Overclockers are definitely exploring new frontiers:

https://www.techspot.com/news/97050-intel-core-i9-13900k-overclocked-world-record-9ghz.html

a record high speed achieved in overclocking - and achieved by use of a
more extreme overclocking technique than previously employed.

John Savard

On Thursday, December 22, 2022 at 4:22:38 PM UTC-6, Quadibloc wrote:
> Overclockers are definitely exploring new frontiers:
>
> https://www.techspot.com/news/97050-intel-core-i9-13900k-overclocked-world-record-9ghz.html
>
> a record high speed achieved in overclocking - and achieved by use of a
> more extreme overclocking technique than previously employed.
>
> John Savard
<
0.4% per degree Celsius.

Quadibloc <jsavard@ecn.ab.ca> schrieb:
> Overclockers are definitely exploring new frontiers:
>
> https://www.techspot.com/news/97050-intel-core-i9-13900k-overclocked-world-record-9ghz.html
>
> a record high speed achieved in overclocking - and achieved by use of a
> more extreme overclocking technique than previously employed.

Liquid helium for CPU cooling is an absolutely massive waste of
resources. Nice publicity stunt, though.

Quadibloc <jsavard@ecn.ab.ca> wrote:
>Overclockers are definitely exploring new frontiers:
>
>https://www.techspot.com/news/97050-intel-core-i9-13900k-overclocked-world-record-9ghz.html
>
>a record high speed achieved in overclocking - and achieved by use of a
>more extreme overclocking technique than previously employed.

Using Liquid Helium for word record "highest frequency" validation
runs has been before, admittedely there's currently only 6 results in
the CPU Frequency: Hall of Fame[1] after you filter for LHe results
but there's been more that has since been beaten by LN2 results by the
same overclocker and therefore are no longer listed.

So, uncommon but definitely not the first big headline LHe overclock
world record. The fact that HWBot actually has a specific filter for
LHe results should tell you that.

Not sure there are any other XO categories where LHe is used, "highest
frequency" is kind of special in that it only has to be stable for a
very short validation, after that it's usually clocked WAY down
(usually using a "slow mode" button which reduces the multiplier) so
it can actually SAVE that validation without crashing!

AFAIK this is common practice for almost all frequency validation
runs, if they're not hellishly unstable and requires slow mode to save
you're not trying (regardless of cooling medium).

I'm told LHe overclocking is rare becase most CPUs won't even run at
those temperatures and many of those that do doesn't scale to
temperatures that low (so no benefit).

9008 MHz is insane score by an all-star team and I assume they with
help of ASUS and Intel (the two large backers) checked lots and lots
of CPUs to find that ONE golden sample that just LOVED LHe.

And despite this they say they almost didn't manage to beat the old
record (which was on LN2), the Youtube description says this was
pretty much their last shot when everything finally clicked.

1. https://hwbot.org/benchmark/cpu_frequency/rankings

Torbjorn Lindgren <tl@none.invalid> writes:
>9008 MHz is insane score by an all-star team and I assume they with
>help of ASUS and Intel (the two large backers) checked lots and lots
>of CPUs to find that ONE golden sample that just LOVED LHe.
>
>And despite this they say they almost didn't manage to beat the old
>record (which was on LN2), the Youtube description says this was
>pretty much their last shot when everything finally clicked.

What I find more remarkable are the clock rates of ~5800MHz that Intel
and AMD achieve with ordinary air cooling and ordinary binning. The
step that AMD did between Zen 3 (below 5GHz with ordinary air cooling)
to Zen 4 is especially big. Yes, they use a new process, but for
Intel that usually means a generation (or more) of lower-clocked CPUs
(e.g. Broadwell for 14nm, Ice Lake and Tiger Lake for 10nm).

Given that the AMD FX-8370 (4.3GHz Turbo) was overclocked to 8.7GHz
(with LN2), more than a factor of 2, while the new record is a little
over a factor 1.5, it seems that the advantage of using very low
temperatures is dwindling.

- anton
--
'Anyone trying for "industrial quality" ISA should avoid undefined behavior.'
Mitch Alsup, <c17fcd89-f024-40e7-a594-88a85ac10d20o@googlegroups.com>

On Tuesday, December 27, 2022 at 11:42:45 AM UTC-7, Torbjorn Lindgren wrote:

> So, uncommon but definitely not the first big headline LHe overclock
> world record. The fact that HWBot actually has a specific filter for
> LHe results should tell you that.

That is interesting. This is just the first one I happened to have heard
about.

> Not sure there are any other XO categories where LHe is used, "highest
> frequency" is kind of special in that it only has to be stable for a
> very short validation, after that it's usually clocked WAY down
> (usually using a "slow mode" button which reduces the multiplier) so
> it can actually SAVE that validation without crashing!
>
> AFAIK this is common practice for almost all frequency validation
> runs, if they're not hellishly unstable and requires slow mode to save
> you're not trying (regardless of cooling medium).

That shouldn't have surprised me, but it did because I never did
pay much attention to overclocking stunts...

Of course, the reason that liquid nitrogen and such is used is not
so much to get the CPU down to those really cold temperatures -
in fact, if that happened, it would damage the CPU - but instead to
remove heat from the CPU quickly.

