One GPU can NOT handle physx + video: benchmarks

Status
Not open for further replies.
J

jebo_4jc

[H]ard|DCer of the Month - April 2011
2FA
Joined
Apr 8, 2005
Messages
14,572
At least, not at a good resolution. Techgage tested in UT3 with hardware physX acceleration on using a 9800GTX and a 3ghz qx9650. They tested first with physics handled on the CPU, then with physics on the GPU, then with physics on the PhysX PPU.

At 1680x1050, the average FPS is:

CPU: 31.4
GPU: 51.05
PPU: 60.5


Once the resolution is cranked up to 2560x1600:

CPU: 30.59
GPU: 30.83
PPU: 41.98

So, this confirms what I and several other people have suspected: when the GPU is under heavy load, it can not handle 3D and physx at the same time.

http://techgage.com/article/nvidias_physx_performance_and_status_report/1
 
Not really. A 9800GTX at 2560x1600? I wanna see all test done at 2560x1600 then.

Edit: Kyle, is there any chance [H] can do a test on this?
 
What I was trying to say was, for what is being tested it is not consistent enough. 1680x1050 compared to 2560x1600 ... BIG difference.
 
jebo_4jc said:
At least, not at a good resolution. Techgage tested in UT3 with hardware physX acceleration on using a 9800GTX and a 3ghz qx9650. They tested first with physics handled on the CPU, then with physics on the GPU, then with physics on the PhysX PPU.

At 1680x1050, the average FPS is:

CPU: 31.4
GPU: 51.05
PPU: 60.5


Once the resolution is cranked up to 2560x1600:

CPU: 30.59
GPU: 30.83
PPU: 41.98

So, this confirms what I and several other people have suspected: when the GPU is under heavy load, it can not handle 3D and physx at the same time.

http://techgage.com/article/nvidias_physx_performance_and_status_report/1
Click to expand...

Did you even read the conclusions ?!
Obviously that @ such a high and demanding resolution, the GPU is too busy with the graphics load. Also, the 9800 GTX is not exactly the most powerful GPU around, much less for such a high resolution. We need tests with a GTX 260 and especially GTX 280.
 
Silus said:
Did you even read the conclusions ?!
Obviously that @ such a high and demanding resolution, the GPU is too busy with the graphics load. Also, the 9800 GTX is not exactly the most powerful GPU around, much less for such a high resolution. We need tests with a GTX 260 and especially GTX 280.
Click to expand...
yes I did read the conclusion. And no doubt you're right that more testing is needed with the high end cards, but how could you conclude the results will be any different? Perhaps at the same 2650x1600 res in UT3 the GTX280 has enough headroom to run 3D and physx at a good framerate, but in a more demanding 3D scenario the results will be the same.
 
Silus said:
Did you even read the conclusions ?!
Obviously that @ such a high and demanding resolution, the GPU is too busy with the graphics load. Also, the 9800 GTX is not exactly the most powerful GPU around, much less for such a high resolution. We need tests with a GTX 260 and especially GTX 280.
Click to expand...

agreed

also notice the large hit the PPU setup took when resolution was cranked......
 
jebo_4jc said:
yes I did read the conclusion. And no doubt you're right that more testing is needed with the high end cards, but how could you conclude the results will be any different? Perhaps at the same 2650x1600 res in UT3 the GTX280 has enough headroom to run 3D and physx at a good framerate, but in a more demanding 3D scenario the results will be the same.
Click to expand...

If it does, then you add a second card, until a more powerful GPU comes along.
 
Silus said:
Also, the 9800 GTX is not exactly the most powerful GPU around
Click to expand...
And they didn't exactly test the most demanding game. Imagine a real next gen game trying to handle graphics and physics on one gpu and I bet the bottleneck will come at a much lower res.

I think physics that can take full advantage of quad cores would be a lot more cost effective than needing a second graphics card.
 
Really, who is going to run a 9800GTX or even the GTX+ at 2560x1600? Lets see it run on a 280 please.

edit: for some strange reason I typed 2560x2400. (I think I did a little too much crack this morning. :()
 
If you have SLI or a dual GPU card will there be a way to dedicate 1 GPU to PhysX and one to graphics.
 
xp3nd4bl3 said:
If you have SLI or a dual GPU card will there be a way to dedicate 1 GPU to PhysX and one to graphics.
Click to expand...

