Shader gives u physx!!

[ajaidevsingh]

Ok from what i know steam processors are used by Nvidia and also by ATI for physx.

Now since the ATI cards say 3870 have more steam processors does that mean that enabled physx will impact lower on the ATI card FPS than on a Nvidia 9800 GTX which has only 128 steam processors as to 320 for the 3870??

If you all say that 9800 gtx will never support physx they dd make a promise that which ever card supports CUDA will support physx!!

If these seems too old a products one can take the example of 4870 "480 SPs" and GTX 280 "240 SPs"??

 

[taltamir]

ATI has WORSE SP performance, not better. while they have more SP units, their units are weaker individually, and are at a much lower clock speed then nVidias.

ATI also will not have physX for a long while...

nVidia bought AGEIA, which owned physX, they opened the standard so that ATI CAN develop their own physX on GPU, but nvidia already has an EXISTING physX engine from AEGIA that has been programmed for more then a year. They ported it to CUDA (C code on nvidia's SP) and will supposedly ship it with the G200, on the 18th of June. ATI will take a long while to come along with it. It will work with every CUDA capable GPU (IE, anything from nvidia with DX10 support)

ATI has to start programming such an engine from scratch, could at over a year, if they will ever do it. (microsoft's directX 11 is supposed to add GPU based physics engine... NOT physX, just physics.)

The G200 will supposedly have twice the shader power of the 9800GTX.
The 9800GTX has 128SP, the G200 has 240SP that are supposedly 50% more efficient, but their clockrate is unknown. The rumer is 2 times the shader power.

 

[BTRY B 529th FA BN]

I'm pretty sure its Stream Processors, not steam processors

EDIT: didn't mean to thread crap

 

[schneiderguy]

Originally posted by: ajaidevsingh
Ok from what i know steam processors are used by Nvidia and also by ATI for physx.

Now since the ATI cards say 3870 have more steam processors does that mean that enabled physx will impact lower on the ATI card FPS than on a Nvidia 9800 GTX which has only 128 steam processors as to 320 for the 3870??

If you all say that 9800 gtx will never support physx they dd make a promise that which ever card supports CUDA will support physx!!

If these seems too old a products one can take the example of 4870 "480 SPs" and GTX 280 "240 SPs"??

The 3870 only has 64 stream processors that can do up to 5 calculations per clock. ATI just advertises it as 320 sp to make it sound more powerful. r600 sp utlization. Average utilization of the stream processors in the 3870 is about 3.5 instructions/clock, which is only about 220 "stream processors" in ATI language.

The number of stream processors in g80/r600 isnt really directly comparable

 

[aka1nas]

Originally posted by: ajaidevsingh
Ok from what i know steam processors are used by Nvidia and also by ATI for physx.

Now since the ATI cards say 3870 have more steam processors does that mean that enabled physx will impact lower on the ATI card FPS than on a Nvidia 9800 GTX which has only 128 steam processors as to 320 for the 3870??

If you all say that 9800 gtx will never support physx they dd make a promise that which ever card supports CUDA will support physx!!

If these seems too old a products one can take the example of 4870 "480 SPs" and GTX 280 "240 SPs"??

CUDA is an Nvidia-specific API that interacts with their own hardware in a more direct manner than APIs such as D3D do. CUDA will not run on AMD/ATI hardware. AMD would need to write their own PhysX wrapper to run it on their cards.

 

[ajaidevsingh]

I did know that ATI SP's were not as powerfull as compared to the Nvidia SP's but i was thinking that decentralization would help in terms of physx...

aka1nas@ Nvidia has made CUDA a base so others can make similar physx tech. but its owned by Nvidia!!

 

[Munky]

"Stream processors" is just a marketing term for the unified shaders present in modern video cards. Nvidia and Ati/AMD follow different architectures and design principles in their gpu's, so just comparing the number of "stream processors" is not an accurate indication of actual performance. However, both camps have been pushing the "physics on the gpu" marketing campaign for a number of years now, and we've yet to see any actual results.

One thing is for sure - all these shaders on the gpu will be hard at work rendering modern games, and don't expect physics on the gpu to be a realistic option, unless you buy a second video card just for physics, AND more importantly, the game developers will code the game in such a way to support physics on the gpu. In regard to my second point, the devs will be much more likely to support such a feature if it's done via a unified API, such as some future version of DirectX, rather than relying on proprietary technology that's exclusive to one gpu vendor.

