Trying to figure out the 16core/24thread Intel Core i9-12900 processor

Message boards : Number crunching : Trying to figure out the 16core/24thread Intel Core i9-12900 processor

To post messages, you must log in.

AuthorMessage
Dougga

Send message
Joined: 27 Nov 06
Posts: 28
Credit: 5,248,050
RAC: 0
Message 104820 - Posted: 15 Feb 2022, 4:25:16 UTC

I'm confused by the strange and vast difference between the different cores on the new i9-12900
Running equivalent work packages each on a single core using the Boinc distributed app, I'm seeing
the following amount of work done in about 4.75 hrs.
Core Load (from app "Open Hardware Monitor)
1 44.924% 100.0%
2 44.227% 100.0%
3 43.979% 100.0%
4 43.318% 100.0%
5 42.678% 100.0%
6 42.570% 100.0%
7 38.194% 100.0%
8 19.265% 100.0%
9 19.189% 0-0.8%
10 19.107% 0-0.8%
11 19.077% 0-0.8%
12 18.800% 0-0.8%
13 18.706% 0-0.8%
14 18.689% 0-0.8%
15 18.688% 0-0.8%
16 18.659% 0-0.8%

Can anyone help figure out what on earth is going on here?
ID: 104820 · Rating: 0 · rate: Rate + / Rate - Report as offensive    Reply Quote
Profile Grant (SSSF)

Send message
Joined: 28 Mar 20
Posts: 1675
Credit: 17,697,747
RAC: 19,279
Message 104822 - Posted: 15 Feb 2022, 7:31:57 UTC - in response to Message 104820.  

Can anyone help figure out what on earth is going on here?
Alder Lake CPUs can have a mix of Performance & Efficiency cores, or just all Performance cores (or with the new Atom CPU, just Efficiency cores). The Performance cores can (but not always) support Hyper Threading, The Efficiency cores don't.
Rosetta Tasks are set to run for a given time period. However they may take longer, or they may end sooner. At present, there's a mix of Tasks, some running 8 hours, others for only 3.

And you are running Windows 10, which doesn't support the new Thread Director that helps decide which cores do what & how hard it should go. And even then as Annand Tech pointed out, cores that are running background Tasks can end up running at reduced clock rates as they are perceived as less important than those that are running in the foreground.

The only workaround that i'm aware of is to set Windows Power Plan to high-performance. That should result in all cores (Performance & Efficiency) running at maximum even when not running a foreground Task.
Let things run for a day or so (if the work lasts that long, which is unlikely) with it set to High-performance & things should improve noticeably.


Oh, and you need to run the BOINC benchmarks as your system is still showing the default values, and they are used to determine the amount of Credit you get for the work you do, .
Grant
Darwin NT
ID: 104822 · Rating: 0 · rate: Rate + / Rate - Report as offensive    Reply Quote
MStenholm

Send message
Joined: 18 Apr 20
Posts: 18
Credit: 25,823,727
RAC: 9,266
Message 104890 - Posted: 17 Feb 2022, 22:50:24 UTC - in response to Message 104820.  
Last modified: 17 Feb 2022, 22:52:20 UTC

You are still running Win 10. Case solved. That and as mentioned above the two different cores run at different speed
ID: 104890 · Rating: 0 · rate: Rate + / Rate - Report as offensive    Reply Quote
Dougga

Send message
Joined: 27 Nov 06
Posts: 28
Credit: 5,248,050
RAC: 0
Message 104891 - Posted: 17 Feb 2022, 23:08:22 UTC - in response to Message 104822.  
Last modified: 17 Feb 2022, 23:08:54 UTC

And you are running Windows 10, which doesn't support the new Thread Director that helps decide which cores do what & how hard it should go. And even then as Annand Tech pointed out, cores that are running background Tasks can end up running at reduced clock rates as they are perceived as less important than those that are running in the foreground.



Interesting, so there's an actual technical reason to upgrade to Windows 11?
I was working at MS when Vista launched. It took continual harassment from my manager to get me to "upgrade".
There were scores of updates and fixes before I took the leap.

I do have things set on Max Performance for what it's worth.

This is your recommendation?
ID: 104891 · Rating: 0 · rate: Rate + / Rate - Report as offensive    Reply Quote
Dougga

Send message
Joined: 27 Nov 06
Posts: 28
Credit: 5,248,050
RAC: 0
Message 104898 - Posted: 18 Feb 2022, 2:38:57 UTC - in response to Message 104890.  
Last modified: 18 Feb 2022, 2:43:55 UTC

Intel Core i9-11900K has 25 cores.
With one core working with the GPU, it leaves 24 for Rosetta.
Calculating relative speed to complete a wu.