If that's a question, the answer is yes, at least as far as using another card in sli. Nvidia is working on it and it will be done soon.
 
jebo_4jc said:
At least, not at a good resolution. Techgage tested in UT3 with hardware physX acceleration on using a 9800GTX and a 3ghz qx9650. They tested first with physics handled on the CPU, then with physics on the GPU, then with physics on the PhysX PPU.

At 1680x1050, the average FPS is:

CPU: 31.4
GPU: 51.05
PPU: 60.5


Once the resolution is cranked up to 2560x1600:

CPU: 30.59
GPU: 30.83
PPU: 41.98

So, this confirms what I and several other people have suspected: when the GPU is under heavy load, it can not handle 3D and physx at the same time.
Click to expand...

It's funny, because I saw this report and came to exactly the opposite conclusion. 50+ FPS with PhysX? Fantastic! And only a 16% loss compared to having a full add-on card.

Of course, you aren't implying that 50 FPS is too little, because if you are, then you really meant to say that the 9800GTX cannot handle the standard UT3 (non-PhysX levels) at 2560x1600 period. Either way, I'm not sure I understand your post.
 
SubtleOne said:
It's funny, because I saw this report and came to exactly the opposite conclusion. 50+ FPS with PhysX? Fantastic! And only a 16% loss compared to having a full add-on card.

Of course, you aren't implying that 50 FPS is too little, because if you are, then you really meant to say that the 9800GTX cannot handle the standard UT3 (non-PhysX levels) at 2560x1600 period. Either way, I'm not sure I understand your post.
Click to expand...
The article states that at 1600x1200 the 9800GTX can handle physx.

However, put a larger load on the GPU and it can't handle it anymore. From the OP:
jebo_4jc said:
At least, not at a good resolution....So, this confirms what I and several other people have suspected: when the GPU is under heavy load, it can not handle 3D and physx at the same time.
Click to expand...
 
jebo_4jc said:
unsubstantiated claim: double fail
Click to expand...
Sure, if you say so. Your own post shows the opposite of what your title claims, except for a ridiculous resolution few people can/do run, especially not with a single card of that generation.

Triple fail.
 
jebo_4jc said:
The article states that at 1600x1200 the 9800GTX can handle physx.

However, put a larger load on the GPU and it can't handle it anymore. From the OP:
Click to expand...

So you're saying that an average 50+ FPS is unplayably slow?
 
Trepidati0n said:
For some people, a min frame rate of 100 is not fast enough.
Click to expand...

Well, bear in mind that switching from a GPU + add-on card solution only goes from 50+ FPS to 60FPS, so if 100+FPS is the standard, then adding the card changes nothing.

Since the OP singled out the GPU-based PhysX solution as being inadequate, and that solution yielded 50+FPS, whereas the GPU + PPU gave 60FPS, I can only presume that 50+FPS is unbearably slow, whereas 60FPS is just fine.
 
pxc said:
Sure, if you say so. Your own post shows the opposite of what your title claims, except for a ridiculous resolution few people can/do run, especially not with a single card of that generation.

Triple fail.
Click to expand...
Logical reasoning FTL. You're not comprehending the conclusion: "when the GPU is under heavy load, it can not handle 3D and physx at the same time"
Forget which game it is, which resolution, etc. The bottom line is when the GPU is under heavy load it can not handle running physx calcs at the same time, evidenced by the fact that the framerate is essentially the same whether physics was run on the CPU or GPU.

SubtleOne said:
So you're saying that an average 50+ FPS is unplayably slow?
Click to expand...
no. I said at 1600x1200 the GPU CAN handle physx. push the 3d settings more, and it becomes "unplayably" slow.
 
Wait, so the op wanted to show us that the higher the resolution of a game, the lower the performance is from the video card?

Didn't we already know this?
 
Logical reasoning FTL. You're not comprehending the conclusion: "when the GPU is under heavy load, it can not handle 3D and physx at the same time"
Forget which game it is, which resolution, etc. The bottom line is when the GPU is under heavy load it can not handle running physx calcs at the same time, evidenced by the fact that the framerate is essentially the same whether physics was run on the CPU or GPU.
Click to expand...

