News for hackers. Stuff that matters.
Re:Nvidia Pascal GPU Architecture to Provide 10X Speedup fo... by modeless on Monday May 18, 2015 @03:16AM (#9563150)
Processing half float data on the CPU will be tricky without hardware support. There's plenty of CPU processing needed for deep learning, for capturing and preparing datasets.

Half precision floating point is really something that should have been in hardware a long time ago. There are other applications besides deep learning that could benefit from the dynamic range of floating point, but don't need 32 bits. For example, imaging and audio.

Re:Nvidia Pascal GPU Architecture to Provide 10X Speedup fo... by minthd on Monday May 18, 2015 @08:15AM (#9563950)
Logarithmic number systems could offer even greater improvement(100x over gpu), if you're willing to live with more errors and/or invest some time in fixing them:

http://www.gwern.net/docs/2010-bates.pdf

Re:Nvidia Pascal GPU Architecture to Provide 10X Speedup fo... by DennisP on Monday May 18, 2015 @09:05AM (#9564162)
Seems like a great idea, and the last couple pages sound like they were making good progress toward commercializing it. But the document's from 2010. I wonder if they're still working on it.
Re:Nvidia Pascal GPU Architecture to Provide 10X Speedup fo... by minthd on Monday May 18, 2015 @09:41AM (#9564330)
I know it was implemented in some military UAV, and achieved great results. So maybe the military took this.

But it still doesn't mean it can't be implemented outside the u.s. or in FPGA's.

Even a simulation of this in software could be really interesting.

Re:Nvidia Pascal GPU Architecture to Provide 10X Speedup fo... by listic on Monday May 18, 2015 @06:20AM (#9563586)
It looks like that's what F16C instruction set was made for http://en.wikipedia.org/wiki/F16C
Re:Nvidia Pascal GPU Architecture to Provide 10X Speedup fo... by Narishma on Monday May 18, 2015 @07:45AM (#9563846)
That looks to be just for converting to and from half-precision floats.
Re:Nvidia Pascal GPU Architecture to Provide 10X Speedup fo... by bitL on Monday May 18, 2015 @05:44AM (#9563506)
For both imaging and audio even 64 bit doubles aren't sufficient if you use many filters together.
Re:Nvidia Pascal GPU Architecture to Provide 10X Speedup fo... by raverbashing on Monday May 18, 2015 @07:46AM (#9563852)
[Citation needed]

It might be not enough if you have a lot of filters and doesn't want to care about their ordering/gains etc

If you have something like 16bit exponent and 48bit mantissa (64-bit) that's absurd