Core i9-12900KS World Record: How I Hit 7.8 GHz on 8 Cores | Hacker News Reader

Core i9-12900KS World Record: How I Hit 7.8 GHz on 8 Cores

tomshardware.com

This was the kind of stuff that excited me back when I was first getting into PC gaming. Nowadays, computers are so fast and games are so forgiving that I haven't had to fuss with overclocking for years.

Back then it felt like the only option, getting maximum frames was a function of my ability to build a highly effective cooling setup and how much patience I had to tweak everything.

My very last foray into custom cooling was a homebrew phase change cooler using a cell I had bought off eBay. This is making me want to jump back in just for the fun of the build. But without the need for performance, it does lack a bit of thst purpose it had back then.

Unfortunately one of my favorite games of recent time (Total War Attila) is almost incapable of running on my current vanilla setup at decent frame rate (not a top of the line gaming PC, but a decent machine with a Ryzen 9 and a 30 series GPU) due to poor optimization and abdonment. The hardware used is years ahead of the game.

Same. I got tired of arguing on forums that some games I play are badly optimised, and it's not my $3000 desktop PC the bottleneck, especially when the vast majority play on a thermally constrained laptop. There's some nasty software around that would have issues running on this nitrogen cooled rocket.

Good hardware in gaming has somewhat become a waste of money because often the bottleneck is just bad unoptimised code, and if you had worse hardware you'd be happier compromising; meanwhile good game engines (like id Tech) fly on any modern hardware.

It becomes ever worse on Linux. I remember being basically unable to play native ports, only saved by proton.

> and if you had worse hardware you'd be happier compromising

hah, when the game in my original post was new I was running it on garbage tier hardware. I got spoiled when I had a top of the line GPU. Nowadays I have to run it like I still have decade old hardware to get decent perf.

I do wonder though, if it's so terrible on top of the line hardware, what do these developers test their games on that they run fine for them?

After 2 months of crunch time, everyone is like: fuck it, let's ship and go home

Even amongst fairly involved programmers I'm usually able to find a speedup just by fiddling with the compiler flags, I suspect a lot of games leave a fair amount on the table.

How many Joules have we lost to Microsoft C++ vs. LLVM for example

> I suspect a lot of games leave a fair amount on the table.

I worked as a programmer for a AAA game. My feeling is that there needs to be someone in the team with enough influence who cares about optimisation enough to invest a lot of time on it.

After working on the game for a while I had enough political budget to spend time fixing ocasional frame stalls that annoyed me, which felt nice. I also worked a bit on general optimisation, but mostly on stalls. There were definitely lots of people who cared about performance, but this probably varies from studio to studio. If I hadn't done this, I'm sure someone else would have fixed some of them, but not necessarily all of them.

Let's also remember that lots of teams are under tight pressure and deadlines to deliver features and they may not have time to work on performance even if they want to - and if you're using an external engine, this gets even harder to work on.

Anyway, just my 2 cents, maybe other people who were in similar contexts will chime in with their own experiences.

Another issue with game dev for games that release on PC is that devs are usually on very high-end machines. NVMe RAIDs, 3080's, etc. Great for productivity, but there's very little dog fooding on low-spec machines.

That is true. The game I worked on was also on PS4 and Xbox One, which are both somewhat similar in architecture to a PC (at least in terms of CPU performance), just slower. If we could make the game perform well on the XB1, we knew it most likely would do well on PS4 and PC as well. But you're right, there could be PC-specific issues that would only show up on low-spec PCs.

In previous console generations things were harder back then as each platform's code was more different than now.

I have been deeply impressed with the unreal engine over the last few years. If a gaming studio picks it, chances are much higher it works well on mid- to low tier machines. It's incredible how well optimized it is. It probably takes the "looks per CPU/GPU cycle" crown if that is a thing. ;-)

I just can't picture how performance couldn't be one of the milestones, surely it's common to have a minimum performance target at least?

Games are a realtime application, so there's always a performance target and some performance work unless the game is very simple. All professional game developers are conscious of performance to some extent.

On the other hand, games are not hard-realtime - if the framerate sometimes drops below the target or there's the occasional stall, not all teams will care enough or be able to fix that.

In my post above I was talking from the perspective of going the extra mile and trying to make performance really good, which not all games do.

Ah I understand, so in the circumstance you're thinking of it does generally run quite well, but there might be hitches or areas of lower performance that would take a bit more work to get performing perfectly. Without someone directly responsible for those areas as that are changes hands a few times no doubt, eventually it's reached launch time and no one took responsibility.

I recently started working with a large organization and it's made me empathize with how these behemoth pieces of work can miss seemingly obvious issues. People probably noticed, but no one did anything about it since it wasn't in their work stream anymore.

