http://www.hotchips.org/wp-content/uplo ... ray-v4.pdfI only recently found out about this, and remembered this discussion thread..
I guess this is exactly what an epiphany PCI card would have been like if anyone ever built one.
They claim it's good at Convolutional Neural Networks, and they claim an order of magnitude energy efficiency improvement over GPUs (12 watts matching what a GPU does in 142 watts, whilst being a little faster, 14%).
Their chip seems a quite a bit more complex, and I guess they only intend to sell it in low volumes for datacentres.
I'd assume the epiphany would perform at least as well, the only unknown being the 32k/core vs their 128k (I'd have guessed you'd want to keep convolution filters on chip whilst streaming the image through). I wonder how the complexity affects it, maybe you could get more e-cores for the same price? they're showing 4x256core chips, maybe you could do 1024 e-cores on a single chip..
r.e. earlier comments that "such a thing would be unprogramable":
I think for convolutional neural nets, assigning cores to the flow between layers programatically would be conceptually easy.
The perception that it's "hard to program" comes from trying to adapt general purpose code to such a thing (which isn't really what it's for, ideally). (that was my fear from gamedev CELL experience.. completely different to this use case)
I think the scaling would work perfectly well: you would benefit from building ever larger arrays for deeper nets & more parallelism across each layer. (supposedly in deep learning, more layers increase versatility & accuracy)
I hope adapteva can get their chips into this market, it would be great to see..
I think it would be perfect for vision in robots too, where the low energy behaviour is even more important (batteries).