So what is XMOS good for?

[Karlos]

I don't really get the point of the XMOS chip, if I'm honest. All I know is that it's a low latency device suitable for handling IO events. In terms of processing power, what can it do that a few more tens of megahertz CPU power wouldn't?

For me, a far more pertinent set of X1000 questions would be:

1) Which CPU will be used and at what speed?
2) Why 4 DDR slots? That suggests a wide architecture. Assuming 64-bit operation is out of the question on the grounds of backwards compatibility, Is a 32-bit PAE style kernel planned?
3) Given the PCIe, which display cards will be supported? (I know RadeonHD drivers are being developed for OS4.x)

[Karlos]

Let's not forget that the XMOS has the Xorro slot available. What about a Xorro board with, say 8 sockets for XMOS chips (so they can be added as required) and a socket linking to another box with more Xorro slots each with a board with 8 sockets...etc...

This is an untried chip for desktop computing so I refuse to write it off quite yet.

8 XMOS chips isn't even going to scratch the surface of what a modern GPU can do in terms of parallel processing, even using basic GLSlang (let alone CUDA, Stream or OpenCL).

[Karlos]

Agreed. However, GPUs can't read back from the destination buffer, so they're not good for Infinite Impulse Response (IIR) filters. It's possible to do with a GPU via tricks (with sizable extra overhead), but this is something that a set of stream processors might have an edge. Of course, you're still contending with a brute of a processor, so 8 XMOS chips still won't do.

IIR filters are used in audio processing, and Gaussian blurs of any size can be done efficiently using IIR filters, whereas a Finite Impulse Response (FIR) based Gaussian blur slows down exponentially with blur size.

Hans

Actually, the CUDA n-body solver that got me into GPU coding initially uses a double buffer where the destination data becomes the source data for the next iteration. It's pretty effective.