new Fermi 480GTX... plus 8800GT physx

Aumakua

Limp Gawd
Joined
Dec 14, 2009
Messages
333
I know they aren't out yet, but this summer I was thinking of upgrading my SLI 8800GT configuration to a 480GTX (or whatever the name was established as). I will only buy one card because of how powerful they should be and the cost.

My question is, should I keep one of my 8800GT cards for physx?
i probably can't get much for them when selling and figured it might be worth it, but will it slow the new cards down? I have read that 8400 and sometimes 8600 cards slow down 295GTX cards when used for physx.
 
My guess is that the 8800GT will not slow down the 480, and if it does, slowdown will be minimal. However the 480GTX isn't out so we don't really know.
 
Depends on what settings you're going to run. If you're not really going to stress your graphics card (low resolutions like 1680x1050), then I'm almost positive that using the Fermi for both graphics and PhysX will be faster than using an 8800GT as a dedicated PhysX card.
 
I wouldn't doubt that a GTX480 + 8800GT PhysX is slower than a GTX480 just running the Physics for you.
 
Don't listen to chris, what he says doesn't make any sense. A GTX 480 + 8800GT will be faster than a 480 by itself. Logically it's the only thing that makes sense. Performance degradation happens at like the Geforce 8200 or whatever the 32SP part is. The 8800GT is a pretty powerful chip.
 
Not trying to start anything here but logic leads me to side with akotlar. Doesnt make any sense that offloading the PhysX to a capable gpu wouldn't help. At the very least it shouldn't slow you down.

The performance gains of using a 8800gtx, if any, may not be worth the extra power it will draw.

Speculation like this, without cards in slots is merely a way to pass time on the internet.
 
8800GT is good for physx, you should try it out when the time comes and make your own tests.
 
I have asked a similar question. I asked if having my 2x GTX280 in sli and a 8800gt for physx is worth it. I already have the parts but if it wasnt worth it then I was going to use the 8800gt as a sard in a secondary rig. Most people have said and logic usually dictates that more is better. Onlything I was wondering is that if the higher end card can handle physx and you throw in another card to help, does the second card(8800gt) help or will it take on all physx responsability itself and hurt? I hope this makes sense.
 
Don't listen to chris, what he says doesn't make any sense. A GTX 480 + 8800GT will be faster than a 480 by itself. Logically it's the only thing that makes sense. Performance degradation happens at like the Geforce 8200 or whatever the 32SP part is. The 8800GT is a pretty powerful chip.

It does make sense in the way he is thinking though. GTX 480=512SP correct? Now just take the 112 SP's away from the GTX480 for Physx(Really, it will probably call for 128SP). The card still has 400SP's left for gaming. And as we all know, no game coming in the next year is really going to push that amount of power on the GTX480. Nothing that can't be run on a 240 SP GTX285 anyway.(Triple monitor gaming excluded)

So basically, the GTX480 can do it all at the same time, where as in his mind there is a slight delay offloading to the 8800GT here lies the "faster on one card" than offloading it to a seperate card. It actually makes sense if you really look at it, and technically I would think there is actually a ms or 2 in that delay to offload.;)
 
Screw amount of SP's surely? Physx is pure crunching for particles etc there's no need for mass rendering of textures etc. an 8800GT would do fine in any current physx enabled game as a dedicated physx cruncher.
 
The reason crappy cards suck for physx is because they are so slow in comparison to a high end card like a gtx285 that it is faster for the gtx 285 to just do all the work.

No one really knows the numbers yet on a 480. It could be the same deal. It's so freaking fast that it is faster to render a game and crunch physx at the same time than wait for the 8800gt to just crunch the physx

Or maybe not. We don't know till the card is actually out.
 
It does make sense in the way he is thinking though. GTX 480=512SP correct? Now just take the 112 SP's away from the GTX480 for Physx(Really, it will probably call for 128SP). The card still has 400SP's left for gaming. And as we all know, no game coming in the next year is really going to push that amount of power on the GTX480. Nothing that can't be run on a 240 SP GTX285 anyway.(Triple monitor gaming excluded)

So basically, the GTX480 can do it all at the same time, where as in his mind there is a slight delay offloading to the 8800GT here lies the "faster on one card" than offloading it to a seperate card. It actually makes sense if you really look at it, and technically I would think there is actually a ms or 2 in that delay to offload.;)

I doubt you'll run into scenarios where anything less than 512SP's will be utilized. It would be a massive oversight on the engineering side if a dynamically allocated processing pool was so mismanaged that you had actually had anywhere near 112SP's sitting idle to offload physics calculations alongside graphics calculations. There isn't some "magical resource ceiling" where games don't use extra resources afforded them; they take as much as they can get (unless you're dealing with a VS/PS fixed setup, in which case compute resources need the correct ratio VS:pS for the average usage scenario).

