Physx Q&A

T

TechonNapkins

Weaksauce
Joined
Oct 20, 2010
Messages
94
I don't have all the answers yet. I'll probably get a GT240 board soon to try it, but if anyone has prior experience, please share.

(1) What is the recommended amount of memory on a dedicated Physx card (ie, when does adding more RAM net no further benefit?)

(2) What is generally the bottleneck on Physx performance: core, shader, or memory frequency?

Shader count and clock are the biggest performance determinants, followed by core clock. Memory bandwidth seems to make little to no impact. While 32 shaders is the minimum recommended by Nvidia, 48 is the effective minimum (GT220), with benefits tapering off rather quickly after normally clocked 96 shaders (GT240). That being said, some situations still seem to benefit significantly from overclocking 128 shader cards (GTS250/9800GTX/8800GTS). Clear as mud yet?

(3) Does a PCIe 1x slot have enough bandwidth to handle Physx properly?

(4) Does a PCIe 1x slot have enough power to drive a mainstream card, say, like, a GT240?

(5) If no to above, what about 16x physical slots that run at 1x data rates (do the extra pins receive power, just not signal?)
 
Last edited:
What is your main video card? For anything newer a GT240 is going to bottleneck it and bring the FPS down quite a bit.
 
cyclone3d said:
What is your main video card? For anything newer a GT240 is going to bottleneck it and bring the FPS down quite a bit.
Click to expand...

My current card is a 6950. I will be using the hacked drivers to get it to work with the gt240.

Here's a Tom's article from awhile back showing scaling of dedicated physx cards:

http://www.tomshardware.com/reviews/batman-arkham-asylum,2465-8.html

It seems after 64 shaders (9600gt) the benefits seem to be minimal. Sure, the results are for Arkham Asylum, which is getting long in the tooth, so I looked around for some Mafia II tests...

http://www.xtremesystems.org/forums/showthread.php?t=257242

http://www.nvnews.net/vbulletin/showthread.php?t=154038

It seems again benefits are minimal after the 9600gt. A gt240 with 96 shaders seems to be the best balance (and it's the most powerful 200-series that doesn't need a 6 pin). Furthermore, I want to test the affects of clock speed on performance. If I throw, say, 384 shaders at it, I won't be able to really determine the shader's clockspeed's impact.

If we find out that core clock, mem clock, mem type, mem amount, and what-not, are all relatively unimportant compared to shader count and clock, then Nvidia could take all their spare gts250 inventory and turn them into cheap dedicated Physx cards.
 
My ideal dedicated PhysX card:
- Equivalent PhysX performance of a 9800GT (or better).
- Low power, with no auxiliary power connectors.
- Single slot, passive cooler.
- Low-profile PCB.
- PCIe 1x connector (so it fits in any spare slot you happen to have).

Now, if only Nvidia would build something like that :p
 
TechonNapkins said:
Hmmm... I could swear I hit submit earlier today, but I don't see the post... anyway, I found out someone already did the research for me. Shader count/clock makes the biggest difference, followed by core clock, whereas memory speed ends up accomplishing bupkiss.

Here's the thread:

http://www.xtremesystems.org/forums/showthread.php?t=263109

And here's a direct link to the chart:

http://img522.imageshack.us/img522/6941/tablel.jpg
Click to expand...

His minimum frame-rates didn't change much at all; making the PhysX card faster only resulted in higher peaks (worthless) and inflated averages (even more worthless).

Looks like he's bottle-necked somewhere else.

I wonder how slow a PhysX card can be before it actually starts to become detrimental to real-world gameplay. By that I mean a card slow enough to lower the minimum frame-rate substantially, or a card slow enough to bring the maximum frame-rate below 60FPS
 
Unknown-One said:
His minimum frame-rates didn't change much at all; making the PhysX card faster only resulted in higher peaks (worthless) and inflated averages (even more worthless).

Looks like he's bottle-necked somewhere else.

I wonder how slow a PhysX card can be before it actually starts to become detrimental to real-world gameplay. By that I mean a card slow enough to lower the minimum frame-rate substantially, or a card slow enough to bring the maximum frame-rate below 60FPS
Click to expand...

Well, I just tried Mafia II on my 6950 with physics set to medium (just using the CPU: 2500K oc'd to 4.45), and it didn't have too many dips below 30fps, so I'm thinking long and hard about this... I would like to know if a GT240 helps at medium (brings those minimums up to over 40 or more), because, as you mentioned, at those single digit minimums on high, who cares.
 
In Mafia 2 you can edit your 'PhysX' settings manually. By deleting parts of the content (character clothes, burnouts, explosions, fire, etc).

I only have expolosions/fire/burnout smoke in the directory and my fps had almost trippled.

