FS has really in depth PhysX Article

[geoffry]

http://firingsquad.com/hardwar...sx_performance_update/

Enjoy

 

[geoffry]

Its interesting how the GPU can spank the PPU even as it is working on graphics as well as PhysX calculations.

Does anyone know how many cores or SP the PPU has?

 

[Denithor]

I wonder why their 4870 underperformed their 4850?

Very interesting results though. Looks like the PPU add-in card is hitting a throughput limit and the overflow work is off-loaded to the CPU instead of to the GPU (based on the fact that the GPU alone performs better than the GPU + PPU).

Now we just need to see a few 'must have' games developed that feature PhysX.

Also have to wonder if/when ATi cards will run PhysX (officially, that is).

 

[geoffry]

Ya, I'm curious if the mostly unused shaders in the ATI architecture are useful for PhysX, if thats the case then they wouldn't take much of a performance hit.

 

[SunnyD]

Originally posted by: Denithor
Very interesting results though. Looks like the PPU add-in card is hitting a throughput limit and the overflow work is off-loaded to the CPU instead of to the GPU (based on the fact that the GPU alone performs better than the GPU + PPU).

More likely the following: The CPU has to prefetch all of the appropriate data from memory, pump it over the PCI bus to the PPU, which then has to push it back over the PCI bus to the system memory pool, which the CPU then has to fetch and use in the rendering path, then finally pushed to the display hardware for rendering.

NVIDIA PhysX more than likely takes the PPU data path and drops it right into the display driver path, meaning that physics "rendering" has 2 less hops across a much slower bus.

 

[apoppin]

Ultimately though we?re still waiting for more engaging PhysX content to be released. PhysX is a nice little bonus worth checking out if you already own a GeForce 8 or better GPU with 256MB of graphics memory, but as it stands now it isn?t compelling enough to make Radeon 4800 card owners jealous in our opinion.

that is the real problem for Nvidia .. not PhysX itself .. *selling it* to the devs while MS/intel and AMD are pushing Havok

 

[Fox5]

Originally posted by: apoppin

Ultimately though we?re still waiting for more engaging PhysX content to be released. PhysX is a nice little bonus worth checking out if you already own a GeForce 8 or better GPU with 256MB of graphics memory, but as it stands now it isn?t compelling enough to make Radeon 4800 card owners jealous in our opinion.

that is the real problem for Nvidia .. not PhysX itself .. *selling it* to the devs while MS/intel and AMD are pushing Havok

Well, they do have one real win with PhysX...Unreal Engine 3 and the games that use it. But honestly, they NEED AMD support on this if they're going to get anywhere.

 

[Elcs]

In my opinion:

The only way this Physics idea will take a strong hold in the gaming market is if the big players work together on the issue to create a standard framework for them to base off.

The move would benefit both AMD/ATI and Nvidia. As Nvidia are seeing, people are buying lesser end 8xxx cards to give themselves a taste of PhysX

 

[Nemesis 1]

Good read . I am a little lazy today. Were can I buy one of these PPU's I would very much like to test it with Nehalem .

 

[BFG10K]

Originally posted by: geoffry

Its interesting how the GPU can spank the PPU even as it is working on graphics as well as PhysX calculations.

http://www.fudzilla.com/index....=view&id=8859&Itemid=1
http://www.fudzilla.com/index....=view&id=8862&Itemid=1

 

[Nemesis 1]

Way to go BFG10K . I been watching this thread. Ya killed it with 2 linlks to the Fud master. Is there more on this. Guys been testing these cards . Why no mention of CPU usage by the guys who were testing this new tech.

 

[ShadowOfMyself]

Interesting links... Usually Id just pass them off as fud but since they have actual numbers to support it, now Im curious

 

[Keysplayr]

Hmmm. Interesting on the CPU usage. Does anyone have any ideas on how this matters?

 

[Nemesis 1]

YA!

 

[Keysplayr]

Well Zinn, what is it?

 

[Nemesis 1]

OK Keys . Since you referr to who I preferr to be called .

We have known each other along time. I have known you to be pretty decent guy. I admired the way you stood up for Rollo and AEG thing . As I felt the same way .

I also know from all those years of reading your post . That your Vary intelligent guy.

There are very few long term members here. That doesn't believe for one second that you can't ans the very question you posed.

So have a go at it . We know you can.

 

[nitromullet]

Originally posted by: ShadowOfMyself
Interesting links... Usually Id just pass them off as fud but since they have actual numbers to support it, now Im curious

I set out to reproduce this myself, and was able to do so. But... I also sort of discovered something else somewhat interesting along the way...

