BFloat16 Support About To Land Within LLVM

phoronix

Administrator

Join Date: Jan 2007

Posts: 64832
- Share
- Tweet
#1

BFloat16 Support About To Land Within LLVM

13 May 2020, 08:10 AM

Phoronix: BFloat16 Support About To Land Within LLVM

The LLVM compiler stack is about to merge its support for the BFloat16 floating-point format, including the BF16 C language support...

BFloat16 Support About To Land Within LLVM - Phoronix

http://www.phoronix.com/scan.php?page=news_item&px=BFloat16-For-LLVM-Landing-Soon

Phoronix, Linux Hardware Reviews, Linux hardware benchmarks, Linux server benchmarks, Linux benchmarking, Desktop Linux, Linux performance, Open Source graphics, Linux How To, Ubuntu benchmarks, Ubuntu hardware, Phoronix Test Suite
Tags: None
coder

Senior Member

Join Date: Nov 2014

Posts: 7848
- Share
- Tweet
#2

13 May 2020, 12:40 PM

Good for deep learning, and very little else.

The half-float format, specified in IEEE 754-2008 and supported by Intel iGPUs since Broadwell, by AMD since Vega, and Nvidia since GP100 (Pascal) is a better all-around compromise between range and precision. But, it's not as good for deep learning, and not as efficient to implement in hardware.

There are also conversion instructions between fp32 and fp16 in both Ivy Bridge and later CPUs (i.e. F16C) and in AVX-512's VCVT group (VCVTPH2PS, VCVTPS2PH).

Last edited by coder; 13 May 2020, 12:50 PM.
Comment

Announcement