The minimum performance target is usually 30fps (more common) or 60fps (less common) as that's all that a consoles budget usually allows. All you need is one thread that comes in just under 32ms on a PS4, but maybe doesn't scale all that well. For example, it's a task that's heavily memory bandwidth or latency bound, which isn't really all that much better on modern top-end desktops. Bam, now you have both a "badly optimized" game and one in which all performance targets are achieved.

Which is not really all that unlikely of a scenario. Just look at the absurd gains that the 5800X3D achieves in some engines over the 5800X, despite the X3D running at lower clock speeds. GTA 5 is a good example here ( https://youtu.be/hBFNoKUHjcg?t=885 ) - the 5800X3D achieves the highest fps on that at 171, a staggering 40 fps higher than the 5800X (by comparison the 12900K is 154fps). Something about that engine just slams into a memory bandwidth/latency wall, at which point it just stops scaling.

The types of games I'm thinking about are actually non-trivial Indie/niche games.

I won't name it but there's a simulator game I play thats very detailed but technically just screams like at some point they either didn't know or didn't have time to care about architecture then just winged it so now you have huge amounts of logic dumped into a basically single-threaded program.

I trust game engine developers a lot, I trust game developers in general not all that much - making games has a very tight feedback loop so a lot of them don't sit back and do it right rather than passing "tests" (not that games have tests most of the time).

I see - I think it's my bias as well as seeing a lot of people complaining about AAA performance that led me to reply the way I did :)

Your example sounds like it could be caused by a program outgrowing its initial design and might almost need a complete rewrite in order to improve, which is an unfortunate thing that happens naturally sometimes.

At one point, not too long ago, if you weren't compiling with Microsoft C++, you were leaving lots of energy on the table.

Is Microsoft C++ faster than LLVM?

There was a time when Intel and IBM (PPC) compilers where unbeatable, but I don’t know if that’s the case anymore with the amount of investment done on LLVM in the past decade or so.

Microsoft C++ has improved a lot but no I mean LLVM being faster than it.

arma3 - 9 years old - still get 40fps in multiplayer on an RTX2080.

Its hilariously badly optimised but so inutterably brilliant that there is nothing* else that compares, it's a genre of 1* - kind of ironic that the game I really look forwards to playing after I upgrade backend of this year next will be a decade old.

arma4 should come up soon

Will there be arma4 on arm64?

Just add the DayZ mod and you'll not only be able to enjoy thrilling Zombie survival bugs, but also dreadful 20 FPS :)

Much like you I look forward to (mostly CPU) increases to get better arma3 performance!

We need all the performance we can get, because even if it was the worlds most optimized engine, we'd just create bigger scenarios. Who wouldn't find a 10,000 unit mission fun, even if it was only at 40fps? :)

Inutterably brilliant and a genre of 1, I couldn't agree more.

I don't know anything about that game in particular, but I have experienced extremely poor performance from old games before. Sometimes it can be fixed by installing an old copy of the DirectX 9 runtime. DX9 to DX10 completely changed architectures and eventually (I don't remember when), the old DX9 stuff was reimplemented as being emulated through the newer architecture. But that emulation is extremely inefficient.

One of the reason why overclocking has gone away is because they boost really high straight out of the factory.

A 12900K will coast at 5GHz if you let it.

I have an 8700K that has been running at a stable 5GHz for several years with just a basic watercooled setup and manual tweaking.

Will the 12900K give a meaningful performance bump at the same clock speed?

About 52% single-threaded, 196% multi-threaded improvement.

https://www.cpubenchmark.net/compare/Intel-i7-8700K-vs-Intel...

And probably also better performance per watt should you care about that.

Not sure about that one; TDP isn't exactly the same as actual power consumption, but for i9-12900K it is more than doubled.

You're comparing the 8700K TDP to the 12900K maximum turbo power (which you can only "achieve" when using all cores fully). In "normal" usage 12th gen is very energy efficient and perform (unsurprisingly) much better than AMD Zen 2/3 SKUs.

You appear to be correct. I dug around a bit, and on Intel's ARK there are 2 numbers: Processor Base Power (125W) and Maximum Turbo Power (241W, same as the one used with the benchmarks). It appears the Processor Base Power might be a closer equivalent to TDP than the other one. If that's the case, the 12900K is much more efficient than the 8700K.

Can't edit the comment anymore, but I made a mistake. The Maximum Turbo Power explanation is as follows: "The maximum sustained (>1s) power dissipation of the processor as limited by current and/or temperature controls."

