only.

Don't tell me that you're complaining that you can't max out the physics on Mafia 2 on your hexcore processor, when the game's physics was designed to be run on a GPU and not a CPU..

Seriously..

[Manicdan]

I can't believe you're comparing Mafia 2 to HL2

The former uses scalable HARDWARE driven physics, while the latter uses software physics only.

Don't tell me that you're complaining that you can't max out the physics on Mafia 2 on your hexcore processor, when the game's physics was designed to be run on a GPU and not a CPU..

Seriously..

you seriously need to read things again. havok is built for CPUs, and Physx is bult for GPUs, we both know this, and we both just acknowledged this.

you seemed to have ignored the last sentence
"havok is not the same as phsyx at all. both have different histories and different methods for implementation, and should be judged independently "

[trinibwoy]

havok is built to work on CPUs and is not overly saturated.

I really think you're confusing APIs with the use of those APIs. There's nothing inherent about Havok that makes it "not overly saturated". You can easily code Havok routines that bog down your CPU.

havok is not the same as phsyx at all. both have different histories and different methods for implementation, and should be judged independently

What does any of that has to do with your laundry list of things PhysX must do but other APIs don't need to, in order to get your approval? You are arbitrarily setting the bar higher for PhysX because that's the only way you can continue to criticize it while ignoring the fact that other APIs are in the same boat.

[Iconyu]

Nvidia didn't choose to use x87, they bought Novodex that way. They didn't go out of their way to do anything, they just didn't go out of their way to improve it. At least try to do some research guys before jumping on the Nvidia hating bandwagon

Any programmer would have at least looked at SSE recompile over x87, even at only a 8% gain the fact the only excuse is the guy we bought it from didn't still isn't enough.

You want to know why I think they won't do it? Because in a couple of years time once AMD and Intel get their 'fusion' act together CPU's will have a GPU (with a much lower latency, and direct use of system memory) on die that will make GPU physics very interesting.

Next year everyone will be a nvidia fanboy again, nvidia always come back from a ass kicking. But will we be talking Physx or will Intel be showing off a next gen Havok built to fusion type CPU's as another direct attack on nvidia?

[Manicdan]

^^i never set the bar for Havok, i ONLY discussed the 3 things that i think would let nvidia gain alot of respect from the public.

and your right that you can make a game that forces Havok to bog down a cpu, but you dont because you know that would be stupid since it cant be supported (unless your goal is to have an ultrahigh setting for future use). however there seems to be a clear hand pushing on developers using physx to do certain things to make the cpu bog. many newer physx games have been constantly identified as bloated. so i suggest that settings should be available to tweak each major piece of the physics (how dense smoke is, how many points are on cloth, should cloth animation be on all users or just main character). let people pick what they like, or let people feel immersed into the game.

my "laundry list" is just for physx because havok has completely different problems and some might be worse, or just as bad, but still different.

SUVs roll over easier
compact cars get completely crushed easily
should we crash test both identically because they are both automobiles? no, you try to identify the weakest things and fix them.

[Manicdan]

darn thing double posted

[trinibwoy]

Any programmer would have at least looked at SSE recompile over x87, even at only a 8% gain the fact the only excuse is the guy we bought it from didn't still isn't enough.

When there's some real analysis on the cost and effort required to improve PhysX performance on the CPU then let's talk. Pulling out a random 8% number out of thin air doesn't work in the real world. For argument's sake let's say there is a magical 8% performance increase and that PhysX is 100% of the CPU workload. So those 15fps numbers become 16fps...hoorah!!!

For the run of the mill stuff that Havok does, PhysX performs just fine. It's only when you try to run the more advanced GPU enabled stuff on the CPU that it struggles. But let's not let those simple facts get in the way of Nvidia bashing.

[RaV[666]]

Well, nvidia, if you cant program , open source it.I bet few folks would make it work well in a matter of months.
But wait, you wont do that because ?

Because you cant cripple non nvidia systems that way.got ya.

[Final8ty]

I really think you're confusing APIs with the use of those APIs. There's nothing inherent about Havok that makes it "not overly saturated". You can easily code Havok routines that bog down your CPU.

Indeed, but its funny that its nearly impossible for developers to write Physx routines that don't bog down the GPU.