So if one wants a practical, stable system where the CPU is
running hot and fast, instead of liquid nitrogen, what one would
try to use is something at a more reasonable temperature,
say chilled water, but moving at as great a rate of flow as could
possibly be achieved, so as to be able to remove lots of heat
from the CPU without risk of the temperature becoming too low.

And IBM does exactly that to achieve 5 GHz in its z/Architecture
mainframes. The CPU chips, also, take quite a few cycles per
instruction, so the design is like the "failed" Pentium 4 and
Bulldozer chips, minimizing the number of gate delays per
cycle.

The idea of having a faster cycle, when the latency for each
individual instruction is the same or slightly worse, is, of course,
to allow a single core to run more threads in parallel (or more
instructions in parallel from OoO).

Since it works for IBM, will this kind of architecture make a
comeback for the consumer? It seems unlikely, as water cooling
is expensive. But how about advanced *server* CPUs for the
business market - with chips embodying the mainstream x86
and ARM architectures?

More instructions in parallel from OoO can't be reproduced in
other ways, but *more threads* certainly can - just use more
ordinary chips, instead of fancy water cooling, and you get the
same result much cheaper. So it _seems_ to me that IBM is
barking up a tree almost no one else would be interested in,
but I may be missing something.

John Savard

On Wednesday, December 28, 2022 at 4:38:08 PM UTC-6, Quadibloc wrote:
> On Tuesday, December 27, 2022 at 11:42:45 AM UTC-7, Torbjorn Lindgren wrote:
>
> > So, uncommon but definitely not the first big headline LHe overclock
> > world record. The fact that HWBot actually has a specific filter for
> > LHe results should tell you that.
> That is interesting. This is just the first one I happened to have heard
> about.
> > Not sure there are any other XO categories where LHe is used, "highest
> > frequency" is kind of special in that it only has to be stable for a
> > very short validation, after that it's usually clocked WAY down
> > (usually using a "slow mode" button which reduces the multiplier) so
> > it can actually SAVE that validation without crashing!
> >
> > AFAIK this is common practice for almost all frequency validation
> > runs, if they're not hellishly unstable and requires slow mode to save
> > you're not trying (regardless of cooling medium).
> That shouldn't have surprised me, but it did because I never did
> pay much attention to overclocking stunts...
>
> Of course, the reason that liquid nitrogen and such is used is not
> so much to get the CPU down to those really cold temperatures -
> in fact, if that happened, it would damage the CPU - but instead to
> remove heat from the CPU quickly.
<
No, transistors are faster when they are colder by 0.4% per degree
Centigrade. Wires are also faster (excepting silver between certain
temperatures.)
<
Maintaining those low temperatures allows those faster cycle times.
Not the rate of heat flow out of the part.
>
> So if one wants a practical, stable system where the CPU is
> running hot and fast, instead of liquid nitrogen, what one would
> try to use is something at a more reasonable temperature,
> say chilled water, but moving at as great a rate of flow as could
> possibly be achieved, so as to be able to remove lots of heat
> from the CPU without risk of the temperature becoming too low.
<
Stanford showed, in the early 1990s, that if the back side of a
silicon chip were cut to for a forest of silicon bumps that one could
remove 1KW per sq-cm and maintain reasonable transistor gate
temperatures with chilled water.
>
> And IBM does exactly that to achieve 5 GHz in its z/Architecture
> mainframes. The CPU chips, also, take quite a few cycles per
> instruction, so the design is like the "failed" Pentium 4 and
> Bulldozer chips, minimizing the number of gate delays per
> cycle.
>
> The idea of having a faster cycle, when the latency for each
> individual instruction is the same or slightly worse, is, of course,
> to allow a single core to run more threads in parallel (or more
> instructions in parallel from OoO).
>
> Since it works for IBM, will this kind of architecture make a
> comeback for the consumer? It seems unlikely, as water cooling
> is expensive. But how about advanced *server* CPUs for the
> business market - with chips embodying the mainstream x86
> and ARM architectures?
>
> More instructions in parallel from OoO can't be reproduced in
> other ways, but *more threads* certainly can - just use more
> ordinary chips, instead of fancy water cooling, and you get the
> same result much cheaper. So it _seems_ to me that IBM is
> barking up a tree almost no one else would be interested in,
> but I may be missing something.
>
> John Savard

On Wednesday, December 28, 2022 at 3:51:05 PM UTC-7, MitchAlsup wrote:
> On Wednesday, December 28, 2022 at 4:38:08 PM UTC-6, Quadibloc wrote:

> > So if one wants a practical, stable system where the CPU is
> > running hot and fast, instead of liquid nitrogen, what one would
> > try to use is something at a more reasonable temperature,
> > say chilled water, but moving at as great a rate of flow as could
> > possibly be achieved, so as to be able to remove lots of heat
> > from the CPU without risk of the temperature becoming too low.

> Stanford showed, in the early 1990s, that if the back side of a
> silicon chip were cut to for a forest of silicon bumps that one could
> remove 1KW per sq-cm and maintain reasonable transistor gate
> temperatures with chilled water.

Very interesting.

Did they need to do tricks with deionized water, or were they able to avoid
leaks well?

Also, of course, getting this close limits the pressure you can use. I was
thinking in terms of small channels in a piece of solid brass or something,
but this is obviously better.

John Savard