Somehow I have misread it to mean that it can peak up to said power limit for UNDER 1 second, which is incorrect - the CPU will run at that limit for as long as cooling allows. The CPU will gulp down ~240W when under full load, as evidenced here at ~24 minutes: https://www.youtube.com/watch?v=fhI9tLOg-6I

“Normal” usage isn’t going to be terribly more energy efficient - efficiency doesn’t change much depending on number of cores loaded. (Unless you’re only loading efficiency cores in the case of 12th-gen Intel.) (More cores might actually be more efficient if it thermally limits the CPU - in that case the frequency will be throttled, which will put you at a more efficient place on the voltage curve.)

Tom’s Hardware 5800X3D review[1] has some efficiency numbers if you scroll down. High-end Zen 3 chips seem mostly a bit more efficient than high-end 12th-gen Intel. By far the most efficient CPU tested is the 5700X, which is AMD’s fastest 65W TDP part. They don’t have any low-power Intel chips (e.g. Core i7-12700T) in the test. (Though I’ve always found retail availability of low-power Intel chips to be terrible.)

1. https://www.tomshardware.com/news/amd-ryzen-7-5800x3d-review

If you are referring to the all-core Handbrake efficiency graph then I'm not sure what's to complain about. The 12th-gen top-end SKUs perform far better while being roughly as efficient as the slower AMD parts, which is a clear win in my book. The 5700X performing well here is not surprising, as it's heavily restricted by the low TDP so in this all-core workload it might even remain under 4 GHz, also resulting in the lowest performance by far.

The situation is very different for lightly threaded loads (e.g. what PCs tend to do a lot of the time, this also includes most games).

I watched GamersNexus' 5800X3D review a few days ago, and the thing that stood out to me was their direct measurement of power draw on the 12V rail; it included a 120W result for the 5950X, and 240W (!) for the 12900KS. These parts are pretty comparable for the workloads I care about (compilation, mostly) but Intel's best part does the same job in roughly the same amount of time, for more money and twice the power draw. Is there something I'm missing, before I just buy the AMD for an upgrade?

For Handbrake the 12700K performs about 12% faster than the 5900X while being about 13% less efficient - I don’t think “perform far better while being roughly as efficient” is a fair characterization there - it performs better by about the same amount as it’s less efficient.

And the 5700X having low performance is taken into account by efficiency metrics - efficiency is performance divided by power draw, so whatever it in performance it more than gains in reduced power draw. (Of course, you might value your time more than your electric bill.)

Thank you.

That link assumes the 8700K is running at 3.7GHz - any idea if there is still that much improvement given that mine is running at 5GHz?

The stock 8700k boosts upto 4.7 Ghz. I don't think your extra 300Mhz will make that much of a difference.

TBH, there are people way more qualified in this area than I, but AFAIK Intel CPUs run at the turbo speed as long as they have sufficient cooling, and the benchmarks run at maximum performance available, which would be 4.7GHz in this case. Therefore, the benchmarks should be pretty close to your OC. You could verify this theory by running the benchmark yourself.

Good to know - thank you!

Edit: I did and you are right. I’m about 10% higher than the benchmark but still WAY below the i9-12900KS (by 25% single threaded)

Time to upgrade :)

I had an i5 9600k that ran 5ghz all cores, and swapped to the i7 12700k. The i7 is still a lot faster out of the box and I undervolted it a bit while running at 5.1 ghz so it's also very quiet.

I benchmarked it. I get a score of 3068 for single-threaded performance, which is around 10% faster than the stock 8700K stats but it’s still 25% below the i-9 12900KS.

I guess it’s time to upgrade :)

If you are asking a general question about single threaded performance, then YES. Modern CPU performance is about so much more than clock speeds, and the TLDR is that there are many non-orthogonal factors that sometimes the only trustworthy quantifier is workload benchmarking.

If you're asking whether that 5Ghz is different than this 5Ghz, the main easy answer IPC (instructions per clock). Unfortunately there are no official IPC numbers, but every new release comes with vague unlabeled slides showing %increase, and then reviewres do their pinning the CPU at a specific clock and measure single-threaded performance to estimate relative IPC improvements (like this https://www.guru3d.com/articles-pages/core-i9-12900k-review,...)

If we go into multi-threaded workloads like compilation, then of course the new one completely throws the 8700k out of the water.

But please don't feel pressured to get on the latest and greatest bandwagon, I'm right here running around with a Thinkpad T460s (with a little bit of a need to upgrade).

_Usually_, if you are not waiting for your computer while CPU is on +90% utilization, there should be little gained from a CPU upgrade. (Workloads that depend heavily on cache size or memory latency are counter-examples)

"Unfortunately there are no official IPC numbers"

Nor can there really be. IPC varies pretty wildly, and depends a LOT on optimization for a particular CPU architecture. The instruction mix should match the CPU model's execution units in addition to not having any dependency stalls.