This messaging app mangles formatting.
Five columns:
1 - % we complete
2 - Time used
3 - Time used (decimal)
4 - Time per wu (decmial)
5 - Speed Category of Core

7 Fast Cores (perhaps the 8th working with the GPU)
61% 5:44:40 5.74 9.37 Fast
59% 5:34:29 5.57 9.41 Fast
59% 5:34:50 5.58 9.44 Fast
59% 5:34:55 5.58 9.43 Fast
59% 5:34:36 5.58 9.45 Fast
58% 5:29:29 5.49 9.49 Fast
58% 5:31:53 5.53 9.49 Fast

8 cores Medium Speed Cores
8% 1:30:20 1.51 18.80 Medium
8% 1:30:20 1.51 18.47 Medium
8% 1:30:20 1.51 18.81 Medium
8% 1:30:20 1.51 18.57 Medium
8% 1:30:20 1.51 18.58 Medium
8% 1:30:20 1.51 18.61 Medium
8% 1:30:20 1.51 18.60 Medium
8% 1:30:20 1.51 18.61 Medium

8 cores are slow as molasses. These must be the 'parked' ones
1% 0:28:31 0.48 33.33 Slow
1% 0:28:31 0.48 33.64 Slow
1% 0:20:19 0.34 38.05 Slow
2% 0:28:57 0.48 31.54 Slow
1% 0:28:42 0.48 33.43 Slow
0% 0:04:13 0.07 14.34 Slow
2% 0:29:10 0.49 31.94 Slow
1% 0:28:59 0.48 33.50 Slow
ID: 104898 · Rating: 0 · rate: Rate + / Rate - Report as offensive    Reply Quote
Profile Grant (SSSF)

Send message
Joined: 28 Mar 20
Posts: 1675
Credit: 17,697,747
RAC: 19,279
Message 104908 - Posted: 18 Feb 2022, 7:15:21 UTC - in response to Message 104891.  

I do have things set on Max Performance for what it's worth.

This is your recommendation?
Managed to lose a detailed explanation.
Short answer- Max performance is best.
Win11 will make better use of the P&E cores for general computer usage, but isn't likely to give any benefit for compute work where all cores are in use. If not all cores are being used for compute work, then Win11 will give better overall system performance/responsiveness.
Grant
Darwin NT
ID: 104908 · Rating: 0 · rate: Rate + / Rate - Report as offensive    Reply Quote
Profile Grant (SSSF)

Send message
Joined: 28 Mar 20
Posts: 1675
Credit: 17,697,747
RAC: 19,279
Message 104912 - Posted: 18 Feb 2022, 7:38:12 UTC - in response to Message 104898.  

Intel Core i9-11900K has 25 cores.
With one core working with the GPU, it leaves 24 for Rosetta.
Calculating relative speed to complete a wu.
etc...
If you have an app_config file on Collatz to reserve 1 CPU core for each GPU Task being processed, that would explain the missing core. If not, then you've probably got your Computing Preferences Usage limits, Use at most xx % of the CPUs set to less than 100%
It's best set to 100% and, reserve a core for Collatz GPU support.



I'd also check Task Manager to see if all cores & threads are running at 100%
That break down is very odd as i would expect E cores (which have hyperthreading) to all have similar processing times. And while slower, i would expect the E cores to do much better than those slow core times.
Grant
Darwin NT
ID: 104912 · Rating: 0 · rate: Rate + / Rate - Report as offensive    Reply Quote
Dougga

Send message
Joined: 27 Nov 06
Posts: 28
Credit: 5,248,050
RAC: 0
Message 104914 - Posted: 18 Feb 2022, 8:26:56 UTC - in response to Message 104912.  
Last modified: 18 Feb 2022, 8:43:51 UTC

The missing core is guiding a single GPU process/wu with another BOINC app.

There is something quite strange with these cores.
I'm not sure what to make of them.
Most of them show idle or near idle, only the 'fast' ones are showing maximum utilization.

I have preferences to use 100% of all cores.
There appears to be some internal CPU-mechanism that is slowing things way down.
These aren't VirtualBox wu, so that isn't relevant.
I haven't studies the nuts and bolts of these new Intel CPU's.
Someone said on here that there are different types of cores within the new processors.
This is news to me but I'll look for verification/more information.

From the Intel website:

CPU Specifications
Total Cores 16
# of Performance-cores: 8 - Max Turbo Frequency 5.10 GHz
# of Efficient-cores: 8 - Max Turbo Frequency 3.90 GHz
Total Threads 24 ??
Max Turbo Frequency 5.20 GHz ?? (Seems to contradict the above)


I belive my calculations clears things up. Given I'm showing 7 high performance Rosetta cores,
these are 7 of the 8 high performance cores. The 9th is doing other work.

