In memoriam: GPU acclerated PhysX

[80251]

Metro Last Light supports GPU accelerated PhysX (I think it's the only game I still have that does). Anyway, when running the game with CPU accelerated PhysX my 4090 was barely getting any GPU load at all, but when I set PhysX to GPU acceleration my 4090 gets pegged at 96%+ GPU load and a peak power draw of 464 Watts with the Metro Last Light benchmark. In the FPS graph though there's a lot of high frequency FPS variation (much more than with CPU accelerated PhysX) but it didn't seem to cause any noticeable spikes and/or lags in FPS (maybe because of Gsync?).

 

[Vayra86]

Heh, interesting find. And damn. Weird

 

[Space Lynx]

this is interesting, i was looking at physx stuff recently cause i am playing the older batman games "batman origins" to be specific and they both had Physx checkboxes in settings that I couldn't turn on since all AMD.

it got me wondering though, has physx been given any love at all on newer cards? or is it like the same physx chip on cards from a long time ago, as far as I am aware physx is dedicated chip on each gpu

 

[wolf]

it got me wondering though, has physx been given any love at all on newer cards? or is it like the same physx chip on cards from a long time ago, as far as I am aware physx is dedicated chip on each gpu

Without googling and just speaking from memory, my impression is that it used unified shaders/cuda cores to process the PhysX calculation, after of course the dedicated card from the purchased company Aegia. It seemed possible at the time to have one GPU do it and all game rendering, but I remember somewhere in NVCP where you could choose which card did the PhsyX, which led to some people keeping older cards in their system/running mismatched cards for dedicated PhsyX.

What OP's talking about though? that's a very interesting one indeed, I wonder if it can be replicated.

 

[AusWolf]

Without googling and just speaking from memory, my impression is that it used unified shaders/cuda cores to process the PhysX calculation

That's right. AMD cards can't run it mainly because they don't support cuda. I don't remember any GPU ever having a dedicated physics chip, to be honest.

 

[Ferrum Master]

Old news. Metro is just stupidly implemented stutter fest if physx are on, there is only one other game having the same.

The fix was to dedicate PhysX to a different GPU. Try using your main GPU and dedicate even a 750ti to do physx only. It will fix all stutter. Is it driver or architecture limitation, I have no idea. But doing physx on same GPU caused unbearable spikes and that's a FPS so not preferable.

 

[80251]

Metro Exodus, even thought it supports PhysX, doesn't support GPU accelerated PhysX, even if you select it in NVCP. I have a few other PhysX titles (incl. Deus Ex: Mankind Divided, Terminator:Resistance), but none of them support GPU accelerated PhysX either. Maybe CPU's are so powerful now GPU accelerated PhysX is just irrelevant.

 

[DemonicRyzen666]

If I remember correctly, it's just supposed to be general purpose, but on some older Cinebench it shows cards that have physX support to have SSE2 for some reason.

Metro Exodus, even thought it supports PhysX, doesn't support GPU accelerated PhysX, even if you select it in NVCP. I have a few other PhysX titles (incl. Deus Ex: Mankind Divided, Terminator:Resistance), but none of them support GPU accelerated PhysX either. Maybe CPU's are so powerful now GPU accelerated PhysX is just irrelevant.

I don't think so
you can benchmark your RTX 4090 & cpu with Fluidmark benchmark.

I test mine with custom in the options part settings with more GPu load checked, MultithreadPhysX checked, Async mode checked # of emitters set to 14 The unlock box checked, & particle count to 256,000. Then hit the Run users benchmark button above the 1080P & 720P buttons

My score barely changed when going from a 5600x to 5800x 3D when doing cpu PhysX sps 16 to sps 18 MSAA was set to 0x for that seeting it back to 16x makes them exactly the same.
However, since I have two RTX 2080 ti's. I have tested with one dedicated to physX I get 68sps with either MSAA setting gpu load never got above 40% for the gpu physX in this benchmark.

