Multi-Chip Intel Core I9-11900K Overclocking Review: Four Boards, Cryo Cooling(anandtech.com)
anandtech.com
Multi-Chip Intel Core I9-11900K Overclocking Review: Four Boards, Cryo Cooling
https://www.anandtech.com/show/16857/overclocking-with-intel-rocket-lake-four-core-i911900k-binned-and-analyzed
45 comments
I don't doubt that it's better for everyone that Intel and AMD are getting closer to the theoretical max performance of chips right out of the box, but I do miss the days when you could work up the nerve to start playing with the voltage and frequency settings and get a free performance boost well after your purchase date.
Well, to satisfy your impulse to optimize your hardware you can look into undervolting. Getting lower temps/noise/running cost. In case of the AMD RX (4|5)(7|8)0 even higher performance, since they can hold their boost because they don't hit their thermal threshhold. For laptops it prolongs battery life.
Not possible on Intel chips any more due to Plundervolt.
<edit>Thank You for bringing Plundervolt [0] to my attention</edit>
They removed the possibility to change voltage on my chip from their eXtreme Tuning Utility in an "update". Had to scour the web for an older version. <edit>Luckily i use Windows around 3% of the time.</edit> Won't buy Intel in the future.
[0] https://plundervolt.com/
They removed the possibility to change voltage on my chip from their eXtreme Tuning Utility in an "update". Had to scour the web for an older version. <edit>Luckily i use Windows around 3% of the time.</edit> Won't buy Intel in the future.
[0] https://plundervolt.com/
I have my Ryzen 9 5950X overclocked to 4.1 GHz from a factory default of 3.4 GHz so I'd say those days are still here. Just needed an AIO liquid cooler.
Just by enabling PBO my 5950x goes automatically to 4.2-4.3 GHz all-core and 4.8-5.0 GHz single-core.
I heard that AIO liquid coolers are not necessary too. I saw some YT videos saying that air coolers were equal or better in performance.
For those who really want to live on the edge and take risks, I heard liquid metal was the what all the "cool kids" were into these days. Just be warned though, like a cool tattoo, some people regretted that decision.
For those who really want to live on the edge and take risks, I heard liquid metal was the what all the "cool kids" were into these days. Just be warned though, like a cool tattoo, some people regretted that decision.
I saw those videos too... a decade ago. Things have really changed. I could cite some benchmarks showing that 2x120mm Watercoolers these days trade blows with high end air, and 2x140mm or 3x120mm usually outperforms air, but consider the following:
1: Air is essentially limited to using two smallish radiators and a couple of ~140mm fans in the airpath of eachother with limited placement options. Water-cooling radiators allow you to run obnoxious amounts of fans with big heavy radiators and very flexible placement options.
2: Cars used to be air cooled and companies eventually gave up and started using radiators because they simply dissipate heat better, despite being extra expense and another point of failure.
As component power draw is going up, and a1o radiators are getting economies of scale, water-cooling has become far more mainstream and easily dominates the >$100 cooling market. We're even seeing watercooling in the datacenter space, Google tensor chips are watercooled.
1: Air is essentially limited to using two smallish radiators and a couple of ~140mm fans in the airpath of eachother with limited placement options. Water-cooling radiators allow you to run obnoxious amounts of fans with big heavy radiators and very flexible placement options.
2: Cars used to be air cooled and companies eventually gave up and started using radiators because they simply dissipate heat better, despite being extra expense and another point of failure.
As component power draw is going up, and a1o radiators are getting economies of scale, water-cooling has become far more mainstream and easily dominates the >$100 cooling market. We're even seeing watercooling in the datacenter space, Google tensor chips are watercooled.
Updated with references.
What matters for thermal performance is thermal resistance at the cold plate. You'll find that the best air coolers end where mid-range AIO coolers begin in this metric.
The NH-D15 sits pretty at about 0.11 degC/W but the Arctic Freezer II line is all sub 0.10 degC/W.
