none (none.delete@this.none.com) on July 26, 2022 2:17 am wrote:
> Bill G (bill.delete@this.g.com) on July 26, 2022 1:40 am wrote:
> > Thank you for the link to the GMPbench results. Intel Alder Lake (i5 12600K)
> > is faster than Apple M1 on GMPbench by the percentages shown below:
> >
> > RSA encryption 78%
> > Multiply 58%
> > Divide 52%
> > GMP-bench 45%
> > Extended Greatest Common Divisor 23%
> > Greatest Common Divisor 22%
> > Pi computation 21%
> >
> > GMPbench is a single-threaded integer benchmark. If Alder Lake is running at its maximum turbo
> > frequency during this benchmark, Alder Lake would have a 53% higher clock frequency than Apple
> > M1 during this benchmark. In that case, Alder Lake and Apple M1 are doing about the same amount
> > of work per clock for multiply and divide. Alder Lake is doing about 31% less work per clock
> > than Apple M1 for the two greatest common divisor benchmarks and the pi computation.
>
> I only compared the aggregate score but you're right that it's worth looking into sub-benches
> to get a better vision.
>
> Now there's something more meaty if you want: https://gmplib.org/devel/asm
> This gives results for low level routines for many CPUs.

Once again I remind you that these numbers all represent GMP 6.2.1 which was released in November 2020 with a very quick and dirty M1 port that has not been close to optimized.
(The ARMv8 work on which it is based does the job, and may even be somewhat optimal for some older ARM Ltd cores, but it's not a great match for M1's capabilities.)
The number are relevant if you want to see what GMP out the box will achieve on M1 right now; they are not relevant if you are interested in any sort of "what could M1 achieve done right?"

> > The web page you linked to says “Most operations in GMP up to quite large operand sizes
> > would scale very well over multi-core systems, but poorly with SMT.” That seems to favor
> > Xeon and EPYC processors, at least until the Mac Pro with Apple Silicon comes out.
>
> I think what they mean here is that when the memory subsystem is not stressed too much (that
> is a thread is not idling too much waiting for data to return from memory), SMT is completely
> useless due to the use of finely tuned routines. They also likely mean that for moderately
> sized operands, multiple cores would not be starved due to memory bandwidth.
>
> As far as I know GMP itself is always using a single CPU even for the most advanced routines.