I once compiled a list of architectures’ relative IPC on the SPECINT 2006 benchmark:

  Cortex-A73 (ARMv8-A)  64.4
  Pentium M   69.0
  Bulldozer (AMD FX)  76.2
  K10 (Phenom II)   76.7
  Goldmont Plus (Atom 2nd gen) 77.5
  Core    80.0
  Bobcat    81.3
  Nehalem    90.0
  Jaguar (PS4)   93.5
  Sandy Bridge  100.0
  Tremont (Atom 3rd gen) 100.7
  Ivy Bridge  105.8
  Cortex-A76 (ARMv8.2-A) 114.7
  Zen   115.9
  Haswell   117.6
  Zen+   119.3
  Broadwell  121.5
  Skylake   124.0
  Zen 2   139.8
  A10 (ARMv8.1-A)  139.8
  Cortex-A77 (ARMv8.2-A) 141.0
  Sunny Cove Ice Lake 146.3
  Cypress Cove Rocket Lake 147.6
  Zen 3   167.4
  A11 (ARMv8.2-A)  186.0
  A12 (ARMv8.3-A)  213.8
  A13 (ARMv8.4-A)  235.0
  A14/M1 (ARMv8.5-A) 236.7

Really highlights the absolutely massive gap between Apple Silicon and Cortex/the-rest-of-ARM. Although you'd really want to classify which Cortex SoCs you're using in that list, as implementers love to chop off the L2 & L3 caches to save die space. This is how we ended up with the bigger.big.little configurations - the "biggest" CPU is just the full CPU with all the caches it was supposed to have had.

Known-(very-)naive question: how and why is the M1 so far beyond both the-rest-of-ARM and both Intel (236.7-(Rocket Lake=147.6)=89.1) and AMD? ((236.7-(Zen 3=163.4)=73.3)?

Of course the technically correct answer I'm not looking for is "because Apple optimized it better" :D, I'm looking more at the (obviously probably extrpolated/guessed) path they took to get there. For example perhaps the increased memory bandwidth contributed (ie $pc go brrrrt :P) but like... this is like the elephant in the room that Apple is basically The Best™ ARM shop in town right now, and I guess I'm trying to reason both about the engineering (how Apple got to where they are now) and the market reaction (how Apple and everyone else are going to go forward).

That IPC figure is a bit misleading because different designs clock differently, for example the Apple designs likely have shorter pipelines which uses cycles and power more efficiently but won't support higher clock speeds. I doubt an M1 could hit 7.8GHz no matter how much cooling you gave it, but the Intel design will never be as energy efficient. Apple has different manufacturing constraints because vertical integration likely lets them spend more on the dies, they are using better fabs, etc. Meanwhile they prioritize having very few variants in a way Intel doesn't (or from another standpoint can't because they need more SKUs to sell). The engineers who do this stuff are living in 100-dimensional tradeoff land though, unlike what popular tech rants tend to imply there are no free lunches.

At a higher level it is just hard to do designs at this complexity so the dominating factor is budget divided by how many different products the team needs to make. Intel is the outlier, they managed to sabotage their own dominant position, but across the rest of the CPU world this holds true.

At very least Apple have the process budget and dollar budget to have really expensive DRAM, huge caches, and huge pipeline/big buffers (big ROB etc.)

It’s described in detail at Anandtech [1][2]. It has mostly to do with the fact that Apple’s chips are so much larger in every way measurable. There’s no magic, just raw performance from comparably massively over engineered designs. At the time of its release their reorder buffer had 630 entries vs Intel’s 352 and AMD’s 256. The floating point units inside the cores could handle quadruple of Intel’s throughput per clock. Etc. Etc.

[1]: https://www.anandtech.com/show/16226/apple-silicon-m1-a14-de...

[2]: https://www.anandtech.com/show/16192/the-iphone-12-review/2

Is this compiled from one source or (say) from individual anandtech bnenchmark results? I just say this to ask if there is repeated readings.

It’s compiled from Anandtech benchmarks and archived forum posts. Unfortunately I didn’t write down each source.

Amazing how much better the M1 is relative to Intel's offerings. I notice that this list doesn't include Intel's latest core (Golden Cove). Do you have those benchmarks?

At the time I compiled it Golden Cove wasn’t out yet. Golden Cove has 19% higher IPC than Rocket Lake which would place it just above Zen 3 at a score of 175.6.

Thank you.

So my real question is: will I see a noticeable improvement when playing games? Especially World of Warcraft, which is a super old codebase that relies heavily on cpu.

