Mirror's Edge PhysX Performance: PPU vs GPU vs CPU

[Final8y]

As we discussed in our Cryostasis PhysX Performance Preview article back in December, 2009 could finally be the year PhysX could begin to matter for PC gamers. With NVIDIA recently signing on game publishers EA, 2K Games, and THQ, plus PhysX being supported by the world’s number one and number two most popular gaming engines – Unreal Engine 3 and Gamebryo – the technology appears to finally be poised for liftoff.

Mirror’s Edge is the first 2009 title to support the technology. Mirror’s Edge also happens to be the first AAA title to support PhysX out-of-the-box: PhysX is an adjustable setting that can be toggled on and off from right within the game’s graphics menu, regardless if you have a GeForce GPU or not. Obviously if you don’t have GeForce 8 or better card installed, PhysX will run on your PCs CPU instead of the GPU. Or if you happen to have an AGEIA PhysX PPU card, you can run Mirror’s Edge on that as well.

http://www.firingsquad.com/hardware/mirrors_edge_physx_performance/
The PPU add in card does very well indeed & out does the NV260-216.

http://www.firingsquad.com/hardware/mirrors_edge_physx_performance/page6.asp

 

[Final8y]

crikey, so the standalone physics pci card beats the GPU physx?

ATM but they should have tested the the faster NV cards as well, and is understandable as the GPU is rendering the gfx as well & shows that there is no spare power on the GPUs for physx or they wouldn't of taken such a hit.

 

[Final8y]

an interesting read, cheers for the link

makes me tempted to boot up win7 and chuck a 8800/9600 to run physx alongside my 4870..

The benchmark results for a 4870 with cpu physx are rather dire, especially when you consider that other games with similar physics effects have nowhere near the processing requirements.

So.... why my cpu load is only 50%~ with crappy fps when there's stuff like flags and windows breaking, and then 35% cpu usage with 60+ fps with no physx stuff on screen ?

Why not use more cpu then ?

The CPU code in the original Ageia PhysX engine solver utilized legacy x87 instructions a lot while barely/not touching SSE at all (ironic, especially as the PPU is supposed to be a SIMD)

It was done so probably because of obvious reasons then- Ageia wanted to sell their card. nVidia likely did not optimize the CPU path for the same reasons.

http://www.xtremesystems.org/forums/showthread.php?t=216703

 

[Final8y]

If they used all the resources on the CPU, they could likely double non-PPU/GPU assisted scores, which in turn would mean an add-in card for PhysX would only add a mediocre boost in overall performance... Heheh, makes me think all this hardware assisted physics stuff is just a load of hot air.

I have had my suspicions of PhysX CPU utilization,optimization for quite some time.

 

[Final8y]

I dunno about that... even with a modified version of the tokamak source with optimization for intel core 2 instruction sets its still not hugely faster than physx for software physics.

Proof is in the pudding which they will not do because anything that improves PhysX on the CPU is a little less reason to own a physx card.
The wider you make the gap the better it looks.
NV are not in it to help the sales of games, they want to sell hardware and you don't put more effort than the bear minimum to function into making software run better on competing hardware .

 

[Final8y]

Yeah I'm just saying with another physics engine without the possible bias... even running less accurate simulations with full mutli threaded and instruction set optimization (not my work I'm just using the API) its not significantly faster than physx for software simulations. And its as like for like as possible as most calls are possible to translate between APIs with marginal difference to the code.

As been stated in posts on the other forum which i have linked to.

I mean seriously, swinging curtains and glass on the floor is a 2009 tech? Give me a break. Hitman 47 could do all that. Then there is awesome Red Faction, Splinter Cell, Max Payne 2. All physically superior to most of today's games.

And to aad BLACK had Glass incredible effects.

CryoS PhysX water effects did nothing for my & im more impressed with ATI's ToyShop random water generator http://www2.ati.com/misc/demos/ati-demo-toyshop-v1.2.exe Edit the screen res in the Sushi Configuration Settings.

 

[Final8y]

Splinter cell (quite a few years back) has pretty awesome cloth simulations (hanging curtains, shower curtains, blinds, etc.) but they were limited to 1 per current scene or performance took a massive nose dive, in the few scenes where there was 2-3 instances performance dropped by 2/3rds. With a GPU you could use that effect extensively throughout the scene.

I have no doubt that it would be possible to squeeze quite a bit more software performance out of physx with the right motivation but it would not be in any way comparable to high end GPU performance. A 260 at 756MHz v a Q6600 @ 3.6gig is between 20 and 60 times faster at doing physics depending on what kind of physics you have in the scene, even with some pretty heavy weight optimization to the software code for extensive effects its still going to be a minimum of 10x faster.

Comparable was not the point because if PhysX on the CPU was, then it would be the number 1 reason not to make it happen on the CPU.

It does not matter on how many scenes it is included on, if it can done then that's all that matters.

 

[Final8y]

lol I'm not too bothered either way (other than I hate game technology being held back because of "beauracry") at the end of the day people can carry on burying their heads in the sand but hardware accelerated physics are the future and aren't just going to go away (unless intel brings out some miracle CPU in the next 12-18 months).

