google 'TPU'

[dobkeratops]

http://www.pcworld.com/article/3072256/ ... l#comments

any thoughts on what this might be?
not as specific as a neural-processing-unit.. could it be something ephiphany-esque? (they say it isn't a DSP, then again I don't think the epiphany is strictly classed as a DSP either) . Or perhaps its something more specific to convolutions, like 'NeuFlow' ?

They seem to suggest it focussed on low precision arithmetic rather than 32bit float ?

I always thought the epiphany idea generally would be good for 'tensorflow' and similar frameworks (as CELL would also have been, if scaled up to todays' processes)

[sebraa]

[Olaf]

I have doubts in this.

GPU's was also supposed to save the world but then it turns out that it excelled on some specific niche problems.
I think Parallella is the best solutions for the future in the sense that you have a combined 16 core RISC together with a FGPA.

In theory you could design your special custom logic/calculations that works at hardware speed and feeds it back to the RISC processors.

I am not far enough into Parallella but I can imagine that you could program a Encryption logic in the FGPA and order the RISC core to use that for calculation and get back the result. Now imagine that every RISC core has his own dedicated FGPA hardware encryption logic.

I recall a time where we wrote self modifying code to alter the assembly code on the fly optimized for that specific function. In our case we could use the FGPA to extend the RISC processor functionality. (I do not know if Parallella could do this) But imagine that in FGPA you create a complex formula in hardware: a = a^2 + 3b + c^4. Then on one or 2 clock pules you get the immediate result.

[dobkeratops]