Air coolers: https://images.anandtech.com/doci/16898/TRvsSPL.png
AIO coolers: https://images.anandtech.com/doci/16427/TRSP.png
What matters for thermal performance is thermal resistance at the cold plate. You'll find that the best air coolers end where mid-range AIO coolers begin in this metric.
The NH-D15 sits pretty at about 0.11 degC/W but the Arctic Freezer II line is all sub 0.10 degC/W.
Air coolers: https://images.anandtech.com/doci/16898/TRvsSPL.png
AIO coolers: https://images.anandtech.com/doci/16427/TRSP.png
How hot does that get at full load? I was worried about having so many cores at factory if not overvolted. For example, you can find plenty of people freaking out about 5800X temps. On the other hand, people claim the high temps are fine.
They will never get close to the theoretical max performance. That's just bad business. You have to cut down the performance if you want to your market segmentation to survive.
>but I do miss the days when you could work up the nerve to start playing with the voltage and frequency settings and get a free performance boost well after your purchase date.
On the flip side, I'm glad that I can get the full performance of the chip without having to spend hours tinkering with it and/or worrying about silent data corruption.
On the flip side, I'm glad that I can get the full performance of the chip without having to spend hours tinkering with it and/or worrying about silent data corruption.
I for one hated gambling on overclocks that could brick my expensive hardware. I’m glad the equipment being sold today comes at max performance out of the box and that the reputable company making it will support it.
Regarding the max-performance-out-of-the-box, unfortunately you're mistaken. Part of the reason people could often get insane overclocks out of their hardware was that AMD and Intel would periodically produce a higher than expected number of processors that could handle a greater clock speeds. And rather than cutting prices on those parts they'd simply configure and sell them as a lower-spec'd part - so back in the day AXIA core Athlons may have been sold at 1GHz, but would often comfortably hit 1.33GHz and beyond[0]. This still happens to this day[1], nothing's different about that.
Whether you're comfortable going down this unsupported/warranty-breaking route is another story, but it's certainly not the case that all processors nowadays are sold at their peak possible performance out of the box.
[0] = https://hexus.net/tech/reviews/cooling/186-axia-overclocking
[1] = https://www.anandtech.com/show/16857/overclocking-with-intel...
Whether you're comfortable going down this unsupported/warranty-breaking route is another story, but it's certainly not the case that all processors nowadays are sold at their peak possible performance out of the box.
[0] = https://hexus.net/tech/reviews/cooling/186-axia-overclocking
[1] = https://www.anandtech.com/show/16857/overclocking-with-intel...
I embarrassed myself recently by trying to get the performance of my AMD Zen 3 CPU beyond what precision boost overdrive and the default XMP profile of my memory were able to achieve. The single threaded performance never came close to it with the kind of old school static overclock + overvolt I was running with my previous Skylake Intel chip.
I did end up getting some promising results by using the curve optimizer feature of PBO which allows you to change the voltage/clock curve on a per-core basis but while it was stable at higher clocks, it did have some trouble while idling or performing some tasks that weren't running the CPU 100%. I had to admit that I was doing all of this more for the excitement of seeing high numbers and not for the actual practical benefits. Testing for stability with non-demanding tasks just wasn't exciting enough to justify continuing.
I did end up getting some promising results by using the curve optimizer feature of PBO which allows you to change the voltage/clock curve on a per-core basis but while it was stable at higher clocks, it did have some trouble while idling or performing some tasks that weren't running the CPU 100%. I had to admit that I was doing all of this more for the excitement of seeing high numbers and not for the actual practical benefits. Testing for stability with non-demanding tasks just wasn't exciting enough to justify continuing.
Yea unclear who this chip is for with 8 cores burning so much power.. especially in the configuration where you overclock it crazily like this? or maybe i'm missing something?
Is power consumption actually a concern for anyone on a desktop chip?
I know I personally care about idle power consumption when looking at PC parts, but I don't really care if peak power consumption is higher if I can compile faster or run games at higher FPS (not that many games are CPU bound).
