Havok's response to the Tim Sweeney Interview.

[Superfly]

Havok is right. I strongly believe Ageia PhysX to be just a piece of marketing nonsense. Nobody needs PhysX HW to calculate physics. We have dualcore CPUs already, we will have quadcore CPUs soon, we have SM3.0 capable graphic cards. Havok FX can use all that. This is the right approach, not adding PhysX decelerator to your PC, occupying additional slot, and being noisy.

Even if Ageia PhysX would cost $1 I would not buy it, as I do not want to support wrong concepts...

Right on the money, lets hope the game dev's agree with us too!!!!

[Oscar]

For consumers looking to spend additional $’s to accelerate physics in their games, why not spend money to purchase a new GPU for graphics, and relegate last year’s model to doing the physics via Havok FX?Doesn't SLI require identical video cards?That's what I keep bringing up in other discusions on this topic. People seem to be under the impression that they can have a 7900GTX and and 6200 in sli and use the 6200 for physics. Sorry, thats not going to work. And even if it did work the 7900/6200 config well be cheaper, but it not going to touch the performance of a 7900/PPU config. Sorry again.Maybe I missed it but where did they say it will require SLI?I think ATI is supposed to allow that. The one article I saw comparing the two said Nvidia saw value in providing that but it didn't say whether or not they'd support it.I wasn't actually thinking SLi, per say...

In reality I know very little about this, but my understanding of the matter is that it's not exactly SLi, but rather using the second PCIe port for another GPU which would be used for physics processing.

.... Buuuut, I'm guessing this would need a new motherboard BIOS or something to be able to handle the second GPU as some sort of a co-processor.

Enlighten me, if you please :D
Can one install and run two different PCI-E x16 video cards on a SLI/Crossfire motherboard at the same time without enabling SLI/Crossfire?

[HighTest]

Havok is right. I strongly believe Ageia PhysX to be just a piece of marketing nonsense. Nobody needs PhysX HW to calculate physics. We have dualcore CPUs already, we will have quadcore CPUs soon, we have SM3.0 capable graphic cards. Havok FX can use all that. This is the right approach, not adding PhysX decelerator to your PC, occupying additional slot, and being noisy.

Even if Ageia PhysX would cost $1 I would not buy it, as I do not want to support wrong concepts...


I understand your point of view and agree. In fact I'll take it a step further. Why The heck do we have 3D GPU's? CPU's can render any of the scenes as shown by the CPU tests in all the Futuremark 3Dmark products.

Can do and performance are two separate issues. I'd never spend $1 to improve the "performance" of what my CPU can do. Geeze, WTH did I ever upgrade from my original Pentium 133 computer? With sufficient memory it'd still have run Windows XP.

All this debate is because not one single title has been shipped yet that truely shows "interactive gamplay effecting physics". I'll conceed that "eye candy visual effects" are not anything ATI nor nVIDIA cant do fantastically without an Aegia PPU. Once compelling titles ship, then the tune will change.

I'm sick of Aegias, ATI's, nVIDIA's and Havoks hype until then.

See my http://www.hardforum.com/showpost.php?p=1029530218&postcount=48 post on HardOCP's forum that actually goes to Havok (straight from the horses mouth) and breaks down the reality of GPU physics. You'll note that outside the features specifically mentioned in the post above, everything else is CPU calculated.

Besides, is Havok FX the only real use of SM 3.0 GPU hardware to date? What's the innovation here? Aegia will surely never be able to support SM 3.0 in their phsyics model?