 

[Ferrum Master]

If I remember correctly, it's just supposed to be general purpose, but on some older Cinebench it shows cards that have physX support to have SSE2 for some reason.

I don't think so
you can benchmark your RTX 4090 & cpu with Fluidmark benchmark.

I test mine with custom in the options part settings with more GPu load checked, MultithreadPhysX checked, Async mode checked # of emitters set to 14 The unlock box checked, & particle count to 256,000. Then hit the Run users benchmark button above the 1080P & 720P buttons

My score barely changed when going from a 5600x to 5800x 3D when doing cpu PhysX sps 16 to sps 18 MSAA was set to 0x for that seeting it back to 16x makes them exactly the same.
However, since I have two RTX 2080 ti's. I have tested with one dedicated to physX I get 68sps with either MSAA setting gpu load never got above 40% for the gpu physX in this benchmark.

Oh can you do fast metro bench with the in build graphs with dedicated PhysX on your second GPU.

 

[80251]

If I remember correctly, it's just supposed to be general purpose, but on some older Cinebench it shows cards that have physX support to have SSE2 for some reason.

I don't think so
you can benchmark your RTX 4090 & cpu with Fluidmark benchmark.

I test mine with custom in the options part settings with more GPu load checked, MultithreadPhysX checked, Async mode checked # of emitters set to 14 The unlock box checked, & particle count to 256,000. Then hit the Run users benchmark button above the 1080P & 720P buttons

My score barely changed when going from a 5600x to 5800x 3D when doing cpu PhysX sps 16 to sps 18 MSAA was set to 0x for that seeting it back to 16x makes them exactly the same.
However, since I have two RTX 2080 ti's. I have tested with one dedicated to physX I get 68sps with either MSAA setting gpu load never got above 40% for the gpu physX in this benchmark.

I don't think games make use of PhysX to the degree your Fluidmark benchmark does. I believe Metro:Last Light was the title that used the most PhysX effects ever and even then it was only for fog, smoke, particle effects (debris from gunfire and/or explosions) and cloth effects.

I remember ten years ago I read someone was running Metro:Last Light w/Titans (I guess they must've been kepler Titans). Anyway, he found out the game ran better w/one of the Titans dedicated to PhysX rather than SLI.

I found with Metro:Last Light that using a dedicated GPU (a 5820 using a dedicated GTX 780 for PhysX with a 980ti, and then the 980ti as a dedicated PhysX card w/a 1080ti) there was more framerate variation in the graph when using a dedicated PhysX GPU, but not any significant difference in performance. The maximum load I saw on the GTX 780 was 33%.

As other people have noted, maybe these days it would be better to have a dedicated GPU RTX accelerator.

 

[wolf]

Mucking about with my GTX 590, I figured I'd revisit the old screens in NVCP, indeed it does let you choose the GPU you want to accelerate PhsyX with. Haven't fired up my main rig in a couple days, not sure what the modern UI shows.

I think I'll install a period specific game, perhaps bulletstorm and test that out, iirc it has physx too.

 

[Shou Miko]

Well if PhysX had taken off the need for dual graphics would still be useful like a powerful gaming gpu and a lower/mid end gpu for PhysX.

I tried it back in the day with 2x8800GT 512mb cards and my fps in games was actually better running one for games and one for physx not in sli.

 

[DemonicRyzen666]

Oh can you do fast metro bench with the in build graphs with dedicated PhysX on your second GPU.

I only have the redux version & not the original as I have it from G.O.G on the Franchise bundle. Plus I have to download it. I'll give it a go.

As other people have noted, maybe these days it would be better to have a dedicated GPU RTX accelerator.

you probably won't see that as the people who complained about microstutter whined enough to kill even new mGPU support for games on DX12, which when used in Aysnc was way better than A.F.R. type multi-gpu.

The cards we have now are basically a half way step. we need to move to all raytracing but you won't be able to play older games as they won't have any type of rasterization & only do pure raytracing. In all honesty that is the only way raytracing will be vaiable at all.