I'm afraid your reasoning is flawed. It is pretty clear that a GPU that is about twice as fast could saturate your refresh rate. And last time I checked a 280 is about twice as fast as a 9800GTX.
 
vengence said:
I'm afraid your reasoning is flawed. It is pretty clear that a GPU that is about twice as fast could saturate your refresh rate. And last time I checked a 280 is about twice as fast as a 9800GTX.
Click to expand...
Think again -- it's your reasoning that's flawed. The conclusion is that the gpu can't handle physics and graphics when under heavy load. You're moving the goalposts by postulating a situation in which the gpu is NOT under heavy load -- a high end gpu running a totally non-demanding game.
 
errr 2560x1600 isnt a practical resolution that so many people run at. also drivers are still in beta are they not?
 
jebo_4jc said:
Once the resolution is cranked up to 2560x1600:

CPU: 30.59
GPU: 30.83
PPU: 41.98

So, this confirms what I and several other people have suspected: when the GPU is under heavy load, it can not handle 3D and physx at the same time.
Click to expand...

Well, correct me if I'm wrong, but it's not *slower* than the CPU, is it?
And it's still doing the physics workload.
So the GPU is doing more than it would with just CPU physics, yet it still gets the same framerate.
So it seems to me that it's still handling its physics. The CPU is now freed of its physics load and can be used for other things, while your performance remains the same.
 
rls669 said:
Think again -- it's your reasoning that's flawed. The conclusion is that the gpu can't handle physics and graphics when under heavy load. You're moving the goalposts by postulating a situation in which the gpu is NOT under heavy load -- a high end gpu running a totally non-demanding game.
Click to expand...

See here's the crux of this. You've taken a weak GPU, playing it at a resolution the GPU was never ment to be played at, on an old game, and now suddenly declaring it has some meaning when scaled to a current game and with a new GPU with a different architecture.
 
Scali2 said:
Well, correct me if I'm wrong, but it's not *slower* than the CPU, is it?
And it's still doing the physics workload.
So the GPU is doing more than it would with just CPU physics, yet it still gets the same framerate.
So it seems to me that it's still handling its physics. The CPU is now freed of its physics load and can be used for other things, while your performance remains the same.
Click to expand...
no. In the benchmarks the physics being run are the same, so when the physics are being run on the CPU, they are the same ones being run on the PPU or GPU (at least, that's what the article indicates).

vengence said:
See here's the crux of this. You've taken a weak GPU, playing it at a resolution the GPU was never ment to be played at, on an old game, and now suddenly declaring it has some meaning when scaled to a current game and with a new GPU with a different architecture.
Click to expand...
This is the only valid criticism I've seen.
However I think there are more than a few people that would argue the 9800GTX is far from weak.
In any case, games in the future will be using different engines which may or may not behave differently, and obviously the hardware in the new cards is different than G92, but what evidence have we seen that would indicate GTX280 would behave differently if it were put in a scenario where it is heavily stressed by a 3D workload?
 
jebo_4jc said:
Forget which game it is, which resolution, etc. The bottom line is when the GPU is under heavy load it can not handle running physx calcs at the same time, evidenced by the fact that the framerate is essentially the same whether physics was run on the CPU or GPU.
Click to expand...
This isn't rocket-science people. A chip can only process so many calculations in a given period of time. If you give that same chip more workload, it will take longer. I don't see why people are surprised (expect the ones that bought into Nvidia's hype for the past 3 years). How could you expect to offload something as complex as physics to the GPU and not have it effect the 3d rendering? That just doesn't even make sense.
 
Forget all this talk of architectures and new games and old games. I think it's pretty straightforward here...

A given device has the capability to do things. If you're using 50% of it's capabilities to do one thing, you have 50% left over. If you're stressing it on thing #1, pushing it at or near %100, alluva sudden it can't really do the other thing very well, now can it?

Shocking!
 
jebo_4jc said:
no. In the benchmarks the physics being run are the same, so when the physics are being run on the CPU, they are the same ones being run on the PPU or GPU (at least, that's what the article indicates).
Click to expand...