[OWA]

Since it's kind of related, figured I'd mention the post here:

http://www.nvnews.net/vbulletin/showpost.php?p=907483&postcount=39

[XGas]

This is kind of funny: Use remaining resources to process physics effects(means eye candy only)

I think some would ask: Is our GPU "free" all the time?
LOL, I dont think so, so dual card configuration is t3h pwnage way to go...... like all of us have enuf money to buy dual card, but ATI enables any combo of SM 3.0 card to able to run on their API, not Havok FX, right? Maybe not, its a guess, but I think nVidia would not be stupid enuf to want you to but two identicle cards to enable physics on one card anyway, from what I have read, Havoc did not mention ATI can have one e.g. X1300, and one X1900 combo, ATI just says they can, plus they have their own API for the physics FX via GPU, they once sent the SM 2.0B API of their GPU to Ubisoft to enable some effects that than be done on a SM 3.0 for the game SplinterCell: Chaos Theory, I think some people mistaken the Havoc FX for their(ATI's) own Physics FX API.


Ok now, lets take a look into the CPU's standard optimization thats been used for more than a decade: SSE.
If my memory serves me well, the SSE optimization reads from the procs cache, then decides the best way of doing the datas then rearranges them in order for better processing, so, if the datas are messy(e.g. physics processing)with this optimization, the CPU can process the data faster than without it, blah blah blah. But thats one of the features.

Then we look at the GPU, does the GPU have this kind of optimization?Not really. I think the ATI Radeon X1900XTX has this pipeline feature that ensures that every pixel and vertex processer works every clock, but the performance differents compare to a 7900GTX is not very much, but still quite good. Even thou the X1900 has the same ROP amount as the 7900GTX, the pixel pipeline is for the 1900XTX is "only" 16, compared to the 7900GTX's 24(or is it 21? Can't seem to remember), since the 1900XTX has this pipeline feature as mentioned above, it ensures that every pipeline and pixel/vertex proc is utilize in every clock, thats why it has the advantage of better physics effects processing.

[Soylent]

Ok now, lets take a look into the CPU's standard optimization thats been used for more than a decade: SSE.
If my memory serves me well, the SSE optimization reads from the procs cache, then decides the best way of doing the datas then rearranges them in order for better processing, so, if the datas are messy(e.g. physics processing)with this optimization, the CPU can process the data faster than without it, blah blah blah. But thats one of the features.

Streaming SIMD(Single Instruction Multiple Data) extension. It's just a hack for speeding up vector math like dot products and matrix multiplications; something that was a big deficiency for CPUs before SSE. In a CPU environment it makes sense to waste huge amounts of transistors to hide latency and similar endevours to increase throughput from single units rather than just slap 10 SIMD units on there. Any optimizations to SIMD instructions on a CPU are very likely to make GPUs perform worse because they don't need all that latency hiding.

Then we look at the GPU, does the GPU have this kind of optimization?Not really.

They don't WANT that kind of optimization! Vector math is pretty much ALL they do and they aren't dealing with terribly latency sensitive tasks so they can just keep streaming data through instead of relying on oodles of cache, OoOE exection, out of order load/stores, register renaming and other ugly tricks that soak up immense amounts of transistors without benefitting them. Instead of one or two very optimized SIMD units, you get a huge amount of pixel and vertex shaders. When it comes to pure brute force vector math(with low latency sensitivity) GPUs are orders of magnitude better than CPUs.

Even thou the X1900 has the same ROP amount as the 7900GTX, the pixel pipeline is for the 1900XTX is "only" 16

No. It uses 48 PS units, and they're what is of interest for physics.

compared to the 7900GTX's 24(or is it 21? Can't seem to remember)

24 shaders but they aren't directly comparable to X1900(they're more complex).

The X1900 is better suited to more general workloads. But look where that got them. It's barely faster than a 7900 with a much larger die area and higher heat output. In fact, it's one of the biggest complaints people have about x1900. You need to be throwing unusual workloads at the x1900 to see the merits of having an architecture like that. You don't want GPUs to be general unless your expected workload is general. Going too quickly towards very general GPUs is a bad thing, because all the consumers will see is hot running, expensive hardware that runs their typical 3d games slowly. Do they care that it might be good at raytracing, AI or physics? Not really if there's a HUGE disparity between their competitors in typical applications.