(This particular configuration is just for fun/show - nobody does liquid nitrogen cooling to actually use the PC, just to compete on benchmark numbers)
I know I personally care about idle power consumption when looking at PC parts, but I don't really care if peak power consumption is higher if I can compile faster or run games at higher FPS (not that many games are CPU bound).
(This particular configuration is just for fun/show - nobody does liquid nitrogen cooling to actually use the PC, just to compete on benchmark numbers)
Seems reasonable as a special part for super enthusiasts yea.. I guess I'm just more excited around the crazy high core count parts I guess than the high peak clock rates... 8 cores just feels like old news at this point for desktop I guess.. gotta have all of those chrome tabs open..
Yeah Intel actually dropped the core count on the 11900k to 8 down from 10 on the 10900k (and it predictably performs worse), which seems like a bizarre choice when they're currently playing second fiddle to AMD who have much higher core counts on their top end parts.
More cores means more heat means throttling for temperature. People get excited about huge core counts but single thread performance dominates in many workloads
Extra power consumption means extra money and/or bulk and/or noise on power delivery and cooling. Then all that heat ends up in my room which is undesirable for ~60% of the year.
Not necessarily if its also faster - something taking 5 seconds at 120w draws the same amount of power as the same thing taking 10 seconds at 60w on a slower but less power hungry CPU.
Try measuring the at wall power draw of your PC while doing some standard tasks + when idling/light web browsing and extrapolate out to monthly power bill - you'll probably find the cost to be pretty negligible even with more powerful desktop PCs.
(Of course none of this is an argument in favor of this particular CPU which has both high peak power draw and worse performance per watt than its competitors).
Try measuring the at wall power draw of your PC while doing some standard tasks + when idling/light web browsing and extrapolate out to monthly power bill - you'll probably find the cost to be pretty negligible even with more powerful desktop PCs.
(Of course none of this is an argument in favor of this particular CPU which has both high peak power draw and worse performance per watt than its competitors).
True, it is called the race to idle. A lot of work went into transitioning between pstates really quickly.
A counterexample would be one commenter here on HN, who got the lowest clocked, highest core count server he could get his hands on, for zipping through the compilation of C++ projects.
A counterexample would be one commenter here on HN, who got the lowest clocked, highest core count server he could get his hands on, for zipping through the compilation of C++ projects.
It is once you properly account for all the issues -
#1 - Higher PSU spend
#2 - Higher cooling spend (case + cooler). Trying to reduce the noise levels of a hotter component gets obnoxiously expensive, and even after this spend you still have to contend with coil whine and relatively large components. There's a reason Apple uses laptop components in its iMacs.
#3 - Higher power spend (total spend varies a lot depending on how often computer is at load)
#4 - Waste heat, which can lead to even MORE power spend if running A/C
#5 - Components generally failure more often when hot, so extra repair/replacement spend
Most of these are admittedly minor concerns (especially if you can live with noise and rarely run at load), but when you see things like Intel drawing 100w more at load than AMD with roughly equivalent performance, and when you consider the average desktop replacement happens after 6 years nowadays, these concerns add up. The savvy PC builder values power efficiency.
#1 - Higher PSU spend
#2 - Higher cooling spend (case + cooler). Trying to reduce the noise levels of a hotter component gets obnoxiously expensive, and even after this spend you still have to contend with coil whine and relatively large components. There's a reason Apple uses laptop components in its iMacs.
#3 - Higher power spend (total spend varies a lot depending on how often computer is at load)
#4 - Waste heat, which can lead to even MORE power spend if running A/C
#5 - Components generally failure more often when hot, so extra repair/replacement spend
Most of these are admittedly minor concerns (especially if you can live with noise and rarely run at load), but when you see things like Intel drawing 100w more at load than AMD with roughly equivalent performance, and when you consider the average desktop replacement happens after 6 years nowadays, these concerns add up. The savvy PC builder values power efficiency.
the rest might be, but #1 and #2 aren't minor concerns. If your old PSU doesn't provide enough power you're sol. And if cooling isn't enough cpu will be throttled, making getting a new powerful CPU meaningless.
