On 11/1/2023 12:52 AM, The Horny Goat wrote:
On Thu, 27 Jul 2023 07:42:25 -0400, knuttle <keith_nuttle@yahoo.com>
wrote:
Could this be true in any significant or semi-significant way?
Anything that can be tied to climate change is always true and SHOULD
NOT BE QUESTIONED.
Thank you for sharing, Greta...
C++ isn't the most efficient language.
It tends to hump around routine names.
Some people seem to leave those in their
release code, because it may aid in bug reports.
In C at least, you seemed to have an option
to strip those routine names.
ASM is of course, the most efficient :-) We
used to have a lot of jokes about that at work.
That's because at first, we were constantly
challenged about our choice of HLL for the
development environment. Basically, the only
overhead, was line numbers. In a HLL, when you
move from one line of HLL code to the next,
a register gets incremented, and that register
content is the line number.
If the machine were to crash at that point, the
trap could print the line number on the screen.
If the executable was stripped, it couldn't print
the routine name, but it could print a line number
for you.
We had one developer, who arrived from the States,
and he was waving a thick program listing around. He
was walking around cube-ville and waving the stupid
thing around. Most employees ignored him, but eventually
he announced that he had "written something in assembler,
to show you fools how a HLL (High Level Language) is
so unnecessary". Once he said that, a few of the
devs started to snicker, because it is at that point,
they knew what kind of froot cake they were dealing with.
There were no hallway discussions with the bum :-) We
didn't go around hallways waving wads of paper to
impress people, where I worked.
And that's back when processors ran at 5MHz or 10MHz
sort of thing. Today, at 5GHz, nobody gives a rats
ass about a line number incrementer. You have more than
one integer unit per core, and the execution is hidden (runs in
parallel) and more than one instruction is retired in
the same 5Ghz clock tick. A modern core can retire four
instructions at once... but it hardly ever gets to do that.
*******
This is what happens on a 16C 32T, as you increase the number
of railed cores. It also gives you some idea what makes the
machine "skittish" between a single core half-used and a
single core fully-used.
idle 48W <--- Hyperthreading -->
1 core 113W 9 core 179W 17 core 223W 25 core
2 core 124W 10 core 204W 18 core 223W 26 core
3 core 125W 11 core 202W 19 core 27 core
4 core 147W 12 core 218W 20 core 28 core
5 core 147W 13 core 217W 21 core 29 core
6 core 161W 14 core 221W 22 core 30 core
7 core 161W 15 core 221W 23 core 31 core
8 core 179W 16 core 223W 24 core 220W 32 core 216W
The computers have closed-loop feedback on the cooling system
and they adjust both VCore value and frequency, to stay within
a "power limit" on VCore. The table does not include "work done".
When a lot of cores are running, it might drop to 4GHz, instead
of just barely managing 5GHz on a single railed core.
It also has policies on what to do when there are two chiplets,
and it does not treat the chiplets equally. On the AMD AM5 7000 series,
the ones with the asymmetric extra L3 cache chip, the same practice
of loading the chiplets differently, can be seen. The processor
above has symmetric L3, and no particular reason to be treating
the chiplets differently. By having one policy to cover everything,
it might simplify the code a bit.
Summary: High end stuff is skittish and wastes power. With enough
effort, an end user can adjust this. I don't know if you can
ever make the controls "completely rational".
The lower core count machines, aren't quite as skittish.
They represent a better value.
Power numbers above, are wall power, Measured by a P3 Kill-A-Watt.
The software that runs my machine, has an "ECO mode" , but
I have not tested it. There have been a few issues with this
kind of software, so I'm not interested in "burning anything"
in the name of science. The SOC power goes up and down, even though
there is no reason for it to do that, and that's a shortcoming
in the way they design these things (shared power converter, shared
with "something that moves"). I've even stuck my finger on the
SOC, and the power number is real.
So yes, the power the machine uses is relevant. If I ran the machine
flat out all year, it could cost me $700 for the privilege.
If you adjusted your machine (ECO mode), the Notepad++ statements
wouldn't matter quite as much.
Paul
--- SoupGate-Win32 v1.05
* Origin: fsxNet Usenet Gateway (21:1/5)