Tbh buying a weak card just for PhysX is not worth it, I'm sure you'l get the same gain if use your CPU then to ur dedicated PhysX card. I ran a GTX 285 & 8800GTS (PhysX) and it showed the same performance as when I switched it to CPU (QX9650 @ 4.2)
 
Linus from Tech Tips tested a few scenarios. He started with a GTX 580 and tested it processing PhysX by itself, then he added a few cards and tested the results. Even when you add a second 580 for PhysX, the FPS only goes up by about 8. 8600GTS as PhysX actually killed the FPS.

http://www.youtube.com/watch?v=cbww3dhzK0M&feature=player_embedded

Nvidia has their marketing all screwed up. Their own cards slow down their products. I think this shows that the current state of PhysX and it's marketing will eventually die off and the product will go away unless Nvidia makes some serious changes and starts being more open and honest.
 
Last edited:
The FPS drops, but does the game look better? I don't know, I never bothered to use PhysX, let alone even knew it was still around.
 
Well, I just finished Mafia II yesterday. Ran medium physx on the CPU the whole time, and never felt the game chug to a halt.
 
This site here...

http://physxinfo.com/news/3628/mafia-ii-demo-tweaking-physx-performance/

... explains the real performance killer in Mafia II: clothes effects.

Apparently most of the clothing physics is done on the CPU, regardless of a dedicated Physx card... which would explain the not-too-impressive benefits people are getting from such a setup.

Medium/High CPU physx = Medium/High GPU physx in terms of effects I believe . And it seems by disabling the clothing effects (on all the NPCs... you can leave it on for Vito), you can even run it on High without GPU physx acceleration.

This is not looking too good for justifying Physx. I'm wondering if I should even waste my time/money on this experiment now.
 
physX is a failed gimmick by nvidia. the multicore CPUs now day can handle the phyX on their own.
 
I don't think there are any games in development right now that are accounced to have hardware PhysX. Maybe Arkham City?
dukerjames said:
physX is a failed gimmick by nvidia.
Click to expand...
+1
the multicore CPUs now day can handle the phyX on their own.
Click to expand...
There was an article up a couple years back showing that the CPU version of PhysX is some old coding language, like developed in the Windows 95 era. Apparently it's highly inefficient on CPU's but can run good on GPU's. Nvidia won't change it. They said the only logical reason that Nvidia won't change it is because they know that modern CPU's could run PhysX by themselves if coded properly. Nvidia doesn't want to have any of that.
 
dukerjames said:
physX is a failed gimmick by nvidia. the multicore CPUs now day can handle the phyX on their own.
Click to expand...
you have no idea what you are talking about. in many games hardware physx will bring any cpu down to pretty low framerates. now Sandy Bridge for some reason has made a massive improvement and can handle it as well as a decent mid range card but no other cpu comes close.

I do agree its pretty much a gimmick though.
 
bigdogchris said:
There was an article up a couple years back showing that the CPU version of PhysX is some old coding language, like developed in the Windows 95 era. Apparently it's highly inefficient on CPU's but can run good on GPU's. Nvidia won't change it. They said the only logical reason that Nvidia won't change it is because they know that modern CPU's could run PhysX by themselves if coded properly. Nvidia doesn't want to have any of that.
Click to expand...

The PhysX code actually dates back to about 10 years ago ('02-'03), before Ageia acquired NovodeX. http://hothardware.com/Reviews/NVIDIA-Sheds-Light-On-Lack-Of-PhysX-CPU-Optimizations/ At this time, SSE and multi-core and such weren't ubiquitous like they are now, so PhysX simply wasn't written for all the extra features that modern CPUs have (because the majority of CPUs at the time didn't have them). Nvidia says that it just hasn't been redone yet (easy to believe if you've ever worked in a company of any size), but that it is being rewritten for version 3 of the SDK. You may want to take that with a grain of salt until you actually see it, but that's what they're saying.

There are also a few other things to take into consideration. All the top developers have access to the PhysX source code and choose to compile it for x87 rather than SSE. Developers can choose to support CPU multi-threading in their PhysX implementation. While it's pretty clear that Nvidia hasn't done much to improve CPU PhysX, many of the things people complain about are controlled by the developer, and are not Nvidia limitations.

There are many PhysX functions that can't run on the GPU, so crippling the CPU implementation just to spite Intel and AMD probably wouldn't be in their best interest. And there are certain calculations that are simply handled better by the massively parallel processing of a GPU than by a few cores of a CPU (just due to the architectural differences of the two). I've run a few different types of GPU clients for distributed computing, and my AMD video cards are consistently faster by orders of magnitude. With Bitcoin, my i7-920 overclocked to 3.5GHz puts out about 5,000 khash/sec. My HD5870 does about 340,000. The HD4350 in my other box does about 7,000. distributed.net's RC5 client shows numbers that are orders of magnitude higher as well (even though the CPU client has been around for almost 15 years, while the GPU clients are only a couple years old).

