GT640 for PhysX - passive operation possible?

[notty22]

There is a chance the 580sli system will respond positively to the dedicated gt 640. Kepler architecture seems to handle physx very well.
When a rig is in SLI, it's physX performance is not anywhere close to having a dedicated card.
This is one of those situations that needs to be tested to confirm the experiment. IMO.

 

[boxleitnerb]

Ok guys, problem solved...kinda.

My board cannot take a third card with only a Sandy Bridge CPU installed. I would need an Ivy Bridge CPU and that would put my 3 slots at 8x/4x/4x. Not worth it imo. Somehow I forgot to check the manual because I got so exited

Never will I buy a crippled board again. The next one has 4 slots and at least one PLX chip, for sure!

 

[Keysplayr]

I am off? you actually think a gt640 can do a better job at handling physx than gtx580 sli can at graphics and physx combined. that is nonsense. a gt640 is probably not much faster than a 9800gt so you are giving it way too much credit.

Toyota have you actually tested this? That GTX580SLI can do a better job at handling graphics & physx better than gt640 can running just the physx? Or have you read any article where someone uses a GT640 for dedicated PhysX?

 

[Keysplayr]

Ok guys, problem solved...kinda.

My board cannot take a third card with only a Sandy Bridge CPU installed. I would need an Ivy Bridge CPU and that would put my 3 slots at 8x/4x/4x. Not worth it imo. Somehow I forgot to check the manual because I got so exited

Never will I buy a crippled board again. The next one has 4 slots and at least one PLX chip, for sure!

WTF?

 

[ShintaiDK]

Because my GTX580 SLI is watercooled and my whole rig is silent.

Before anyone asks:
This setup is not enough. With PhysX at high I get asymmetric GPU load and fps drops to 45fps in Borderlands 2. With PhysX at low I get a steady 60fps. PhysX+SLI = not optimal. I could dedicate the second 580 to PhysX, but then I would have to cut back on the settings (SGSSAA), which I quite frankly don't want to do

Either fit it to the loop or forget it. socalled "passive GFX cards" need active airflow from other components. Specially CPU cooler.

But why not just run PhysX on the 580s? You aint gonna get magically good speed doing it on a GT640. Specially when you gonna way for the transfers and latency between the cards. And a GT640 might actualy be too slow.

Only reason to do this is essentially if you use an AMD GPU as primary.

 

[railven]

So sorry to thread steal, swapped out the 9800 GTX+ for the GTX 460 and saw a performance boost from tanking to 30 FPS with no offloading (before I figured out how to work the mod), to tanking to 40 FPS with the 9800 GTX+, to sliding down to about 57 FPS with the GTX 460.

I ran into another issue which isn't PhysX related (at least that I think of it) and I wonder if it is my PC or the settings I use. Will test tomorrow on the GTX 680 system.

With the FPS limit set to 60 through the settings, I'd notice an occasional dip to 45 FPS when just starting into the horizon and turning the camera a tad. I took of the FPS limit and the drop is still *felt* but the FPS counter never drops below 80 (it dips from 100 to 80s though.)

Oh wells. I benched Batman: AC with the 460 and saw no gain during the hall way scene, rather disappointing.

 

[boxleitnerb]

WTF?

Lol, nobody's perfect, stuff like that happens...sometimes

Next setup will be 3xGK110, then I will not need a dedicated extra PhysX card anyway.

 

[Dark Shroud]

Ok guys, problem solved...kinda.

My board cannot take a third card with only a Sandy Bridge CPU installed. I would need an Ivy Bridge CPU and that would put my 3 slots at 8x/4x/4x. Not worth it imo. Somehow I forgot to check the manual because I got so exited

Never will I buy a crippled board again. The next one has 4 slots and at least one PLX chip, for sure!

Generally you want to avoid PLX chips if possible.

They add latency and do not add bandwidth, they only manage it hence the higher latency.

We shouldn't have this problem on LGA 2011 socket CPUs or Intel's "Haswell."

 

[boxleitnerb]

How will Haswell be different? As far as I know it will only offer 16 lanes just as Ivy.

