| |
BGonline.org Forums
Neural Nets in Hardware Prototyped
Posted By: Frank Berger In Response To: Neural Nets in Hardware Prototyped (dlevy)
Date: Friday, 19 August 2011, at 9:30 a.m.
Thanks for the link. They achieved 3.6 speed improvement about an Pentium III 800 Mhz and estimated 2 orders of magnitude with a more capabable FPGA. Estimations are always dubious and the evaluation in endgames is trivial in Othello. So they have some succes but a lot of work is still to do. Don't get me wrong: I don't say this is nonsenes, I'm not qualified to judge about things like FPGA programmin, but there is always the reality which makes problems ;)
2 years ago I spend quite some time studying GPU to speed up the evaluation. Hundreds of CPU at your disposal, High level APIs finally available...... just it don't works (yet). When I let a midclass GPU (NVidia GT240) do the vector/matrix multiplication, the CPU was faster by a factor of 2. Only if I pumped up the Matrix from 250 to 5000 the GPU was a little bit faster. Xavier made the same experience with CUDA (I tried OpenCL). I talked some years ago about that with Chrilly Donninger (father of Hydra) about this and he said that GPU programming is a lot easier than FPGA and he would do Hydra today with GPU. So I think we have a problem not fitting very well with GPU. It might be that with the next geeration Nvidia or AMD chips it might be different. It looks to me that they put quite some effort in making the GPU more versatile (again I might be wrong) but currently I think it doesn't work. (If I could only get me fingers on a Intel Knights Ferry card.... It might be a fail on the GPU market, but for our purpose it could be a perfect device. Anyone listening from Intel?)
Related to the neural chips: I'll have an eye on this, but I doubt anytime soon we'll see something an mere human (outside IBM labs) can use.
| |
BGonline.org Forums is maintained by Stick with WebBBS 5.12.