While a CPU can do much more specific calculations, the thousands of shaders on a GPU may do a much better job of applying one simple physics-related mathematical formula to a million different particles simultaneously. I'm not saying that CPU PhysX couldn't handle the physics of a current game, but as we get more and more realistic, there will be more and more physics calculations applied to more and more objects, so it's possible a massively parallel GPU could very well spank even the best CPU available.


As for it being a gimmick, yeah, it is somewhat. Nvidia seems to want to use PhysX to force people to buy only Nvidia GPUs. However, devs want their game to work with the widest selection of systems, so they're not willing to completely exclude ATI cards. Because of this, PhysX is used for a little extra fluff, rather than using it in the core of the game. Since PhysX adds only a little eye candy to games, consumers don't have a whole lot of reason to rush out and buy PhysX cards. So the devs continue to build games that aren't completely based on GPU PhysX, and people continue to not buy PhysX cards, ad infinitum. I think it's more an economic chicken & egg situation leading to tacked-on PhysX features rather than PhysX itself being a gimmick, but the fact remains that GPU PhysX doesn't currently do a whole lot.
 
InvisiBill said:
As for it being a gimmick, yeah, it is somewhat. Nvidia seems to want to use PhysX to force people to buy only Nvidia GPUs. However, devs want their game to work with the widest selection of systems, so they're not willing to completely exclude ATI cards. Because of this, PhysX is used for a little extra fluff, rather than using it in the core of the game. Since PhysX adds only a little eye candy to games, consumers don't have a whole lot of reason to rush out and buy PhysX cards. So the devs continue to build games that aren't completely based on GPU PhysX, and people continue to not buy PhysX cards, ad infinitum. I think it's more an economic chicken & egg situation leading to tacked-on PhysX features rather than PhysX itself being a gimmick, but the fact remains that GPU PhysX doesn't currently do a whole lot.
Click to expand...

This.

That's what bothers me with Physx games. I find them LESS interactive than ones that shun Physx and use their own engine. Physx can't be part of the gameplay too much, or you would exclude all the non-nvidia audience. The Red Faction series is the ultimate counter example: even from the first game years ago, they had a more interactive world than even the latest Physx game. Take Mafia II for example: Want to ram your car through that store window? Too bad... but if you follow the linear path you're supposed to, you'll encounter windows you CAN break, and it'll be prettier! wow...
 
Keep your physx card a Tier just below your main card. If you got a 570 then a 460 will give a lot of improvement. For the 460 a 430 will give a good improvement as well. When you see a 5fps increase its not worth it all that heat and that power for such a small increase which also means the physx card a bit on the slow side.
 
Spooony said:
Keep your physx card a Tier just below your main card. If you got a 570 then a 460 will give a lot of improvement. For the 460 a 430 will give a good improvement as well. When you see a 5fps increase its not worth it all that heat and that power for such a small increase which also means the physx card a bit on the slow side.
Click to expand...
um a 430 is way way below a 460.
 
Can I piggyback on this thread? I have a new machine and two spare nVidia cards lying about, a 512MB 8600GT and a 1GB 9800GT, but I know next to nothing about multi-card setups and PhysX. I came snooping around here thinking to find out whether the 8600 would suffice or if I should really go for the 9800, but reading some of what you guys are saying and about how too slow a card might actually backfire and slow my system down I'm starting to wonder whether I should do it at all. There's more riding on that decision on some space considerations, too, I'll get into that in a bit.

Right, my system's running on an Asus P8Z68-V with a GTX570. I wanted to go with the 8600 since I figured it's smaller, runs cooler and drains less power. Also, it would make my old system more valuable on resale if I left the 9800 in there, so there's that too. But would that actually result in no benefit, or even work against my performance? With current games, or future ones? And if so, what about the 9800, if I used that instead would it be a sure-fire way of increasing performance in games using PhysX? Both cards are really a good deal older than the main graphics processor, so it's rather concerning.

See, I gotta admit I don't really know that much about how the PCIe lanes work. The 570 is in the standard x16 slot, about which the manual says can be used in x8 mode with dual setups. The second slot on the mainboard is listed as an x8. Now, does that mean that the simple act of plugging in any second graphics card, regardless of what it's to be used for, is going to right off the bat half the bandwidth available for my 570?

Now, there's also a third PCIe 2.0 slot available, labelled as x4 mode and supposed to be able to take an x4 or x1 device. Could I use that slot instead? And would that have any impact on the performance I'd be getting out of the main 570? Though I'm pretty sure I couldn't fit the 9800 in there anyway, the cooler would be right up against the PSU if it fit in at all, and even the 8600 would be hard pressed for breathing space.