Really slow cards decelerate PhysX compared to a single performance GPU because they don't reach a critical threshold of processing power & especially bandwidth. It is not a relative measure, so its not like 8800GT + 280GTX is good but 480 GTX + 8800GT is somehow magically decelerating PhysX calculations.

Look at it this way. In what case is 512SP's + 112SP's slower than 512SP's for an architecture with similarly balanced SP/GB/s?
 
Last edited by a moderator:
Why not just wait until it is out to see how it really performs and if it is worth the price vs performance

Curiuous, how many games these days are using PhysX fully and in a way that makes the game better?
 
I think you should sell both, If you havent seen physics videos on fermi, they are pretty amazing.
 
not sure what outputs Fermi will have. I always keep extra cards for replacements in case of failure and to use triple monitor support. Until now, a single GTX280 cannot drive 3 monitors (conventionally speaking, not sure about the s-video). Anyway, keep the card. The money you make isn't worth the trouble of having a new card not working and finding a temporary replacement.

Besides, it really is a worthy keeper and still strong enough to run most games very well, especially at 1680*1050.
 
I can almost guarantee you of the results you'll see...

A GTX480 with 8800GT dedicated to PhysX will have higher maximum framrates, but lower minimum frame rates. With PhysX dedicated to the 8800GT, you've got 112 SP's to work with, and that's it. If you need more power for PhysX, there's nowhere to get it, so your frame rate gets dragged down even if there's extra processing power available on the GTX480. Meanwhile, the GTX480 is free to render like crazy when the 8800GT isn't bogging it down, but getting 200FPS sometimes is kinda worthless when your minimums are still below 60.

A GTX480 doing Graphics + PhysX will have lower maximum frame rates, but higher minimum frame rates. You'll have more total stream processors to balance with, so the only thing that puts a ceiling on PhysX performance is how intensive graphics rendering is at that very moment. Unless games start getting a lot more demanding really soon, that's not much of a ceiling.
 
That was my thinking and I have a 285 GTX. G400 apparently handles PhysX more efficiently than G200 and previous gens, like 3x better according to Nvidia. I had plans to keep the 285 as an overkill PPU, but now I'm just gonna part it out and save some bux on Fermi (or something ATI DX11 should Nvidia shit the bed). No point crippling your min FPS when a single GPU can handle it with grunt to spare.
 
Why not just wait until it is out to see how it really performs and if it is worth the price vs performance

Curiuous, how many games these days are using PhysX fully and in a way that makes the game better?

If you mean GPU PhysX, almost none. http://www.nzone.com/object/nzone_physxgames_home.html is a list of GPU PhysX games. 3-5 AAA titles on that list since 2007. How much GPU PhysX improved game play or did not is not a debate I wish to enter into again.
 
For those doubting that a 8800GT would benefit a 480GTX for physics performance, here is a review where a GTX 275 is paired with a GT220, and shows dramatic improvements in framerate. The relative performance parity is about as large between the GTX 275 and the GT220 as between Fermi and 8800GT (about 1/4 the compute resources), and obviously the objective performance, as in what the resource ceiling is for PhysX effects, is in a completely different league on the 220 vs the 8800GT.http://www.xbitlabs.com/articles/video/display/evga-gtx275-coop-physx_8.html#sect3
 
Last edited by a moderator:
480GTX + 8800GT will only boost your framerates, minimum frames included. If you honestly think that the 480GTX will have anywhere near 112SP's just idling when physics effects are going on, you're just wrong, no offense.
You kinda missed the point. With an 8800GT, you have 112 SP's to work with, period. Some PhysX games have settings that demand more physics processing power than an 8800GT can provide, and therein lies the problem.

This creates a bottleneck, where the GTX480 is waiting on the 8800GT. This caps off your framerate to the speed at which that 8800GT can churn out those physics calculations. Ironically, because the GTX480 is now forced to sit and wait for the 8800GT to catch up, it now has free steam processors anyway! :p

Moving PhysX onto the GTX480 allows it to load balance PhysX across more stream processors (if need be), meaning you won't experience capped framerates in situations where the 8800GT can't keep up. There's also the option of using a faster dedicated card; something around a GTS250 or a GTX260 should prevent bottle-necking of a GTX480.
 