I have a 3090 as my GPU which is overkill for WoW so I’m mostly wondering about whether the single-threaded performance of the new chip at 5GHz will be much more than my current chip, given that it’s running at 5GHz vs. the default 3.7GHz.

The new gen of Intel chips is somewhat notable because despite AMD fanboys (that's what they are) saying hurr-durr TDP they are very efficient gaming chips (i.e. realistic workloads), so yes it will be a win. They are absolute monsters on a single-thread. Might be worth waiting for Zen4 / raptor lake, at very least for a price drop.

I really doubt you will experience any improvement.

What kind of noticeable improvement do you expect? Are you below 60fps today? Are you waiting for game to load? Do you want to upgrade from 1080 to 4K?

Basically, what is your bottleneck? If you don’t have any visible such, what’s left to improve? I was under the impression WoW was so old that any gaming rig from the last decade would have no problem to run it.

In essence you get a CPU that is 4 (!) architectures newer.

Here are the architectures of Intel (according to Wikipedia):

1. i686 (Pentium Pro, Pentium II, Pentium III)

2. Core (Core 2)

3. Nehalem (Core i7-920, Core i5-750)

4. Sandy Bridge (Core i 2000 series)

5. Haswell (Core i 4000 series)

6. Skylake (Core i 6000 series)

7. canceled due to low yield

8. Sunny Cove (Ice Lake mobile 10xx and desktop 11xxx)

9. Willow Cove (Tiger Lake mobile 11xx)

10. Golden Cove (12xxx and 12xx)

Did you purposefully leave out Broadwell, or was it a mistake? (If it wasn’t a mistake, I’d be interested to hear you expand on why, I know it was a bit of a weird release.)

I think broadwell was a process shift, not a microarchitecture shift.

They also left out the whole Netburst (and it's decedents) arch too.

I think Intel would rather we all forget about netburst.

Netburst is not based on P6 (i686). Core is a development stemming directly from Enhanced Pentium M which stems from Pentium III. All others build on this. Broadwell is Haswell on a new node (14nm instead of 22nm).

Yup. I would rather undervolt, than overclock. An extra couple of frames isn't worth being uncomfortable while gaming in the summer.

Not a gamer but when I swapped my Ryzen 3700X out for an M1 Mac mini the temperature difference on my legs was the best feature.

I noticed a different going from an i9 MacBook Pro to an M1 Pro model. It’s faster, much cooler, I don’t have to listen to fans blasting all days, and I even have a visible dip on in my homes electricity usage.

I sympathise. A colleague of mine has an i9 MacBook Pro and it’s horrible and noisy.

This, x100. Most people would know of overclocking but not undervolting, with the irony being that the latter can be far more useful for the average user.

For example, I have a stock air cooled Nvidia GPU. Standard overclocking doesn't get me much because I get temp throttled. However undervolting, i.e., you boost clocks at lower voltages while not at higher voltages, leads to higher sustained clocks with _similar_ power/temps.

This is a consequence of power generally being linearly related to frequency and super linearly (even quadratically) to voltage.

Dumb question: why don't they do this from the factory?

Maybe stability margins, undervolting can also introduce instability.

Xbox 360, PlayStation 3 and Nintendo Switch emulators are very much CPU bottlenecked. Simulator games too, though many of those are memory performance bottlenecked.

Yes and no. They're CPU bottlenecked but not due to the lack of raw performance, but due to the fact that those custom architectures had some IP blocks that are so radically different that what you can find on ARM and X86, that can't be easily emulated in SW no matter how much raw CPU power you throw at it.

This is where you need to use FPGAs, but getting a 1:1 soft-core clean room reverse engineered clone of a PS3 Sony Cell processor would be a monumental effort we're unlikely to see anytime soon.

It would probably be easier to reverse engineer the games themselves and use hooks to replace the API calls that use the custom IP blocks with standard Vulkan API calls that achieve the same functionality for example.

AVX-512 maps really well to the SPEs inside of Xbox 360 and PS3. Just enabling AVX-512 increases frame rates at almost 20% in some instances!

Yeah. Back in the 90s, especially in the days of software rendering, a combination of overclocking and config tweaking was often the different between "playable" and "unplayable" framerates for something like Quake. The difference was hugely important.

These days, generally, you'll be getting playable frame rates no matter what in modern games. Powering up your rig or your config can help to eliminate stutters, push higher levels of detail, etc. But it's generally not needed just to play the dang game in some kind of baseline reasonable way.

A faster CPU would help me write code faster, especially when I'm working on a big hairy Rails monolith or whatever. But when you're working on something that slow, a 10% or even a 50% CPU usually doesn't help that much. You need to address the root causes in the software itself - fixing the test suite or whatever.

Feeling this vibe, back when overclocking meant changing pins or dip switches.

