[libre-riscv-dev] [isa-dev] Re: FP transcendentals (trigonometry, root/exp/log) proposal

[lkcl]

On Monday, September 9, 2019 at 10:37:41 AM UTC+1, lkcl wrote:
>
> On Monday, September 9, 2019 at 7:03:58 AM UTC+1, Allen Baum wrote:
>>
>> And that may be an excellent reason for a custom, rather than a standard 
>> extension. 
>>
>
> i am relieved you said "may".
>

Machine Learning, on the other hand, yes absolutely, the reduced accuracy, 
the fact that approximations fit perfectly well for most tasks, the 
"temptation" to use weird data types (8 bit IEEE754, 10-bit numbers, 
fixed-point FP), and the need for absolutely insane computational rates 
[1], these all mean that specialist custom extensions are not just likely, 
they're a necessity.

high-end ML is therefore definitely *not* a "user" of this proposal, and 
there is no point trying to target them.

on the other hand: low to mid-end ML *might* be, purely in an incidental 
fashion, because it will be what people can "get hold of" if you know what 
i mean.  i.e people use NVIDIA GPUs for ML right now not because it's the 
best, most efficient design, they use them because they can afford a 
Graphics Card and easily program it, i.e. they can't afford to do a full 
custom massively-parallel processor ASIC and certainly haven't the time to 
develop an entire new CUDA / OpenCL software stack.

l.

[1] the numbers i heard from someone putting in a proposal for a 
Machine-Learning-centric GPU were bananas.  4096 parallel processors with 
64-wide SIMD FP16 ALUs.  that's 256 teraflops on a single chip!