Last edited:
Moving PhysX onto the GTX480 allows it to load balance PhysX across more stream processors (if need be), meaning you won't experience capped framerates in situations where the 8800GT can't keep up. There's also the option of using a faster dedicated card; something around a GTS250 or a GTX260 should prevent bottle-necking of a GTX480.

No, you missed the point. The 480GTX has 0 SP's to work with for PhysX. Any amount of work it does on PhysX is going to be deducted from graphics process time with the addition of some time to switch context. You're imagining Fermi as some endless well of performance. It isn't. Fermi's "cap" as you call it, is 1500Gflops. That's already barely enough for it to get to + 30% of 5870 (probably the average performance advantage), and that's for graphics only tasks.
 
Last edited by a moderator:
8800 GT provides 500Gflops (or 333 if you can't utilize MUL) of compute resources for PhysX. Fermi will not have a spare 500 or even 333 gflops of compute power, ever.

So if you want to talk about 8800GT being a performance "limiter" why don't you show me...
Sure, no problem. As I said before, in situations where graphics aren't very intensive, an 8800GT very quickly becomes a performance limiter (even with a last-gen video card such as the GTX260 being used as the dedicated rendering GPU).

Lets use fluidmark, a PhysX benchmark, as an example. The graphics aren't very demanding, but the PhysX effect it uses is. We see a 8800GT very badly bottlenecking and holding back a GTX260 here. The GTX260 has power to spare with the simple graphics of the benchmark, but is left sitting and waiting for the 8800GT to catch up (exactly as I described).

Graphics + PhysX both running on a GTX260 = Minimum FPS 104
Graphics on a GTX260 + PhysX on a 8800GT = Minimum FPS 73

Well, would you look at that. The GTX260 has so much extra power to spare in this situation that it can run the benchmark faster without the 8800GT holding it back! There will, no doubt, be games that fit this profile. Not very visually demanding, but requiring a lot of physics processing. Such games will most likely perform a lot better with a GTX480 doing graphics and PhysX, rather than using a relatively underpowered dedicated PhysX card like an 8800GT.

If you want a dedicated PhysX card for the GTX480, I would recommend using something much faster than an 8800GT
 
Last edited:
im going to say a GTX480 wont need a 2md card for physx atm
even on high the one thing its good as GPGPU stuff and the card has a ton of power for that kind of stuff but not so hot at every thing else my guess is the card will shine when you can use it for more then just 3D be it tessellation or physics or some thing else wile doing 3D
 
I guarantee that you won't find a PhysX game where Fermi has so much power to spare that it can render visuals and calculate physics at a clip faster than it can render visuals alone with an 8800GT holding it back, which is what you're suggesting.
Added in (in red) the part you failed to mention.

Fact remains, under the right circumstances (which is what I've been saying all along) an 8800GT dedicated to PhysX can hold back faster cards. If it can drag down a GTX260, what do you think it'll do to a GTX480?

This is going to be one of those hard-to-benchmark things in real games, because it completely changes the performance curve. There will be situations where having the GTX480 free to render only graphics with a dedicated 8800GT for PhysX will improve performance (maximum framrates), and there will be situations where the 8800GT will hold it back because it can't keep up (minimum framrates)

So as much as you want to believe otherwise, I don't see any evidence that 333-500Gflops of compute power dedicated to physics will hold back a Radeon 5800-class gpu.
Well, as it just so happens, I just upgraded to an HD5850. This left me with a GTX260 and a 9800GT to decide between for use as a dedicated PhysX card.

I can tell you right now, the 9800GT (a supposed 333-500Gflop card) can hold back an HD5850 rather badly; it simply wasn't enough. With an HD5850 doing graphics and a 9800GT doing PhysX, minimum FPS in Batman Arkham Asylum showed no improvement over the GTX260 + 9800GT setup I was running before (13FPS minimum).

I swapped the 9800GT for the GTX260 dedicated for PhysX, and finally saw performance jump up (36FPS minimum). That faster PhysX card really let the HD5850 stretch its legs.

So as much as you want to believe otherwise, the 8800GT / 9800GT is now sub-par as a dedicated PhysX card for highend video cards like the HD5850. The bare minimum I would go when your main graphics card is this fast, is a 9800GTX / GTS250.
 