 

[aka1nas]

Originally posted by: ajaidevsingh

aka1nas@ Nvidia has made CUDA a base so others can make similar physx tech. but its owned by Nvidia!!

No, they didn't. Nvidia has pledged to keep PhysX an open API so other companies can implement wrappers for it. There's nothing in CUDA for them to "Open up", it's basically just a bunch of C libraries that compile to Nvidia GPU machine code.

The best(but still flawed) analogy I could give would be that it's like picking on Intel for having an x86 assembler that doesn't create PowerPC machine code.

 

[SunnyD]

Why would ATI support PhysX anyway when they already support Havok?

 

[Grinja]

I wonder if you could add a low end (or older) vid card to do the physics. For example when you upgrade to GTX 280 or something in future but keep your existing 9600GT for physix (not necessarily SLI)

It would be nice to be able to 're-use' old tech.

 

[Munky]

Originally posted by: Grinja
I wonder if you could add a low end (or older) vid card to do the physics. For example when you upgrade to GTX 280 or something in future but keep your existing 9600GT for physix (not necessarily SLI)

It would be nice to be able to 're-use' old tech.

What, and keep you from having to buy a second new video card? No, that would make far too much sence...

But seriously, that might work as long as the video card is not too slow or too old. The practical minimum requirements would probably be a DX10 card, say a 8600gt or better. The older DX9 cards don't offer as much flexibility for general purpose computing, and a bottom of the barrel card will probably slower than a cpu for physics. For example, in my own testing, I found that a 7300gs is actually slower in number crunching than a moderately-clocked C2D cpu.

 

[taltamir]

DX9 cards can't do CUDA, so it isn't even an option... when the drivers are out all DX10 cards will be able to do physX.

It does make sense to do physics on the GPU... i imagine games will look a lot better if you substitute AA and AF for physics calculations, and so on.

And the G200 is setting to more then double the shader power of the 9800GTX. So there is plenty of juice to go around.

As for directX... DX11 will add physics on GPU...

So we are looking at physX on nvidia, Havok on ATI, and DX11 with something unified "in the future".

PhysX would also work on linux etc, so that means openGL games could choose to use physX.

 

[AzN]

Originally posted by: taltamir
ATI has WORSE SP performance, not better. while they have more SP units, their units are weaker individually, and are at a much lower clock speed then nVidias.

ATI also will not have physX for a long while...

nVidia bought AGEIA, which owned physX, they opened the standard so that ATI CAN develop their own physX on GPU, but nvidia already has an EXISTING physX engine from AEGIA that has been programmed for more then a year. They ported it to CUDA (C code on nvidia's SP) and will supposedly ship it with the G200, on the 18th of June. ATI will take a long while to come along with it. It will work with every CUDA capable GPU (IE, anything from nvidia with DX10 support)

ATI has to start programming such an engine from scratch, could at over a year, if they will ever do it. (microsoft's directX 11 is supposed to add GPU based physics engine... NOT physX, just physics.)

The G200 will supposedly have twice the shader power of the 9800GTX.
The 9800GTX has 128SP, the G200 has 240SP that are supposedly 50% more efficient, but their clockrate is unknown. The rumer is 2 times the shader power.

Doesn't necessarily mean ATI has worse SP performance. They are different architecturally. 3870 is actually has little more power than G92 far as SP goes.

 

[ViRGE]

Originally posted by: SunnyD
Why would ATI support PhysX anyway when they already support Havok?

Havok FX was bunk, Quantum Effects was bunk, and every other physics solution currently or previously advertised is bunk. There is no GPU accelerated physics solution currently on the market.

 

[ajaidevsingh]

As i recall Havok was bought by Intel and its not open..

That leaves Nvidia's physX "Which is all ready open and others can make their own phytech upon it so that most games support a single phytech"

As taltamir pointed out DX9 cards can not support physX as non of them have SP's to work on!!

As far as ATi SP vs Nvidia SP go....ATi SP have lesser lanes as compared to Nvidia but they do have a bigger pocket....!!! So i was thinking that decentralization would help in terms of physx or physx wrapper!!

aka1nas@ Nvidia will not share any tech developed by PhysX, they will only share a same base... Its like Nvidia has allready build a fort and now telling ATi that they can use the same schematics to build a similar one!!!

