how much could an E5 fit onchip - vision nets

[dobkeratops]

https://arxiv.org/abs/1510.00149 ... claims to get the weights for alexnet into 7mb

so the E5 has 1024 x 64kb = 64mb on chip;
I imagine most of that will go on code and workspace.

But perhaps it's reasonable to expect that you could set aside 8mb to keep compressed weights permanently onchip? ( i suppose you'd scatter that as 12.5% of each core)

I imagine a few of the cores to one side could keep program code for decompressing weights, whilst most of them are churning away convoluting. I think you would only have to decompress infrequently, because once decompressed the kernels are applied across the whole image.

Could this be a killer app for the architecture - compared to dedicated neuromorphic chips, you have the flexibility to implement any compression scheme you can come up with

[jar]

I like your thoughts. I think you may not be estimating just how little space is needed for programs if you do it right...16 kB for many kernels may be generous. When you go larger, software instruction paging techniques can be used usually with minimal performance impact.

I've got a paper coming up in the next month or so that you, in particular, might enjoy reading...related to an idea you had a while ago.

And a funny one for everyone in the next couple weeks.

[dobkeratops]