Tuesday, March 15, 2005

Possible synergy ahead?

Cell, the new processor architecture co-developed by IBM, Sony and Toshiba, looked to me at first glance as the next Crusoe, Emotion Engine or Transputer - past harbingers of revolutions that never were. Uniting on a single piece of silicon a mid-grade PowerPC processor with eight simpler flexible co-processors is an interesting idea, and might even be useful for some purposes, but I didn't see any real general applicability for the design for most of what we do or think we should be doing near-term with computers. Yes, I know, the Cells are supposed to open up some kind of magic window for sharing computing resources with everyone else out there, but that kind of thing takes considerable coordination and cooperation, and I don't see any of the needed groundwork in OS or application-level support for the hype. And what would we really do with the eight co-processors in most computing applications? I foresaw lots of idle time, and idle transistors are a waste of resources.

Two unexpected (to me, anyway) recent announcements have me wondering about my initial reaction, though. First, Ageia announces the PhysX 'physics processor', designed to off-load from the main processor all of the calculations required for solid and fluid behavior in simulated environments, like video games. And they were even smart enough to pick an established API (NovodeX) to access these features, so developers can write to the API and the application will run with or without the PhysX chip, though at vastly different speeds. Needless to say, electronic game developers are very interested. And now a German University has revealed a real-time ray tracing chip called SaarCOR for rendering graphics in a completely different way than the rasterization methods used in essentially all modern graphics processors. If the reports are to be believed, and they look pretty convincing, this technology is going to force a complete reevaluation of imaging engines. Like PhysX, the SaarCOR chip is accessed via an open API - OpenRT. So if you write for OpenRT, you can use or not use the chip and still function, though if you go SaarCOR-less, you'd better have a large network of very fast conventional processors at your back.

So, where does this lead? Maybe nowhere. But, maybe, Cell is perfectly positioned to take advantage of two coming transformational technologies. Those eight co-processors, 'Synergistic Processor Elements', according to the Cell group, sure look like the kind of resources an enterprising developer could leverage to do some NovodeX and OpenRT work, paving the way for Cell2 with on-die PhysX and SaarCOR SPEs in addition to the general purpose ones. No other architecture currently in production could match Cell's advantage in taking this approach, because no other design has so much spare computing power designed specifically for the kind of highly regular, parallel processing necessary for real-time realistic physics and ray-tracing.

But what do I know?


Post a Comment

<< Home