Its not about THE hardware its about the PhysX which is not impressive enough for people to bother with adding hardware to run it.
These are games, not nuke simulations.

ATM the other software physics engines are quit capable of rending what has been seen in Physx games up to this point the level of accuracy of eyecandy physics is not as important as you make out to most of the worlds gamers.

 

[Final8y]

There is no software engine capable of doing what you see in cryostasis above single digit fps and most of the effects in mirrors edge while not running the most optimal of performance would tax a software implementation too serverely to be useful.

It does not matter what it takes to what cryostasis does as it still looks unimpressive to me. The ATI Toy Story give me a bigger wow factor when i first ran it.

I have seen just as good or better than mirrors edge for 90% of the effects besides the tearing of cloth, & the only thing obviously different is that on PhysX it takes allot more resources. I don't care nore does the majority if its real-time or pre calculated as long as it looks GOOD.
I see many games with cloth that looks real enough to me that even wrap around body parts.
As things stand now, the only reason that users want to run PhysX is because of more eyecandy on screen & not because of more accurate physics, because even the review says that its only really eyecandy & not the physics.
Even Painkiller looks better physically & that's all CPU.

And even if we don't take flashy physics effects into account 3D fps game engines are converging on a point where software physics can't run fast enough to simulate physics on simple bodies with a moderate degree of accuracy.

Thats your problem right there accuracy, the avg gamer would not know what accurate physics is, so why waste hardware & programming resources doing so when the gamer will never appreciate it.
Lots of simple convincing physics is enough as this is about games having fun or if you keep pushing for high accuracy you end up with a simulator which does not have a broad appeal across game types & becomes boring for many.

 

[Final8y]

Yeah one thing that made me chuckle about this physx stuff is something I saw with those weird 'plastic strip curtains' (don't know what the correct term is) which you can brush through and make them move around - I'm sure Splinter Cell (or possibly Pandora Tomorrow) managed the same thing 5+ years ago in some kind of hospital lab / mortuary or something. Perhaps not as 'realistic' but certainly good enough to be convincing when you are just playing a game and not trying to nitpick every tiny flaw in the engine.

Yep And is like a person in a Motorcar race still at the starting line to busy looking under the bonnet while the rest are having fun & nearly finished & is moaning that the cars are actually electric with a loud speaker for the motor sound.

 

[Final8y]

You'd have to be a game developer to really appreciate what I'm saying here but the difference in accuracy that I'm talking about is the difference between 2 systems, one the "old" way that just isn't good enough any more and the new method with proper interaction with the scene.

As an example guns in a game like quake 3 would fall straight to the floor - always at an exact perpendiuclar angle towards "straight down" and if the surface was a slope or uneven they wouldn't align themselves to it - in a newer game using a proper physics engine they would interact with the geometry to align themselves properly and that code takes a lot more processing however inaccurate you make the precision.

Which 99% of the gaming community are not developers so we will not be nit picking.

And the rest you described would not mater to a large amount of games. And gamers would not even be looking out for it in the first place in the types of games where that would have an influence unless it was pointed out & there is nothing worse then having to point it out for it to be noticed.

Realism wasn't the issue with those cloth effects (which were very nicely done in splinter cell - and it was quite an old version of splinter cell) the problem was you could only use a very limited number of those affects active in any one scene before performance plummeted from nice and smooth to single digit framerates. With hardware physics you can have say 20 of those effects running in a single scene and it will still be giving you smooth framerates.

Why would you need 20 of them in the first place, you would end up with a CellFactor with lots of just because we could, not because its needed to be convincing.

 

[Final8y]

Ill take to seconded one, because my mind will be on the task at hand & not how well the gun was resting on objects.

 

[Final8y]

Thats the problem... you don't understand how games are made so you aren't seeing the huge issues... which I can assure you are very real.

No! I play the games & not analytically study them & the point of the games maker is to make games for the games players & not to try and impress other games makers with inclusions that only gamers makers would notice & appreciate.

 

[Final8y]

I rest my case - if only it was that simple.

No it is not! The point is your trying to convince gamers what we need to have a good time & we are telling you that we don't need that sort of accuracy that hogs so much recourse as for what is in Mirrors Edge is a joke for what power it used visually.

Your case should have never been open on a non developer forum because you cant see it from a gamers perspective & to be a successful one, you have to see it from our view. I don't need to see a game from a developers view to buy & play it.

You cant tell us that something is good just because it uses a technology.
It has to be implemented in a way that we openly & obviously can see for ourselves.
If improvement in accurate physics has to be pointed out all the time to be noticed then it has failed for its intended audience, the gamer.


I'm all for good psychics, but that's not what I'm witnessing atm. all I'm seeing is big talk with lacklustre results.
You have to sell it to us by results, not talk.

 

[Final8y]

Exactly! Physx as it stands is a joke

 

[Final8y]

I dunno, our systems analysis and design tutor always told us the client doesn't know what they want! Or that might've been web design. Either way!

Its not the same.
As what were are being told is not what we are getting.