You're right, I don't believe you. Mainly because the GTX 260 has nowhere near 3x the performance of a 9800GT.

Graphics performance does not directly equate to physics performance. The GT200 core on the GTX260 is quite a bit more optimized for CUDA processing (which includes PhysX) than the old G92 core. You get a lot more PhysX processing power per-clock and per-stream-processor with the GT200 core.

Going from 13FPS minimum to 36FPS minimum was a 27.6% increase.

The GTX260 has 216 stream processors, the 9800GT only has 112. If everything else remained a constant, that alone is a 19.2% increase. Architectural improvements easily make up the last 8.4%
 
Graphics performance does not directly equate to physics performance. The GT200 core on the GTX260 is quite a bit more optimized for CUDA processing (which includes PhysX) than the old G92 core. You get a lot more PhysX processing power per-clock and per-stream-processor with the GT200 core.

Going from 13FPS minimum to 36FPS minimum was a 27.6% increase.

The GTX260 has 216 stream processors, the 9800GT only has 112. If everything else remained a constant, that alone is a 19.2% increase. Architectural improvements easily make up the last 8.4%

FYI, 9800GT to GTX 260 192 theoretical math improvement is not 192%. It's 41%. 192/112 * 1242/1500 = 1.41. I'd go ahead and check your math. Per clock improvement is almost linear with respect to increase in SP vs hot clock, which is nowhere near 2.76x. It's close to 1.4x. Theoretical math improvement (MUL isn't utilized for any of the physics code I've seen, explained below) for 260GTX 192 vs 9800GT ~ 1.41; for 280GTX vs 9800GTX it's ~ 1.45 (these cards used in the test below)

Most of the CUDA performance documents I've seen treat both GT200 and G80/G92 as 2 flops/c. Go ahead and look at speedup in some CUDA apps:

http://www.ks.uiuc.edu/Publications/Papers/PDF/HARD2009/HARD2009.pdf
http://compview.titech.ac.jp/Members/itoht/7b2c556de/5c0f5ddd-6167/handout-ogawa
http://www.prace-project.eu/hpc-training/training_pdfs/2650.pdf
https://yallara.cs.rmit.edu.au/~pknowles/GPGPUPS.pdf

There are a few cases where GT200 really outperforms G80 (no atomics) and G92 (max 50% mul co-issue), but I haven't seen any evidence of that kind of gain for SPH; and navier-stokes stuff is probably the only thing that really stresses these gpu's (it certainly isn't rigid body low 10's of thousands particle systems)

edit: Just to make it clear: If you had a 100% linear speedup, ceteris paribus like you say, you would see a jump from 13 fps to ~18.5. That's 41%. You're seeing a 176% gain. You'd need roughly a 824 core GTX 260 to pick that up. Even if this PhysX driver ran on rainbows and fairy tales, and fully utilized MUL on GT200 but somehow got 0 MUL utilization on the G92, you'd still need ~ 520 cores.
 
Last edited by a moderator:
260GTX 192 vs 9800GT
You missed the part where I said I was using a GTX260 216, not the gimped 192, as a PhysX card.

Also, just to be abundantly clear, I'm using a newer 9800GT which lacks a 6-pin power connector. It's clocked slightly lower than the older version so it can run on the 75w provided by a PCIe slot.

Most of the CUDA performance documents I've seen treat both GT200 and G80/G92 as 2 flops/c.
They treat the G80 and the G92 the same? That's a laugh; it's pretty well known the old G80 core isn't great at handling PhysX compared to G92 and GT200.

edit: Just to make it clear: If you had a 100% linear speedup, ceteris paribus like you say, you would see a jump from 13 fps to ~18.5. That's 41%. You're seeing a 176% gain. You'd need roughly a 824 core GTX 260 to pick that up.
You're assuming the GTX260 is over-performing. Had it occurred to you that the GTX260 is operating normally, and the 9800GT is under-performing? I'll omit and ignore your rudeness, but you seriously need to come down off your high horse.

Now, given Nvidia's recent track record as far as drivers go, I could believe it's a driver issue gimping PhysX performance on G92 based cards. I'll run some tests to see what I can get out of the 9800GT; according to your numbers (adjusted for a GTX260 216 instead of a 192) the 9800GT should be able to pump out 32FPS minimum (12.5% slower than a GTX260) if we have completely linear scaling.