The 24 threads is the curious bit.
This suggests that the 'efficient cores' are multi threaded 2 threads each which vastly under-perform.

What's not described is how the hyper-threading works.
It seems there are primary threads and secondary threads just for the 'efficient cores' with the latter vastly under-performing the primary which gives us three sets of 8 threads and three different performance metrics which is very odd indeed.
It's not clear at all what logic is used to determine which type of core/thread combination is designated for what workloads.
ID: 104914 · Rating: 0 · rate: Rate + / Rate - Report as offensive    Reply Quote
Profile Grant (SSSF)

Send message
Joined: 28 Mar 20
Posts: 1675
Credit: 17,697,747
RAC: 19,279
Message 104915 - Posted: 18 Feb 2022, 8:38:35 UTC - in response to Message 104914.  
Last modified: 18 Feb 2022, 8:46:56 UTC

Most of them show idle or near idle, only the 'fast' ones are showing maximum utilization.

I have preferences to use 100% of all cores.
So "Use at most xx % of the CPUs" and "Use at most xx % of CPU time" are both set to 100%
Power settings are set to High Performance.
But Task Manager is not showing 100% CPU load?

Have you got something like CPUID HWMonitor to check the CPU temperatures. Thermal throttling is the only other thing that comes to mind, but i would expect it to limit all cores & threads, not just some.
But given it's a new CPU type, and the OS doesn't fully support it's new features, anything is possible i guess.



Someone said on here that there are different types of cores within the new processors.
This is news to me but I'll look for verification/more information.
As i've been mentioning.
There are Performance cores and Efficiency cores.


And is there any particular reason you haven't bothered running the BOINC benchmarks?




The 24 threads is the curious bit.
This suggests that the 'efficient cores' are multi threaded 2 threads each which vastly under-perform.
What it suggests is that something is extremely screwed up.
As i posted several times before, and in a more recent post- the P cores are the ones that support Hyperthreading, the E cores don't.
Which is why i would expect 15 of the cores/threads to have similar processing times, with the 8 Efficiency cores having reduced output.
Grant
Darwin NT
ID: 104915 · Rating: 0 · rate: Rate + / Rate - Report as offensive    Reply Quote
Dougga

Send message
Joined: 27 Nov 06
Posts: 28
Credit: 5,248,050
RAC: 0
Message 104916 - Posted: 18 Feb 2022, 8:42:41 UTC - in response to Message 104915.  
Last modified: 18 Feb 2022, 8:54:07 UTC

Any thoughts on how to 'un-screw this up'?
I'm headed out of town asap so won't be coming back here for a couple weeks.

Thanks for your insight and help.

I currently have 32GB which doesn't seem to be an issue unless Virtual Box is involved.
I'll be upgrading to 64 at some point.
ID: 104916 · Rating: 0 · rate: Rate + / Rate - Report as offensive    Reply Quote
Dougga

Send message
Joined: 27 Nov 06
Posts: 28
Credit: 5,248,050
RAC: 0
Message 104918 - Posted: 18 Feb 2022, 8:47:10 UTC - in response to Message 104916.  
Last modified: 18 Feb 2022, 8:52:21 UTC

deleted
ID: 104918 · Rating: 0 · rate: Rate + / Rate - Report as offensive    Reply Quote
Profile Grant (SSSF)

Send message
Joined: 28 Mar 20
Posts: 1675
Credit: 17,697,747
RAC: 19,279
Message 104920 - Posted: 18 Feb 2022, 9:04:32 UTC - in response to Message 104916.  

Any thoughts on how to 'un-screw this up'?
I'm headed out of town asap so won't be coming back here for a couple weeks.
Just checking the CPU temperature is all that comes to mind at present.
You've got enough RAM, so all the cores & threads are being used.
BOINC CPU usage isn't limited (just double check that the values are correct- local preferences override web based ones).
OS is set for maximum performance

But something is limiting it.



A quick thought just occurred- running a different programme to make sure it's not just some BOINC weirdness with the new architecture.

Exit BOINC, install & run Cinebench, multithreaded test & check Task Manager that the CPU is at 100% load- all 24 cores & threads in full use.
Cinebench Release 23 Installers

If it runs all cores & threads at 100%, then it's something to do with BOINC/Win10/Alder Lake.
If not, it's something to do with the system- check the temps when running Cinebench.
Grant
Darwin NT
ID: 104920 · Rating: 0 · rate: Rate + / Rate - Report as offensive    Reply Quote

Message boards : Number crunching : Trying to figure out the 16core/24thread Intel Core i9-12900 processor



©2024 University of Washington
https://www.bakerlab.org