My first overclocking foray was a 333mHz Celeron up to 416 (if I remember correctly) and that was a performance improvement proportional to the base clock. But heat was then an issue and the cpu card (as it was) with the integrated cooler was getting too hot for comfort.

My last system I built was for an 4970k but the system doesn’t get used in anger anymore and so it sits at stock clocks under 600$ of custom water cooling gear, with me afraid to push it in case the 8 year old hardware blows a capacitor or mosfet.

I’m glad you went and did the phase change cooling. My dream next step was to put a peltier under the water cooling setup and get sub ambient temps but the power bill and possible condensation issue kept my plans as just being plans.

If you're curious about absolute maximum clock frequency records, the top is at another GHz more:

https://hwbot.org/benchmark/cpu_frequency/halloffame

Almost entirely AMD FX-8xx0, with two Celerons.

Those Celerons were not known for running cool when stock, either.

Evidently the Celeron D is very overclockable. Still, 8ghz. Wow.

Having a longer than average pipeline for the frequency (and the longest pipeline ever made for a production CPU, in the case of the Celeron) certainly helps. They have a 31-stage pipeline, and the AMDs have 20, while other CPUs are around 15:

https://en.wikipedia.org/wiki/Comparison_of_CPU_microarchite...

Imagine a liquid deuterium cooled Mill CPU running Half Life 4!

that one person on the table using liquid helium - bet they did it for a laugh

...a very high-pitched laugh.

indeed! I was confusing it with nitrous oxide

Actually they did it for a PR event with ASUS to get the record when the platform launched.

Liquid Helium is too expensive and difficult to handle otherwise. Even for extreme overclocking.

Will is strong, but the body is soft and flabby

To the home experimenter, please do this in a well ventilated area. Nitrogen asphyxiation is no joke.

Does anyone have an idea how would the situation would be with Apple M1 if actually pushed as hard as this CPU? Sure, Apple doesn't allow it out of the box but are we out of luck for hacking it? Would it actually have dramatic performance increase when overclocked? Is Apple leaving some performance on the table in order to achieve silent operations?

All the tests I've seen with Apple M1 vs Intel, have been between Apple M1 chilling at full load vs Intel being cooled with a loud and powerful cooler(or liquid nitrogen in this case). What if M1 was held at 100C with loud and powerful cooler or liquid nitrogen?

Apple is using a high density library on TSMC 5n. I'd be surprised if golden samples could hit even just 4GHz on LN2.

You would theoretically get something akin to the latest Intel processor running at over 10 GHz.

I fully expect that once Apple releases their server class CPUs at the end of this year we’ll get something that’s equivalent to the latest Intel processor running at 7 GHz.

> I fully expect that once Apple releases their server class CPUs at the end of this year

I assume you mean MacPro by this. I haven't seen anything indicating something bigger than the M1 Ultra coming anytime soon(pre-2024). Hell, the M1 Ultra itself was a surprise to me.

I saw M2 mentioned somewhere, but not sure if we'll see it this year.

When you say server-class, do you mean for desktops? Thought I just read recently that Apple is ending Mac OS server.

It depends on if the M1 can clock that much higher or not when given the thermal & power headroom. The M1 architecture internally could have other bottlenecks that limit how high it can clock.

See for example why clock speed records are filled with things like the AMD FX, even though Zen{+,2,3} is on paper (and in practice) a way better CPU - the older AMD FX just clocks really really well, even though it's otherwise a pretty crap CPU. Or for another example from history, the disaster that was P4 Prescott compared to the Core 2 that replaced it.

I'm sorta surprised there is nobody running liquid nitrogen cooled overclocked chips 'in production' and rentable.

For some usecases, such as running a single threaded application you just can't optimize any more, it would be worth the added cost and complexity to get a chunk more single threaded performance.

I think it would be good for server applications.

Possible reasons why it's not done are the very high energy cost of the cooling setup. I'm not sure if it's possible to bring the cost down enough to a practical level.

Another reason might be lower lifetime of many of the parts involved.

I think that this would be great for HFT, but ... what do I know.

Not nitrogen-cooled, but liquid-cooled overclocked rackmount servers (with consumer gaming CPUs rather than the usual Xeon CPUs) are a thing in HFT.

They’re crazy because they are so unstable that you basically can’t reboot them and expect them to reliably come back up, so you are stuck on that given kernel version and firmware etc (important for eg the custom network cards that often want firmware updates).

You just turn it on, run it for like a year, and then it dies and you trash it and get a new one delivered.

We use heavily overclocked 10980XE and beyond common teething issues with voltages and some specimens just frying the survivors definitely offer stability if you don't push them insanely far.