I would love for the 9800GT to be that close in performance to a GTX260, I really would. It would mean I could use a 70w card instead of a 182w card for PhysX and still get decent performance. Honestly though, that number seems pretty far-fetched for a card as slow as the 9800GT, but like I said, I'll see what I can do :rolleyes:

Please for the love of not turning my brain to complete sludge, don't reply telling me that you're getting 800 cores worth of compute power from your 260 GTX 192 :p
I never claimed to have a mythical uber-GTX260. I've been saying that the 9800GT is too slow this whole time. :rolleyes:

I'm just telling you what I saw with batman's built in benchmarking tool:
GTX260 + 9800GT = 13FPS minimum
HD5850 + 9800GT = 13FPS minimum
HD5850 + GTX260 = 36FPS minimum

I'm going to do some more testing with the 9800GT in my system. Older drivers, overclocking it, etc. If the 9800GT suddenly jumps up to over 30FPS minimum, I'll be sure to let you know.
Oh, by the way, I'm not the only one around here seeing weird PhysX performance. SirKronan over in this thread can't seem to get minimum framrates in Mirrors Edge over around 55FPS with a bunch of different dedicated PhysX cards tested, no matter what he does (he's also mentioned weird slowdowns all the way to 3FPS).
 
Last edited:
FYI, 9800GT to GTX 260 192 theoretical math improvement is not 192%. It's 41%. 192/112 * 1242/1500 = 1.41. I'd go ahead and check your math. Per clock improvement is almost linear with respect to increase in SP vs hot clock, which is nowhere near 2.76x. It's close to 1.4x. Theoretical math improvement (MUL isn't utilized for any of the physics code I've seen, explained below) for 260GTX 192 vs 9800GT ~ 1.41; for 280GTX vs 9800GTX it's ~ 1.45 (these cards used in the test below)

I would suggest using reaver tuner and setting the SP clock to 1500 and the core to 600, the 192 can handle that very easily.
 
I've been messing with the 9800GT again, and after both updating to the latest PhysX system software and overclocking the 9800...the minimum frame rate in Batman Arkham Asylum has gone up to 15FPS. Older drivers only made PhysX performance worse.

Any suggestions before I put the GTX260 back in?

G92 and G80 both have very similar CUDA capability. G92 adds atomics support, but afaik it doesn't matter a whole lot for physics performance, where you aren't bandwidth starved.

Then there's something that you aren't accounting for that leads to seriously slow PhysX performance on G80, because it does not perform the same as equivalent G92 and GT200 based solutions. Period.
 
I've been messing with the 9800GT again, and after both updating to the latest PhysX system software and overclocking the 9800...the minimum frame rate in Batman Arkham Asylum has gone up to 15FPS. Older drivers only made PhysX performance worse.

Any suggestions before I put the GTX260 back in?



Then there's something that you aren't accounting for that leads to seriously slow PhysX performance on G80, because it does not perform the same as equivalent G92 and GT200 based solutions. Period.

Err, it scales almost linearly from G80 to GT200 compared to ALU*hotclock? That was well demonstrated in my 2nd to last post. Period

And G92 to GT200 scaling will be slightly less impressive.
 
I just put the GTX260 and 9800GT together again in a different computer entirely; that should rule out any other hardware or software, specific to my configuration, that might be holding the 9800GT back.

Once again, I ended up with about the same result, around 14FPS minimum.

I really don't think a 9800GT is capable of what you're claiming. There's something you haven't accounted for that's leading to much slower PhysX performance on the older G80 and G92 cores.

Edit: SirKronan, from the thread I linked previously, has ordered a GTS250 to test as a dedicated PhysX card. That should make a nice comparable; it's the fastest G92 card, and its performance lands somewhere between a 9800GT and a GTX260.

Err, it scales almost linearly from G80 to GT200 compared to ALU*hotclock?
Obviously not. PhysX has always performed slowly on G80 based cards, especially when you force them to load balance between graphics and PhysX.

That was well demonstrated in my 2nd to last post. Period.
The link in your previous post showed the G80 getting whooped by the GT200. It even starts out with a bulleted list of all the major improvements that GT200 offers over G80...
 
Last edited:
You argument pretty much rests on the idea that I'm not accounting for something. I have nothing else to add. After accounting for clock and alu count it is not much faster for the only physics cuda examples I could find. It's posted. You can't analyze the data I gave you. I'm really sorry about that, but I don't have any more time. Plus I really dont care.
 
Back
Top