I would love to go socket 2011, but I don't like Intel's policy of releasing the enthusiast parts so late. I don't want to buy obsolete stuff. I had planned to get a new system when Haswell and GK110 are released and I'd rather not wait for Ivy-E until Q3 2013.

 

[Ancalagon44]

its a fact that based on benchmarks, a gt640 has about the same power as a 9800gt. well here is some reality for you. here you can see a gtx480 with different cards being used for physx. now factor in that a gtx580 is 20% faster than a gtx480. that would make the baseline about 60fps. so a gt640/9800gt level card would be below what that 9800gtx is given here in Mafia 2. and this with just ONE gtx580, with TWO gtx580s a freaking gt640 would be slower.

Yeah but thats not what the OP is asking. He is not asking "should I buy a second 580", he is asking, "would a 640 handling the PhysX increase my FPS".

Based on the chart you posted, the answer is yes. A GTX 480 on its own gets 50 FPS average, and a 480 with a 240 running physics gets 77. Thats more than a 50% improvement - quite impressive if you ask me. Plus it pushes the average above 60 and the min gets a lot closer to 60.

Why are you against having a second, weaker card for PhysX? To me it looks like a slam dunk.

 

[Smoblikat]

um a gt640 a will actually slow down your system as opposed to letting the gtx580 sli run it. a gt640 will offer very little gain over having even a single gtx580 run physx and graphics.

I dont think you understand whats happening..........

Yes, a 640 would be perfect. You might look into a 450/550 though, thet overclock fairly well and come in single slot flavors.

 

[toyota]

Yeah but thats not what the OP is asking. He is not asking "should I buy a second 580", he is asking, "would a 640 handling the PhysX increase my FPS".

Based on the chart you posted, the answer is yes. A GTX 480 on its own gets 50 FPS average, and a 480 with a 240 running physics gets 77. Thats more than a 50% improvement - quite impressive if you ask me. Plus it pushes the average above 60 and the min gets a lot closer to 60.

Why are you against having a second, weaker card for PhysX? To me it looks like a slam dunk.

he already has TWO gtx580 cards.

 

[toyota]

I dont think you understand whats happening..........

Yes, a 640 would be perfect. You might look into a 450/550 though, thet overclock fairly well and come in single slot flavors.

now it would not. using one of his current 580 for graphics and other for physx would be faster than trying to offload physx onto a 640.

 

[boxleitnerb]

he already has TWO gtx580 cards.

Yes, but...

...they are running in SLI and are being taxed quite well with graphics as it is (I use 4xSGSSAA, not FXAA).There is not enough power for PhysX anymore. Why don't you believe me when I tell you I have nasty slowdowns to 45fps which really disrupt gameplay (otherwise locked at 60fps).
Using the second 580 dedicated to PhysX is not an option. I hope it is clear now.

However, I believe something is a bit broken with PhysX, because when I turn off Ambient Occlusion which should give a nice boost, these slowdowns still exist, nothing changes. This happens mostly when looking at cloth and fighting (many particles).

 

[Ancalagon44]

now it would not. using one of his current 580 for graphics and other for physx would be faster than trying to offload physx onto a 640.

I would think an entire 580 would be wasted on PhysX.

OP, why not see if you can borrow an Nvidia card to test? Pretty much any recent card will give you an idea whether a 640 would speed things up or slow them down.

 

[Smoblikat]

now it would not. using one of his current 580 for graphics and other for physx would be faster than trying to offload physx onto a 640.

Youre right.........a 580 is faster than a 640.............but that leaves him with half of the graphics power he had, meaning it will actually decrease performance more than having no physX card at all. These arent difficult concepts to grasp.

 

[boxleitnerb]

I would think an entire 580 would be wasted on PhysX.

OP, why not see if you can borrow an Nvidia card to test? Pretty much any recent card will give you an idea whether a 640 would speed things up or slow them down.

Well, the original question was about the 640 running passive
But since I found out I cannot populate the third PCIe 16x slot with my Sandy Bridge CPU, that has become irrelevant for the moment.

Are there more PhysX titles on the horizon? I would want to get a decent mainboard with sufficient PCIe lanes for 3-way SLI anyway. Problem is latency, but would that really matter much? Does someone have 2-way/3-way SLI benchmarks on Z77 vs. X79 with the same CPU clock speed?

