By: Sylvain Collange (firstname.lastname.delete@this.gmail.com), August 16, 2011 3:04 am
Room: Moderated Discussions
Mark Roulo (nothanks@xxx.com) on 8/15/11 wrote:
---------------------------
>Do we expect x86 numerics to be *WORSE* than GPU numerics?
Oh yes, definitely.
For double precision, OpenCL mandates support of the FMA, subnormals and the 4 basic rounding attributes.
Not that this cannot be implemented on x86 (FMA can be emulated in software, subnormals are typically handled by microcode, and rounding modes can be switched at the expense of a pipeline flush).
But it is painfully slow, enough to discourage the use of such features in any serious computation.
By contrast, all current high-end GPUs have full-speed hardware support for these numerical features.
"GPUs have bad numerical behavior" is now just an urban myth.
---------------------------
>Do we expect x86 numerics to be *WORSE* than GPU numerics?
Oh yes, definitely.
For double precision, OpenCL mandates support of the FMA, subnormals and the 4 basic rounding attributes.
Not that this cannot be implemented on x86 (FMA can be emulated in software, subnormals are typically handled by microcode, and rounding modes can be switched at the expense of a pipeline flush).
But it is painfully slow, enough to discourage the use of such features in any serious computation.
By contrast, all current high-end GPUs have full-speed hardware support for these numerical features.
"GPUs have bad numerical behavior" is now just an urban myth.