>>Is power consumption actually a concern for anyone on a desktop chip?
Yes, because it means my office becomes an actual oven after playing games for 30 minutes. The heat from all of this has to go somewhere, and I don't have a very big room.
Yes, because it means my office becomes an actual oven after playing games for 30 minutes. The heat from all of this has to go somewhere, and I don't have a very big room.
Electricity in the UK is pretty expensive, so energy cost is definitely a consideration for me.
seems a waste to do this with intel when the 5950X from AMD is sooo far ahead in both total performance and performance per watt.
Granted the 5950X is a good $200 more expensive than the 11900k [1], but even the similarly priced 5900X has this Intel chip beat.
For reference the very best score they managed to get on the Pov-ray benchmark, after all their extreme overclocking and carefully selecting the CPU/Motherboard combo was 6872. The 5900X scored 7472 out of the box and the 5950X scored 8869.
Even more 'weird' is that the overclocked 11900k isn't even beating the score the last generations 10850k got out of the box. And that CPU can still be bought for $100 less than the 11900K
[1] granted again, they almost certainly spent more than $200 on cooling, so that isn't that big a difference.
For reference the very best score they managed to get on the Pov-ray benchmark, after all their extreme overclocking and carefully selecting the CPU/Motherboard combo was 6872. The 5900X scored 7472 out of the box and the 5950X scored 8869.
Even more 'weird' is that the overclocked 11900k isn't even beating the score the last generations 10850k got out of the box. And that CPU can still be bought for $100 less than the 11900K
[1] granted again, they almost certainly spent more than $200 on cooling, so that isn't that big a difference.
AMD might bench higher in multi-threaded benchmarks, but Intel dominates in single core benchmarks because of the smarter silicon and higher clock speeds.
That makes them still faaaar more suitable to certain workloads... like Engineering, where most of the tools designed in MATLAB (ie Simulink) are still single threaded.
I recently upgraded from an i7-10750 to an i9-11950 and the different is very noticeable IRL even though most of the cores sit around doing nothing all day!
That makes them still faaaar more suitable to certain workloads... like Engineering, where most of the tools designed in MATLAB (ie Simulink) are still single threaded.
I recently upgraded from an i7-10750 to an i9-11950 and the different is very noticeable IRL even though most of the cores sit around doing nothing all day!
If you only care about single core performance you dont need a flagship chip, you could buy an entire M1 system from Apple with the same single core performance for the price intel are charging just for the CPU.
you could buy an entire M1 system from Apple
Unless you need a reasonable amounts of RAM, a decent graphics card, or to be able to run a large class of professional software.
Plus if you check benchmarks you'll see that the 11900k still beats the M1 in several single thread benchmarks
Unless you need a reasonable amounts of RAM, a decent graphics card, or to be able to run a large class of professional software.
Plus if you check benchmarks you'll see that the 11900k still beats the M1 in several single thread benchmarks
If you cared about performance/benchmarks you'd switch to a multithreaded/parallel alternative and have the compute finish in 1/16th of the time on a 5950X.
Matlab/simulink make pretty heavy use of multithreading: https://www.mathworks.com/discovery/matlab-multicore.html
Even Excel runs its computations in parallel these days.
Matlab/simulink make pretty heavy use of multithreading: https://www.mathworks.com/discovery/matlab-multicore.html
Even Excel runs its computations in parallel these days.
All the toolchains I use (ie Simulink plugins) only support Windows unfortunately.
That must be torture. Hope it wasn't your idea to lock yourself in like that. You should try R, you'll probably find there are better versions written in that, probably even ones that support distributed compute.
Whoa, 200 extra watts just for the cooling system :)
That cooler looks much more impressive than that CPU. But 200W... whoosh!
its extra 200W cooler + up to 400W CPU
Wouldn't it be preferable to just have a 400W CPU :)