Results with four instances of CPU Burn In running quad core cpu at 100%:

PhysX FluidMark v1.0.0
[ SCORE: 1245 o3Marks ] - Time: 60000 ms
Res: 1920x1200 (FS) - MSAA: 4X
FPS: min=11 max=22 avg=21

Without CPU Burn in Running:

PhysX FluidMark v1.0.0
[ SCORE: 8161 o3Marks ] - Time: 60000 ms
Res: 1920x1200 (FS) - MSAA: 4X
FPS: min=93 max=342 avg=136

Without CPU Burn in running, looking at the CPU Usage History graph in Windows show core 0 running at 70-80% load during the FluidMark run.

Results with three instances of CPU Burn In running quad core cpu at ~77% (affinity 0,1,2):

PhysX FluidMark v1.0.0
[ SCORE: 1384 o3Marks ] - Time: 60000 ms
Res: 1920x1200 (FS) - MSAA: 4X
FPS: min=11 max=29 avg=23

Interestingly enough, core 3 doesn't show a significant increase in usage. Cores 0-2 were running at 100%, and core 3 was staying at ~10% during the entire benchmark, with a quick spike to 70%. So, I decided to try setting the affinity for CPU Burn In differently...

Results with three instances of CPU Burn In running quad core cpu at ~77% (affinity 1,2,3):

PhysX FluidMark v1.0.0
[ SCORE: 8124 o3Marks ] - Time: 60000 ms
Res: 1920x1200 (FS) - MSAA: 4X
FPS: min=102 max=343 avg=136

Hmmmm... interesting. This time core 0 was at 75-80% during the benchmark and cores 1-3 were all at 100%

Not sure what exactly to make of this, except that it looks like the benchmark might its affinity set to core 0.

 

[Nemesis 1]

Thats interesting. THanks. So does this mean than that a 3 or 4 core cpu is required so as not to bottleneck?

 

[Keysplayr]

Originally posted by: Nemesis 1
OK Keys . Since you referr to who I preferr to be called .

We have known each other along time. I have known you to be pretty decent guy. I admired the way you stood up for Rollo and AEG thing . As I felt the same way .

I also know from all those years of reading your post . That your Vary intelligent guy.

There are very few long term members here. That doesn't believe for one second that you can't ans the very question you posed.

So have a go at it . We know you can.

Nevermind, I'll ask anyone else the same question who does not go off on wild tangents.
Anyone?

 

[Keysplayr]

Originally posted by: Nemesis 1
Thats interesting. THanks. So does this mean than that a 3 or 4 core cpu is required so as not to bottleneck?

It means, as he said, that he doesn't know what to make of it. I've run benches on Dual Core as well as Quad. Didn't really pay attention to the CPU usage, cause I was benching or gaming.
I was not also trying to decode DVD's while I played. So I can't tell you.

 

[nitromullet]

Originally posted by: Nemesis 1
Thats interesting. THanks. So does this mean than that a 3 or 4 core cpu is required so as not to bottleneck?

No I think it means, that the benchmark will run better on a multi-core (2 or more) system than a single core.

I also checked the affinity on the the benchmark, and sure enough, its affinity was set to cpu 0. Allowing the benchmark to actually use all four cores by setting its affinity to 0,1,2,3 increased its performance slightly...

PhysX FluidMark v1.0.0
[ SCORE: 8261 o3Marks ] - Time: 60000 ms
Res: 1920x1200 (FS) - MSAA: 4X
FPS: min=95 max=341 avg=138

So, I think what this means is that the benchmark does have some CPU overhead, which may be increased by NVIDIA's implementation of PhysX. The important thing to note is that I had to artificially force all four cores to run at 100% to slow the benchmark down.

I also I don't think that NVIDIA has developed a method of offloading physics calculations to the CPU that apparently outperforms by 3x (or more) software implementations that do the same thing.

If they had, this seems to me that it would be an even greater accomplishment than running PhysX on the gpu. I'm pretty sure they could license that technology for a little bit of cash a lot easier then trying to use PhysX as a selling point for their video cards.

 

[BFG10K]

Originally posted by: Nemesis 1

Way to go BFG10K . I been watching this thread. Ya killed it with 2 linlks to the Fud master. Is there more on this. Guys been testing these cards . Why no mention of CPU usage by the guys who were testing this new tech.

Well I don't know if it's true or not, but given it's posted by Lars and not the usual "fraud clones" (Lars used to write for Tom's Hardware) I thought it might have a bit more credibility.

What I found most interesting was that he was reporting much less CPU usage with the PPU (again, if he is to be believed).