Where this gets interesting is Physics on GPU. Both Nvidia and AMD/ATi had projects underway using HavokFX, a version of the Havok physics engine accelerated by GPU computations. Several doom and gloom analysts are worried that Intel's purchase of Havok will make it difficult for Nvidia and AMD/ATi to pursue Physics on GPU, and that Intel has already locked in the market for their upcoming add-in graphics cards. If you want Physics on GPU, you'll have to go with Intel. If you want the best Havok physics... you'll have to buy Intel.
Well, I've already been over Intel's chances of making a decent graphics card before. Graphics have been ancillary to Intel for literal years, and I don't think Intel has the development staff or engineers to compete with AMD/ATi and Nvidia.
That, however, isn't the gaping hole in the idea that Intel owns physics.
That... is the gaping whole in the idea that Intel owns physics. AMD is opening up the specifications for how their graphics cards actually work. Now, if you know how a graphics card works, and if you have access to the BIOS/Firmware, which in AMD's case developers do have access, what's to stop using the Graphics card for items other than Graphics?
As I see it, Intel's purchase of Havok was in direct retaliation for AMD opening up the specifications and BIOS for AMD graphics cards. Intel sees the scenario I am about to lay out and did what they could to upset it.
With the specifications for the graphics card in hand, what is to stop AGEIA from optimizing their Physics Software Developer Kit to leverage AMD/ATi hardware? Presuming that Nvidia wakes up and opens their specifications and BIOS as well... what's to stop AGEIA from leveraging Nvidia hardware?
The following is something that I sent ATi last year... in 2006. This was sent before the merger between AMD and ATi was announced... now... after the merger and after the opening up of the specifications and the BIOS... these questions take on a whole new light.
been reading through the reports from Computex about ATi and the Asymmetric Physics Processing and I was wondering... In the Hypertransport Specification 3.0, part of the specification dictates an external cable link which allows for boxes outside the computer to connect directly into the Hypertransport internal links. Also, AMD recently announced Torrenza, an internal connection technology to allow internal devices direct access to Hypertransport links on the motherboard itself.
Some of the rumblings I've seen from Torrenza specifically involved AGEIA using the Torrenza platform to connect their physics processor to the system. Would it be possible to utilize a re-purposed R5xx GPU in the same manner and directly connect the GPU right into Hypertransport for direct access to the system? If that is possible, would ATi also consider utilizing re-purposed GPU's and selling them in external boxes using HyperTransport Link cables? While I realize that external HyperTransport interfaces are not yet in production, I also must note that motherboards using 3 PCI x16 slots are not very widespread either, so both types of connections have time to grow.
The argument I would make to manufacturers in reference to including an external HyperTransport cable link is that most users are not comfortable with opening their computer up and adding components. While I have little doubt that Asymmetric Physics Processing with Crossfire would sell very well in the high end gaming segment, thats only, what? 5%? Maybe less? of the available market that is actually sold to? I would think that these users, if given the opportunity to buy an external box that plugs into the computer and immediately creates a performance difference, would probably buy the product. Re-purposing older x1600 gpu's, as mentioned in the tech brief, to external HyperTransport boxes might be a revenue stream that I don't think I've seen any manufacturer or hardware developer actively look at for the home market. I think this would also be of interest to the Hometheater market, as well as student markets, which may not have the physical space to put a full tower. Speaking for myself, I know there's no way I could fit 2 double slot graphics cards and a 3rd card into a MicroATX case.
I've also seen some rumblings going around about using GPU's to process sound commands, which leads me to pose this question: Presuming that GPU enabled physics take off and are must-haves for the mass user. Lets say that both AGEIA and Havok get behind GPU enabled physics. Lets say that in one years time everybody is using GPU enabled physics or needs to use GPU enabled physics, or have a dedicated PPU inside the case. AGEIA partially tackled the problem of getting hardware physics to consumers by releasing on the aging PCI slot. Granted, it's slow, really, really, really slow, but just about every computer built and shipped today has a standard PCI slot. Has ATi considered allowing manufacturers like PowerColor and Sapphire to release PCI versions of current cards that could be re-purposed for physics use over the PCI bus? Has ATi considered building versions of Radeon Xpress that would support an AGP slot for the expressed purpose of using the AGP card to accelerate physics?
Are there also considerations to release Physics ONLY cards that strip items like the RAMDACS, display ports, and other items from the card needed for the visual display itself? In such a scenario the physics only cards could sell for less than their visually enabled cousins. That would help clear GPU inventory, and GPU's that didn't make the cut to be used in graphics could be binned to be put into a physics only card. If a physics only card has been considered, going back to the question about processing sound commands, has ATi considered releasing cards that have two re-purposed GPU units? For example, an x1600 unit stripped of visual display for physics, and an x300 GPU for processing sound? Or have there been considerations to add Sound support into Asymmetric Physics Processing and have the Physics GPU also handle sound processing?
While I understand that most of these questions are probably going to be met with the typical "I can't comment on that at this time," part of me hopes that at least some of these items have already been circulated through ATi.
Okay... so what would prevent AGEIA from using the GPU specifications to port their physics SDK? What would prevent PowerColor, Sapphire, XFX, Leadtek, or any other hardware vendor from re-purposing older chips onto stripped cards just for physics acceleration?
No. Intel doesn't own physics, and since I know that ATi has had this brought up before, I do not doubt that somebody in AMD/ATi is already thinking about this.