 

[toyota]

Yeah but thats not what the OP is asking. He is not asking "should I buy a second 580", he is asking, "would a 640 handling the PhysX increase my FPS".

Based on the chart you posted, the answer is yes. A GTX 480 on its own gets 50 FPS average, and a 480 with a 240 running physics gets 77. Thats more than a 50% improvement - quite impressive if you ask me. Plus it pushes the average above 60 and the min gets a lot closer to 60.

Why are you against having a second, weaker card for PhysX? To me it looks like a slam dunk.

you need to look at that again. thats a 9800gtx getting 77 fps not a gt240. and a single gtx580 would be getting about 60 fps so the baseline would be higher for the OP.

 

[RussianSensation]

I would think an entire 580 would be wasted on PhysX.

Not even close. It appears that a modern Intel CPU can handle the entire PhysX of BL2 with a Radeon HD7970 dedicated to graphics. You get the same performance as having a GTX680 doing graphics + PhysX.

^ If that is not an accurate review, let's use a secondary source. They used a dedicated GTX690 for PhysX in BL2.

GTX690 (Card B) is dedicated entirely to PhysX

12% GPU usage......

It's way better to ditch those 580s and get 680s in SLI instead of wasting $ on a $100 GT640. Also, SuperSampling can get very demanding as well. That in itself can bring a huge performance hit at 4x. Why? Because with 4x SSAA, GTX580 is 37% slower than a 680 and 50% slower than an HD7970 GE in games. So you are looking at roughly a 37% performance increase right off the bat going from 4xSSAA 580s to 680s. And that's going to be everywhere not just when there are PhysX on screen.

You can't expect to run modern games with GTX580 SLI with 4xSSAA and not have situations where massive slowdowns occur, such as cases of 45 fps drops.

Supersampling is an entirely different ballgame:
http://www.computerbase.de/artikel/grafikkarten/2012/test-grafikkarten-2012/18/

 

[RussianSensation]

They add latency and do not add bandwidth, they only manage it hence the higher latency.We shouldn't have this problem on LGA 2011 socket CPUs or Intel's "Haswell."

This is not true on 3rd/4th videocard's end though. I believe what you are talking about is the entire platform's PCIe bandwidth is fixed but that doesn't matter since even Haswell Quads will be limited to 16x PCIe 3.0 iirc.

The whole point of the PLX chip is that by redistributing PCIe lanes, it indirectly results in effectively more PCIe bandwidth to the 3rd and 4th GPUs. The side-effect is added latency (1-3% performance hit) but you do get an additional added bandwidth for the 3rd GPU for sure. Otherwise, there would be no point to the PLX chip at all.

If you have a motherboard with an IVB CPU, it has 16x PCIe 3.0 lanes. Modern GPUs do not need 16x PCIe 3.0. With the PLX switch, you can redistribute that bandwidth to 8x/8x/8x PCIe 2.0 and feed all 3 GPUs. You didn't create new bandwidth but because of redirecting it more effectively, the end result is more bandwidth for the 3rd GPU. This isn't as good as having 32 PCIe 3.0 lanes on the X79 platform but even PCIe 2.0 8x is still not a huge bottleneck for modern GPUs.

http://www.anandtech.com/show/6170/...x-8747-featuring-gigabyte-asrock-ecs-and-evga

The PLX switch can get take IVB's PCIe 3.0 x16 and give you PCIe 2.0 x8/x8/x8/x8 [because PCIe 3.0 x16 = PCIe 2.0 x32 = PCIe 2.0 x8 (x4 PCIe slots)].

What you end up with is GTX680 Tri-SLI for example on PCIe 2.0 x8, or a 5-6% performance hit vs. GTX680 Tri-SLI on X79 platform. On the Tri-Crossfire, the performance hit will be more like 3-4%.

While it is true that PLX switch doesn't magically create more PCIe 3.0 x16 lanes for the platform as a whole, the the way it redirects the PCIe lanes does result in more bandwidth for the 3rd/4th videocards which means it actually works as intended, but is of course not as good as having native PCIe 3.0 x32 lanes.

