Re: Python recompile

On Tue, 18 Mar 2025 16:27:00 -0000 (UTC)
antispam@fricas.org (Waldek Hebisch) wrote:

bart <bc@freeuk.com> wrote:
 
Actually 20 lines per per second would be not bad.  Early Turbo
Pascal was considerd very fast and IIRC did 8000 lines per minute,
that is about 130 lines per second.
 
Modern machines are more than 1000 faster than early PC-s, probably
closer to 10000 faster.  If you belive in Dhrystones, slow RISCV
board is doing 1820.7 DMIPS, slow Atom 2952 DMIPS, And Zen at about
3 GHz 30501 DMIPS.  VAX in 1 DMIPS and is quite a bit faster
than early PC-s.
 

Not sure how VAX compares to early PCs in general, but if my deem
memories of casual 1st hand experience are worth anything then I'd dare
to say that compiling Pascal program on 780/11 with VMS Pascal was at
least 10 times slower than doing the same on PC/AT clone with Turbo
Pascal. Possibly, up to 50 times slower, but certainly not less than 10.
It is possible that AT clone was running faster than original IBM AT,
but it was 286, not 386.

OTOH, compiler are bad case for modern machines.  gcc folks
observed that gcc execution time correlates better with
databases than with compute intensive programs.  In particular
there are a lot of cache misses which are costly on modern
machines.
 

SPEC CPU benchmarks suit has gcc-based subtest. By this subtest, modern
CPUs are quite a lot faster doing gcc compilation than even circa-2006
Cor2Duo that back then was considered new class for this sort of tasks.
All result below are Base scores.

SPEC CPU2006
2006:
3.0 GHz, Intel Xeon 5160 403.gcc - 15.2 (i.e. 15.2 times faster
than Sun Ultra Enterprise 2)
2017:
4.20 GHz, Intel Xeon E3-1270 v6 403.gcc - 54.8

SPEC CPU2017
2017:
4.20 GHz, Intel Xeon E3-1270 v6 602.gcc_s - 11.0
2024:
5.4 GHz, AMD EPYC 4364P 602.gcc_s - 20.6

So, total speed up of gcc compilation in 18 years:
(20.6/11.0)*(54.8/15.2) = 6.75
Not too bad, I should say.

Something like Apple M4 or fastest Intel and AMD desktop chips are
probably even faster, but nowadays vendors don't submit SPEC results
for client-class CPUs.