quote: Original post by Sandman
[offtopic techie rant]
AI accelerators are a pretty stupid idea if you think about it. Graphics accelerators work because output to the screen is a one way operation - the game tells the card what to draw and the card goes away and does it. The game itself doesnt care what the outcome is. Advances like TnL and programmable shaders enable more stuff to be done on the card without the app needing to know about it. Note that if you ever read stuff back off the card, you get massive slowdown because the upload bandwidth across the agp/pci bus is so crap. Also note that if you run a demo with just graphics (no AI/physics or anything) youll find that hardware TnL is actually slower than software TnL - you only get a speedup in games because offloading it to the card gives the CPU more cycles to do other stuff like AI and physics etc, and ultimately less data needs to be sent across the bus each frame.
AI on the other hand, is a two way thing. The game needs to give the AI engine inputs and then needs to read the outputs. There is absolutely no point in a AI card because the bandwidth problems, memory latencies, download and upload speeds, etc. are going to prove far slower than just doing it in software. More hassle than it is worth, at least with conventional computer architectures. The best we can really hope for is some specific opcodes coded into CPU''s to make standard AI functions easier, a sort of AI Now! feature in cpu''s.
[/offtopic techie rant]
OK, now it is my turn for the offtopic techie rant... I don''t see this "AI Card" as something along the lines of an Accellerator. I see it as a parallel processor. It basically runs the same way that a normal CPU should function running a threaded function on its devoted AI place. This card wouldn''t necessarily be required for ONLY AI, and could be used to parallely speed up using a threading routine for Any function.
Unfortunately, as I am well aware, there would be no support for allowing threading, but IIRC a few years ago there was no Autodetect for Cards and PnP didn''t exist, so I expect that threading for functions onto parallel CPU''s could be possible from a theoretical standpoint.
The drawback I can see immediately is that it would require some devoted RAM that would have to be removed from access by the standard CPU, or it would have to be onboard the card...
But I would still definitely like to see somebody look into such an idea. It may not have the one way functionality that 3D accs do... but it should never be designed in the same way.
But it is all speculation and maybe I am just a dreamer
-Chris Bennett of Dwarfsoft - The future of RPGs Thanks to all the goblins in the GDCorner niche