You're not getting what I'm saying.
When the GPU does physics it STILL gets 30 fps, even though it's doing the same graphics work as with the CPU physics, AND it is doing the physics workload that the CPU was doing previously.
So the GPU is still doing more work than it was doing with the CPU physics, because in that case it didn't have to do the physics next to the graphics. But performance remains the same. So saying it cannot handle it is wrong, if you ask me. The framerate is not affected by adding the physics workload to the GPU, so you get it 'for free'.
 
Scali2 said:
You're not getting what I'm saying.
When the GPU does physics it STILL gets 30 fps, even though it's doing the same graphics work as with the CPU physics, AND it is doing the physics workload that the CPU was doing previously.
So the GPU is still doing more work than it was doing with the CPU physics, because in that case it didn't have to do the physics next to the graphics. But performance remains the same. So saying it cannot handle it is wrong, if you ask me. The framerate is not affected by adding the physics workload to the GPU, so you get it 'for free'.
Click to expand...
Ok I see what you're saying. However the bottom line is this:

PhysX calculated on CPU with 3D on GPU: 31 fps
PhysX and 3D on GPU: 31 fps

So in this test, there is no benefit to running physX on the GPU.
 
xp3nd4bl3 said:
If you have SLI or a dual GPU card will there be a way to dedicate 1 GPU to PhysX and one to graphics.
Click to expand...

If this is the case and you have a GTX280 SLI setup, then Nvidia has made you paid $600+ for a dedicated Physx card disguised as a GTX280. I mean, we would've been better off in just running a single GTX280 and a standalone Physx addon card for much less and accomplish the same thing.

I personally don't see the logic in getting a second GTX280 just so you can run Physx on it.
 
jebo_4jc said:
So in this test, there is no benefit to running physX on the GPU.
Click to expand...

There is no loss either. Considering the rather extreme resolution, I'd say this is a perfect example of the opposite of what you claim.
Namely, the GPU is very well capable of handling a physics workload alongside the graphics work. In the worst case it's as fast as the CPU (which is a very high-end CPU, a regular dualcore Athlon or Core2 would perform much worse), in the best case, it's 66% faster.

I would also like to note that these were already test with a workload aimed at a separate processor. Until recently, it was impossible to play these levels without a PPU. And even this super-fast quadcore only gets 30 fps, completely limiting the videocard.
The GPU just tears through these physics at lower resolutions.
This means that it can run any 'regular' CPU physics-based game with ease, as that workload will only be a fraction of this test.
 
jebo_4jc said:
Ok I see what you're saying. However the bottom line is this:

PhysX calculated on CPU with 3D on GPU: 31 fps
PhysX and 3D on GPU: 31 fps

So in this test, there is no benefit to running physX on the GPU.
Click to expand...

We can all agree that the GPU is too busy doing graphics at high resolutions to do physics at the same time. Let's see what happens when physics is run in parallel on a gpu as opposed to a cpu with SLI. I think THAT'S what everyone wants to see.
 
vengence said:
See here's the crux of this. You've taken a weak GPU, playing it at a resolution the GPU was never ment to be played at, on an old game, and now suddenly declaring it has some meaning when scaled to a current game and with a new GPU with a different architecture.
Click to expand...
So what? The point is, a heavy graphics load is a heavy graphics load. What constitutes that will of course be different in each situation depending on the card, game, resolution etc. That doesn't change the fact that there is only so much processing power available and at some point splitting it between two tasks is going to hurt performance. Then you've got the choice of giving up the gpu physics, lowering your graphics settings, or getting another card.
 
jebo_4jc said:
Ok I see what you're saying. However the bottom line is this:

PhysX calculated on CPU with 3D on GPU: 31 fps
PhysX and 3D on GPU: 31 fps

So in this test, there is no benefit to running physX on the GPU.
Click to expand...

There is no problem running it on the GPU. Also, that is one case. I'd venture that at 1920x1200 you would still see improvements over running off the CPU. Also, there is no evidence to show that you would NOT get improvements from a 280. The 280 remember was designed to preform better at higher resolutions, AA, and cuda applications. The 9800 GTX was simply a die shrink of a midrange graphics card.
 
Status
Not open for further replies.
Back
Top