Parallella Community

sorry, sebraa put it in a way that could be misunderstood. to keep the discussion short:

no, it is not feasible to put epiphany on pcie with a display-port. theoretically it could be possible to increase bandwidth to the requirements of display-port, and to add epiphany chips up to a max memory-size of 4k epiphany cores times 32k memory each=128M, that's likely about 50 bytes per pixel in whatever highest resolutions. but as I said, I have no competence to make such a suggestion, even less to flesh out the details how this could be implemented in hardware. the point is as sebraa tried to say, any kind of graphics card has more memory available, and a caching-system which basically is able to store the whole screen at a place where it's available on each of the processing-units. not to mention all the textures and 3d-models.

at first I too thought epiphany could be used to implement a voxel-based grphics-card, but now I realize it's more fit for a physics-engine at best. the point is that simd chips like in gpu are better fit for whatever graphics application, even in terms of power-consumption they're on par with epiphany when all cores are in use. and also for basic acceleration in a pc, there are better solutions, and nobody will buy them either since current pcie is too slow in comparison to whatever co-processors built into the cpu-chip. if you want an alternative product to the parallella based on epiphany chips, it needs to be a wearable or otherwise mobile device, where the low-power-profile comes into play...

again, just look at the numbers, you'll see the effort put into an epiphany-card is in no relation to what you could gain out of it. a lot of things are possible, being feasible is a rarity among the stuff that actually would look cool...

And again, you miss the point by a mile, piotr5. Why don't you get the hint? *sigh*

Putting 256 Epiphany chips on a single PCIe card is stupid.
The result would take some 500 W of power, would not fit in any PC case and would be virtually unprogrammable.

Putting 4096 Epiphany cores in a single system is stupid as well (apart from some scalability reasonings).
The result would suffer from the limited external bandwidth.

This sort of thing would be useful for AI workloads - training & running neural nets ?

see how Facebook are running multiple GPUs plugged into a single server board. I'm sure epiphany architecture (scaled up appropriately) could handle them better (and there are other 'dataflow' oriented chips appearing for this sort of thing)

I realise adapteva may be better to focus their time on other areas; I would personally rather see a 'parallela 2'

http://www.hotchips.org/wp-content/uplo ... ray-v4.pdf

I 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).

I disagree with your point about off chip/on chip? The architecture only distinguishes between local memory (inside the core) and remote memory (not inside core)
There is no difference (except for performance) between 4096 cores on a chip and 256 chips of 16 core chips in an array.