 

[railven]

Not even close. It appears that a modern Intel CPU can handle the entire PhysX of BL2 with a Radeon HD7970 dedicated to graphics. You get the same performance as having a GTX680 doing graphics + PhysX.

^ If that is not an accurate review, let's use a secondary source. They used a dedicated GTX690 for PhysX in BL2.

Good to know those morons at Tech Spot did a shoddy review and now that information is being posted as fact.

YOU CAN NOT RUN ALL PHYSX EFFECTS ON CPU. AS A HYBRID USER I KNOW.

TecH Spot didn't properly test PhysX, by their own admission in the comments, the author of the article only shot walls and cloths. Once you get to the fluid portion of PhysX YOUR SYSTEM WILL TANK.

Need proof?
PhysX Blood effect tanking Intel i5 2500k @ 4.4ghz w/ Radeon 7970 @ 1125/1525:

Same scene, blood clears, FPS rebounds:

Here is me testing with PhysX Mod working, notice the FPS is still crippled (this is with 9800GTX+ offloading):

Do not go by that Tech Spot article, it is poorly done and doesn't properly test PhysX in a Hybrid system.

EDIT: Updated picture with GTX 460 offloading, exact same area as the picture directly above [if you look, I'm standing on the metallic object looking at where the barrel is as I detonate it to get goo every where]

You can not run PhysX high with just a Radeon, it will tank FPS the moment you get any liquid on screen.

 

[boxleitnerb]

It's way better to ditch those 580s and get 680s in SLI instead of wasting $ on a $100 GT640. Also, SuperSampling can get very demanding as well. That in itself can bring a huge performance hit at 4x. Why? Because with 4x SSAA, GTX580 is 37% slower than a 680 and 50% slower than an HD7970 GE in games. So you are looking at roughly a 37% performance increase right off the bat going from 4xSSAA 580s to 680s. And that's going to be everywhere not just when there are PhysX on screen.

You can't expect to run modern games with GTX580 SLI with 4xSSAA and not have situations where massive slowdowns occur, such as cases of 45 fps drops.

Supersampling is an entirely different ballgame:
http://www.computerbase.de/artikel/grafikkarten/2012/test-grafikkarten-2012/18/

I'm upgrading when GK110 is here. I will definitely not switch to AMD for several reasons and the 680s are a joke for the money. I paid 1000 Euros for 2x580 3GB and I'll only upgrade when I get about 70% more performance for the same price.

As for SSAA - the game runs fine with 4xSSAA as long as I don't have PhysX enabled. I get 70+ fps all the time. Why should I not not use the power that I have to get better image quality? I'm playing with VSync anyway, so anything above 60fps is wasted. I'd rather have those fps dips than live with the massive aliasing which personally I find horrible.

Thanks for the explanation with the PLX chips!
I'm still contemplating what my plattform will be exactly. If Ivy-E were to be released with Haswell, that would be a no brainer. But I don't like the wait until Q3. But I also don't like to buy an "outdated" CPU/plattform (Ivy/Sandy vs. Haswell; and there supposedly will be a second revision of X79 boards) or lose 5% performance either. I mean if I get 3 GK110, why make compromises elsewhere that minimally change the budget. Fanboys bash their heads in over 5%
Aside from that, I have heard that PCIe 3.0 isn't properly working in X79 in all configurations, hence the PCIe 3.0 patch by Nvidia. Not exactly enticing.

 

[ViRGE]

As an aside, has anyone recently benchmarked PhysX on an SLI system and compared it to a single GPU system to measure the impact of PhysX? The last time I looked, PhysX did evil, evil things to SLI performance since PhysX is confined to one GPU (thereby messing up frame metering).

 

[f1sherman]

Messing up AFR itself you mean, never mind frame metering.

I doubt 15ms, 30ms, 15, 30, 15, 30 etc ... is any worse than say 23, 30, 23, 30

 

[boxleitnerb]

As an aside, has anyone recently benchmarked PhysX on an SLI system and compared it to a single GPU system to measure the impact of PhysX? The last time I looked, PhysX did evil, evil things to SLI performance since PhysX is confined to one GPU (thereby messing up frame metering).

I can try that later today. I'm pretty sure you're right - I don't expect better than 60% scaling.