But I'd agree on the bleeding edge being generally more bleeding than edge (esp. from a performance engineer perspective) with problems ranging from kernel not supporting TSC calibration correctly to acpi issues to faulty (fried) instruction cache.

I work in finance (HFT) and some shops have been using liquid cooled, heavily overclocked gear for quite some time now. Typically with all cores locked to avoid c/p state transitions.

Even immersion cooling is getting traction and showing up in data centres here and there.

For higher power chips, we're running into limits of air cooling. You can only blow so much air and make heat sinks with enough surface area that are still able to fit into chassis. At some point, we'll have to change cooling mediums, use chip materials that can run at high temperatures, and or improve performance per watt.

Doesn't even need to be nitrogen, while that gives better possible OC there are easier to handle liquid gasses in common use (liquid ammonia).

Though ammonia is somewhat unpleasant (and toxic) it's also common in industrial refrigeration so supplies are steady and it's cheap - in a DC environment solvable.

Not something I'd fuck with at home though.

> there are easier to handle liquid gasses in common use (liquid ammonia).

I’m irrationally angry reading this comment.

I’d prefer to risk nitrogen asphyxiation that deal with the smell of ammonia. A cat has been peeing on my door step for the last few years.

Since when is liquid ammonia easier to deal with then LN2?

Ammonia is nasty stuff. Like way worse then a solution of ammonium hydroxide (household ammonia).

What is the chance that the ammonia itself would damage parts of the computer? Surely few, if any machines are rated to deal with liquid ammonia immersion with direct PCB contact.

Also, afaik Ammonia dissolves zinc into zinc hydroxide, which is a semiconductor, and liquid ammonia may also be electrically conductive even if gaseous ammonia is not. There's so many things that could go wrong with that in my mind, but I would really like to know where you heard about this or if you're doing work with this somehow and fill in my own knowledge gaps if you will.

The extra cooling means extremely inefficient power-wise. The power gains are not linear.

So past 1st to ring the bell, e.g. some high frequency trading setups - it makes absolutely no sense at all, otherwsise. Now the real kicker is that such overclocked setups are inherently unstable - so no mass deployment for you.

AIUI, if you build a custom chip to use low-power async/clockless logic, cooling it with LN2 would absolutely make sense. It would run at high speeds while sipping power. Not for commonly available chips, though - those will waste power with every clock cycle and release that as heat. (It's gotten to the point where chips cannot be fully powered on at any time - some part of the chip will always be power-gated and kept "dark".) So any practical subambient cooling system would simply be overwhelmed.

Look up Kingpin Roboclocker. That's the closest anyone has gotten to develop a fully automated system that could be developed into something you could run for extended periods of time in the right environment.

It's extremely hard to run chips at sub-ambient temperatures for a sustained period of time, as you have to get extremely careful about condensation.

I'm not sure how you avoid condensation damaging your motherboard at that temp?

You cover it with kneaded eraser or petroleum jelly or some other sealer, and cover that with lots of towels.

You can see on the pictures the MB is modded to avoid this.

you prepare it using wax/conformal coating (+towels) - it's quite a standard procedure for sub-ambient cooling.

You could use one of those non-conductive liquids [1]. They have products with pour points going down below -100C. Depending on the interface between the liquid nitrogen and the coolant it might still flow well.

[1] https://www.3m.com/3M/en_US/data-center-us/applications/imme...

If you're interested, here's a video of when a youtuber did it live on a 5900x about a year ago. They go into fine detail about the process.

https://www.youtube.com/watch?v=ISmyfzlf8SY

5950X*

And it's GN, a very rigorous testing, review, teardown channel with a team behind them. Feels like "a youtuber" sort of does em a bit of injustice :)

Saying GN doesn't really mean anything, unless you already know who they are, which obviously the person you replied to doesn't.

It's just their name; saying anything else would be equally meaningless. The point was just that it's a well regarded name within the 'scene'.

Their name is Gamers Nexus...

Thanks that answers my question.

Much appreciated .

You use Vaseline, electrical tape and duct tape, paper towels, and good airflow for extreme overclocking. And there will be condensation, but a motherboard won’t last long anyways, burnt out vrms and fried cpus

Insulation. Here you can see the motherboard getting prepared for LN2 overlock: https://www.tomshardware.com/reviews/overclocking-amd-ryzen-...

Can extremely dry air be used for it, or will it have some other side effects on Thermal conductivity?

You apply a conformal coating to the PCB.

Well, partially with good insulation, Vaseline or plastidip coating.

And partially, you just don't. It'll freeze over and since this is for competition and not daily use, that is okay. Usually nothing gets damaged, sometimes it does. Extreme OC is expensive.

