Mark Cerny has already said it would do the processing for hundreds of audio streams
Mark Cerny was talking about decoding hundreds of MP3 streams. It's probably just a decoder chip, like XMA on the 360. That's the biggest bang for your buck, very cheap to do in hardware, and expensive in software. (I say "just", but it would be almost a requirement for a machine without SPEs to be able to decode hundreds of MP3 channels simultaneously.)
And I really wish people would stop throwing around the GFLOP number for SHAPE its practically meaningless. Its a fixed function device, you can't do anything else with the flops and we have no idea how the number was calculated, nearly all of its cores are integer after all.
To do the majority of the normal audio functions of SHAPE (i.e. none of the kinect stuff) would probably take less then a single jaguar core I'm led to believe.
You're led to believe by whom? And while FLOPS is not accurate, it is useful if you're tying to gauge how much CPU you would be saving by using the fixed function hardware. On a CPU, you would have to use the vector instructions to even approach it, and they're listed in FLOPS. Also, 24bit integer audio is bitwise identical to 32bit float audio, as long as you have a few extra bits available for mixing (which SHAPE does - it uses 28bit internally) So they are comparable. You write a float version of your fixed function algorithm, measure it, and that's how much CPU you're saving with the hardware.
And no, a single jaguar core could not even begin to process as much as the SHAPE block, not counting the AVPs and ASP and ACP, those cores are good, but not that powerful, it's the fixed function stuff that provides most of the systems advantages.
I've seen no one say that a 8 core CPU couldn't reach the same number of voices after all SHAPE does much more then voices, i could believe maybe that a 8 core CPU couldn't do _ALL_ of shape, but I have to see anyone provide any evidence that a CPU couldn't do all of the _NORMAL_ audio processing.
Define "normal" audio processing? Just reading streams off disc, mixing them, and outputting them is not what a normal audio engine does. If that were so, no one would need FMOD or WWISE.
Audio is cheap, the voice + kinect stuff isn't.
Actually, in an average 360 game using voice commands, the "normal" audio processing takes up to 10 times more CPU than the kinect voice pipeline.
I don't expect very many people to make their decision on which version to purchase based off an audio spec and if someone truly thinks that is relevant in the mind of consumers should show where audio has successfully marketed before as key differentiator.
Excellent point, and probably why the audio hardware has not been mentioned publicly except by the FMOD guys.
Anyone got any idea how this shape chip compares to a EMU20K1 dsp (in the xfi)
It is significantly more powerful, by about an order of magnitude.