I also mentioned space earlier, and that's to do with the onboard audio for P8Z68-V being crap - so I want to get an audio card. Here's the catch, if I use that second PCIe slot, the x8, the 9800 will definitely obstruct the second regular PCI slot (the 570 has the first one covered in that department), and I'm pretty sure that even the 8600 will do the same. In which case I need to buy a more expensive audio card for the PCIe x1 slot instead of the PCI one. Basically, if I end up putting in a second video card for PhysX acceleration I want to be sure I'm getting some benefit from it, otherwise I'd rather just save some money on the audio.

So yeah, that's about where I am. Can anyone offer some advice?
 
FelixC said:
Can I piggyback on this thread? I have a new machine and two spare nVidia cards lying about, a 512MB 8600GT and a 1GB 9800GT, but I know next to nothing about multi-card setups and PhysX. I came snooping around here thinking to find out whether the 8600 would suffice or if I should really go for the 9800, but reading some of what you guys are saying and about how too slow a card might actually backfire and slow my system down I'm starting to wonder whether I should do it at all. There's more riding on that decision on some space considerations, too, I'll get into that in a bit.

Right, my system's running on an Asus P8Z68-V with a GTX570. I wanted to go with the 8600 since I figured it's smaller, runs cooler and drains less power. Also, it would make my old system more valuable on resale if I left the 9800 in there, so there's that too. But would that actually result in no benefit, or even work against my performance? With current games, or future ones? And if so, what about the 9800, if I used that instead would it be a sure-fire way of increasing performance in games using PhysX? Both cards are really a good deal older than the main graphics processor, so it's rather concerning.

See, I gotta admit I don't really know that much about how the PCIe lanes work. The 570 is in the standard x16 slot, about which the manual says can be used in x8 mode with dual setups. The second slot on the mainboard is listed as an x8. Now, does that mean that the simple act of plugging in any second graphics card, regardless of what it's to be used for, is going to right off the bat half the bandwidth available for my 570?

Now, there's also a third PCIe 2.0 slot available, labelled as x4 mode and supposed to be able to take an x4 or x1 device. Could I use that slot instead? And would that have any impact on the performance I'd be getting out of the main 570? Though I'm pretty sure I couldn't fit the 9800 in there anyway, the cooler would be right up against the PSU if it fit in at all, and even the 8600 would be hard pressed for breathing space.

I also mentioned space earlier, and that's to do with the onboard audio for P8Z68-V being crap - so I want to get an audio card. Here's the catch, if I use that second PCIe slot, the x8, the 9800 will definitely obstruct the second regular PCI slot (the 570 has the first one covered in that department), and I'm pretty sure that even the 8600 will do the same. In which case I need to buy a more expensive audio card for the PCIe x1 slot instead of the PCI one. Basically, if I end up putting in a second video card for PhysX acceleration I want to be sure I'm getting some benefit from it, otherwise I'd rather just save some money on the audio.

So yeah, that's about where I am. Can anyone offer some advice?
Click to expand...

8600gt would be worthless; 9800gt you would see some benefit, but would probably be marginal with a 570.

if you don't have any gpu physx games, it would not benefit you whatsoever - see below:

http://physxinfo.com/

would use the x4 over the x8 slot.

wouldn't hurt to try if you have a compatible game to see how performance fairs, but i don't think it's worth the trouble since a 570 is more than adequate on its own, imo.

if you don't have the space then don't bother.

you can just leave the x8 and the slot below free in case you ever decide to sli or xfire.

pcie usually will not cost more than pci for audio cards if they are similar.

in conclusion, i would just buy an audio card; if you want pci for cheap, see the following:

http://www.amazon.com/ASUS-Sound-Ca...HJSS/ref=sr_1_1?ie=UTF8&qid=1309653200&sr=8-1
 
As an Amazon Associate, HardForum may earn from qualifying purchases.
Thanks for the advice, mate, that helps. I knew it's only certain games that are purpose-built to take advantage of PhysX acceleration, I just figured that since I had the cards on hand I might as well make use of one. But if even the 9800 would only bring about a minor benefit then it's probably not worth the hassle. I've got Mirror's Edge and Metro, the latter at least is bound to be a good benchmark, so I'll give it a shot when I get the time this upcoming week and see how it goes.

On the audio card, yeah, I had me eyes on something similar. Wanted to go for the Xonar DS if going PCI or the DX if I had to go PCIe. I'll probably end up picking up the DX to keep my second PCIe slot accessible even if I don't use a second graphics card now, will see.
 
You must log in or register to reply here.
Back
Top