 

[Fallen Kell]

Originally posted by: ajaidevsingh
Ok from what i know steam processors are used by Nvidia and also by ATI for physx.

Now since the ATI cards say 3870 have more steam processors does that mean that enabled physx will impact lower on the ATI card FPS than on a Nvidia 9800 GTX which has only 128 steam processors as to 320 for the 3870??

If you all say that 9800 gtx will never support physx they dd make a promise that which ever card supports CUDA will support physx!!

If these seems too old a products one can take the example of 4870 "480 SPs" and GTX 280 "240 SPs"??

The ATI stream processors are very floating point operation limited... What do you think physics calculations will be? You are talking about force vectors at specific angles all being calculated against the mass and friction resistances of moving the object. If you even get 1 whole number (int) operation out of 100,000 calculations, you will be lucky.

 

[taltamir]

well... that 320 SP cluster ATI has is something like 5 integer SP to ever 1 floating point SP or something like that (it differs by model from what i heard)... and integers have no room in physics calculations.

@Azn, I know that doesn't mean AMD has weaker shaders, I say that they have weaker shaders because comparing the 3870 to the G92 they do (to the best of my knowledge)...

 

[hooflung]

Originally posted by: taltamir
well... that 320 SP cluster ATI has is something like 5 integer SP to ever 1 floating point SP or something like that (it differs by model from what i heard)... and integers have no room in physics calculations.

@Azn, I know that doesn't mean AMD has weaker shaders, I say that they have weaker shaders because comparing the 3870 to the G92 they do (to the best of my knowledge)...

It is open to debate what 'weak' shaders boil down to. AMD has a VLIW tech which means that they are very simple compared to how they work directly compared to Nvidia's version. However, Nvidia has built theirs with more complexity and thus require far, far more space on the die of the GPU for less units.

We have seen a lot of speculation on how AMD will or will not scale. But they haven't really done anything out of the ordinary through their product lines. The 3650 has less tech and less memory bus. The 3850 and 70 just have 'more,' while the X2 is just adding single slot SLI.

Where as Nvidia is all over the place trying to figure out the sweet spot for their die realestate. The 8800GS vs the 8800GT vs the 9600GT vs <this months variation of the G80/92> We simply do not know if the Texture units in AMD tech is the flaw or if its Shader power. We do know that bandwidth is real deal in 8/9 series Nvidia tech. The 9600GT has less GPU power over a 8800GS but still holds its pace well.

I just wonder if people are going to be able to use AMD cards as primary GPUs on Crossfire/SLI boards and an 8 series card for physics all at the same time with no head aches.

 

[ajaidevsingh]

ATi's Very long instruction SP units are the very thing that may help the calculation of phy because i have seen phy calculation on a PS3 for a game demo on linux. The phy was done with the SPE's and as they are not ultra complex thus i think even simple VLIW's should do!!!

 

[Munky]

I can't believe the amount of misinformation I'm seeing in this thread. First, Ati used floating point data formats throughout the shader core, same as Nvidia. Secondly, Ati shaders are not "weaker" than Nvidia's, it's just that getting the most utilization of them is more complicated than on Nvidia hardware. In addition, neither camp has any "stream processors" in the technical sense, because these ALU's are grouped by threads, and Ati can no more execute 320 separate instructions simultaneously than Nvidia can execute 128.

The rv670, for example, has 4 groups of 16 shaders, with each shader being a 5-component superscalar vector ALU. The g80, on the other hand, has 8 groups of 16 scalar shaders. Moreover, Nvidia serializes multi-component vector instructions, so a 4-component vector would execute on the same ALU, 1 component at a time. This makes getting optimal utilization of the shader core easier than trying to pack vector instructions in parallel, but this method has its own pitfalls, such as the possibility of bubbles occurring in the shader pipeline, where certain ALU's remain unused at a given time.

The bottom line is that the rv670 has more than enough raw shader power to compete with Nvidia's current gpu's, but in typical game scenarios, there are other weaknesses of the Ati architecture which put it behind the competition.

 

[ajaidevsingh]

munky@ ppl must mean that ATi SP implementation is much too complex than nvidia..Thus in a matter of speaking they are weaker "At least i was thinking that they are" As for the 4 group and 8 groups packs i have allready refered to them in am lay man lang. in my older post!!

From what i could make out ATi's SP's can processes phy. much faster mainly because of the decentralized structure!!