Arm Talks Up Their BFloat16 / BF16 Support For Upcoming Processors

[phoronix]

Phoronix: Arm Talks Up Their BFloat16 / BF16 Support For Upcoming Processors

While we've known Arm would be adding BFloat16 (BF16) support to their future processor designs, on Thursday they publicly provided more details on their plans for this new floating-point format to help AI / machine learning workloads with training and inference...

http://www.phoronix.com/scan.php?page=news_item&px=Arm-BFloat16-Support

[tildearrow]

Is it Copper Lake or Cooper Lake? (or do both exist?)

[ShizzyG]

Is it Copper Lake or Cooper Lake? (or do both exist?)

Cropper Lake*

[ldesnogu]

Is it Copper Lake or Cooper Lake? (or do both exist?)

Michael is correct: it really is Cooper Lake. I always read it as Copper Lake. Silly internal names; we have the same issue at my company where I have a hard time remembering what is what. Or perhaps am I just getting old?

[ms178]

From that blog post: "Experienced developers in the floating-point arithmetic field will appreciate that some results of a matrix multiplication performed using Arm’s BF16 instructions may differ marginally from other ISAs. However, such marginal differences occur with a similar frequency and magnitude to those expected due to the use of different instructions, software algorithms or hardware accelerators that change the order of calculations, and which do not materially affect the accuracy of neural networks." - is this marketing speak for: "Dear developers, excuse us for being less accurate with our new instructions, but the area and power savings totally make up for it (at least for us, YMMV)".

[FireBurn]

What's this supposed to say?

It's looking now like Arm might be AMD in to supporting BF16 on their processor designs.

[raster]

is this marketing speak for: "Dear developers, excuse us for being less accurate with our new instructions, but the area and power savings totally make up for it (at least for us, YMMV)".

No. It's choosing rounding mode for something that is already inaccurate. Round up, down, to nearest, to odd or to even... ARM has chosen the one that has a better performance/power consumption profile and testing has shown that whilst the choice is different to others who may round differently, the end results are within the same margins of error as other things like using GPU vs CPU to do the same thing etc.

[ms178]

No. It's choosing rounding mode for something that is already inaccurate. Round up, down, to nearest, to odd or to even... ARM has chosen the one that has a better performance/power consumption profile and testing has shown that whilst the choice is different to others who may round differently, the end results are within the same margins of error as other things like using GPU vs CPU to do the same thing etc.

I am not a developer, but don't they expect the same results from all implementations? Or is it up to them now to deal with these different rounding modes on different hardware?

[reavertm]

What's this supposed to say?

I think Michael wanted to say "might beat AMD in" as in supporting bf16 earlier?

[Michael]

I think Michael wanted to say "might beat AMD in" as in supporting bf16 earlier?

Whoops yeah was supposed to be "be[at]". Fixed now, thanks.