The other answers all cover the prep, but the other key part here is you just don't run like this for very long. You run in like 1 hour bursts, and let everything dry out (and warm back up) between runs. So you're just not giving condensation enough time to build up enough to cause shorts in the areas that are not fully sealed off, or at least that's what you're banking on.

Intel fused off AVX512 instructions completely on the Core i9-12900KS

Am I right that this is something to do with scheduling and the Efficiency cores not supporting AVX512? So a process that's in the middle of running an AVX512 instruction might get scheduled to an E-core? And so with better OS schedulers this could be fixed?

I don't think Intel ever intended AVX-512 to be used by people who buy non-enterprise chips. People getting them enabled with workarounds (disabling E-cores) probably signaled to Intel that they needed to shut it down.

It's nasty, but it's how Intel operates. AVX-512 and ECC are some of the few value adds for their Xeon line. They couldn't get away with charging 3x as much for the same processor otherwise.

AVX-512 in consumer chips was not new to Alder Lake, though; the previous generation did support it without any workarounds (there were no E-cores to disable to begin with). Intel may have changed course regarding AVX-512 but I wouldn't be so sure about the reasons.

That's true, but it seems like the consumer chips that did have AVX-512 came out when Intel did not have the single threaded performance crown. We can only speculate on what Intel's motives and plans are.

I think they should embrace any ISA advantage they can muster in every CPU they sell.

The Xeon exclusive for this generation is AMX, and I believe Sapphire Rapids has 2 AVX-512 pipelines (sorry, too early to remember the correct term) vs 1 on Alder Lake / Rocket Lake. Disabling it on the consumer chips is related to the heterogenous architecture, since E-cores don't have support. Nobody knows exactly why they decided to fuse it off to disable with E-cores turned off, my guess is because it isn't fully validated/tested (save money) and a very unusual config to start with. Sucks for the 50 people who wanted to run that one emulator, and "save money" is a pretty bad reason for a halo KS product.

ECC is available on the consumer chips as of Alder Lake, it's supported with the W680 chipset.

Interesting to see they fused off the avx512 set. You just can't run that kind of code on these.

That jives with my experience with the 10?00K series in the Intel extreme compute element. I can glide cool at 3.5ghz x8 and get some thermal throttling up to 4.5x8, except when running avx512 workloads. Then it's much hotter.

Title is a bit confusing. When using "L" for liter it must be capitalized, plus the mention of 70 liters is not in the actual title only the article itself

This seems wasteful and pointless. I have a compiling machine with an AMD Ryzen 7 3700X 3.6GHz 8-Core, the AMD cooler, an M.2 SSD, a relatively efficient PS, and a very simple video card. Draws a silent 25W total when idle; hums quietly at 125W when compiling on all sixteen threads.

Switch this system back to water-cooled; what would it draw while idle? While fully loaded?

I’m a curmudgeon, but I still think the gaming industry has a lot to answer for in terms of power consumption. Steam even has a ‘Earth Day’ banner today, why should probably read: ‘turn off your PC’.

Author was given these parts for free. Basically worthless test. If he was given the funds to buy parts at random somewhere, it would be interesting.

But to get parts sent to you from Intel and then say "omg look how overclockable" is advertising at its finest and by finest I mean worst.

I'm not sure why the price matters. And reviewers don't tend to get golden samples in general; it's possible overclockers get some but I'd want to see evidence. Since they sent 14 this is probably legitimate.

It's not the Price that matters, it's the source. "Since they were sent 14 this is probably legitimate"??? Are you saying intel couldn't find 14 of the best of the best to send him? 2 they could pull off, but not 14?

It's inherently worthless because none of us are going to overclock our PCs and cool them with liquid nitrogen in daily use. It's a sport and top athletes routinely get their tools for free from manufacturers. In some cases they are custom built. An example where tools matter much more than the people using them: Formula 1' 1.6 liter engines are built, tested on a bench and only the best ones go into cars, 3 of them per car per year if nothing goes wrong. The author basically did that with CPUs.

According to this logic, life is meaningless

Try not to think about that too much. We’re all going to die, our planet will be consumed by our dying star and our galaxy will end up forever alone due to expansion of space. Life is an accident of chemistry, an inevitable outcome of an infinite universe.

Which is objectively true. What is your point?

Meaning is inherently subjective, so there is no point in the appeal to objectivity.

If you're gonna go the effort of doing this I see no reason at all why playing the silicon lottery is a bad thing.

It's not a consumer test. Of course it's advertising

This is a top class race car competition, not a family van shoot out. Top 0.001% percentile. No compromises.

Overclockers buy (or receive) binned CPUs all the time. The cost of buying top of the lines CPUs in bulk for overclocking is simply not feasible.

Reply on news.ycombinator.com