Discussion 12700k vs 5900x/3900x/5950x DC benchmarks info, and some build info.

[StefanR5R]

Also, I run geekbench, and single was less(1718) but multi was more(12831)

OK, here is a screen shot. cpuz and geekbench5. On to linux install.
View attachment 58554

OK, rebooted, then had to revert to novueu drivers (spelling). An now I get this. Now installing boinc....

1923
Single-Core Score

13138
Multi-core

Running 5.13 linux kernal.

So, geekbench is a bit silly, but anyway:

• original Windows score (Win 10 or 11?), DDR4-4000 c18(?)
  1718 single (100 %), 12831 multi (100 %)
• Windows 11, DDR4-3200 c17
  1755 single (102 %), 13094 multi (102 %)
• Linux Mint 20.1 (kernel 5.13), DDR4-3200 c17
  1923 single (112 %), 13138 multi (102 %)

Did I get this right?

Note that all these scores, especially the 'single' scores, are perhaps a bit a luck of the draw from run to run, as they depend on how much the P cores and E cores are accessed respectively. The previously discussed change of scheduler policy in upstream kernel 5.16 (change is enabled by default, but could be disabled by downstream distributors at compile time) would shift the emphasis to P cores somewhat.

If lscpu -e still gives the same ordering as in post #3, this command would launch a geekbench run on the P cores only:
taskset -c 0-15 Geekbench-5.4.4-Linux/geekbench_x86_64
Or on the E cores only:
taskset -c 16-19 Geekbench-5.4.4-Linux/geekbench_x86_64
No guarantee that this works as intended though, as I don't know whether or not geekbench picks up that the multithread runs will only have access to 16 or 4 instead of 20 logical CPUs. (More thorough would be to go into the BIOS and boot with specific cores disabled, but it may not be possible to set up an E-cores-only system that way.)

Edit:
Oh wait, Alder Lake and Linux kernel ≥4.10 support 'Turbo Boost Max Technology 3.0', a.k.a 'preferred cores'. This should cause the scheduling of lightly threaded workloads to be biased towards the preferred cores (which in case of Alder Lake, happen to be P cores). In the lscpu output of post #3, the logical CPUs 4,5,12,13 correspond to four threads of two preferred cores, I presume.

 

[StefanR5R]

Here are LWN's reports on the Linux kernel 5.16 "merge windows": part 1, part 2
The only relevant change is this:

Core kernel
[...]

• The CPU scheduler has gained an understanding of "clusters", a hardware arrangement where multiple cores share the same L2 cache. The cluster-aware scheduler will take pains to distribute tasks across all clusters in the system to balance the load on caches across the machine.

The change above affects Alder Lake insofar as 4 E cores share one L2 cache, while each P core (hence each pair of threads on a P core) has got its own L2 cache.

There was in fact another relevant change released with kernel 5.16: In many cases, it is best for performance to spread a workload across the P cores (but loading only one out of two SMT threads of the P cores), then additionally across E cores, and finally also across SMT sibling threads on P cores. This policy is followed from kernel 5.16 on. With previous kernels, SMT threads of P cores would typically be loaded before the E cores. (Source: LWN article)

Both the "clusters" related change and the SMT vs. E cores related change affect only partially loaded systems (i.e. systems which still have idle logical CPUs), not fully loaded systems.

 

[Markfw]

There was in fact another relevant change released with kernel 5.16: In many cases, it is best for performance to spread a workload across the P cores (but loading only one out of two SMT threads of the P cores), then additionally across E cores, and finally also across SMT sibling threads on P cores. This policy is followed from kernel 5.16 on. With previous kernels, SMT threads of P cores would typically be loaded before the E cores. (Source: LWN article)

Both the "clusters" related change and the SMT vs. E cores related change affect only partially loaded systems (i.e. systems which still have idle logical CPUs), not fully loaded systems.

The highest I can upgrade to is 5.13

 

[igor_kavinski]

The highest I can upgrade to is 5.13

At some point, you should disable the E-cores in BIOS as they may be holding the P-cores back. Then run GB5 in Linux and see if your DC points increase too.

 

[StefanR5R]

Phoronix has benchmark data with different amounts of P and E cores enabled via the BIOS:
Intel i9-12900K Alder Lake Linux Performance In Different P/E Core Configurations
Caveat: As usual, it is not immediately obvious which of these tests are lightly/ medium/ heavily parallel. Besides performance in numerous tests, some power draw figures are presented too.

Depending on the BIOS, disabled E cores may enable AVX-512 support in the P cores. Igor'sLAB tested what would be better: All P+E cores available, or just P cores but with the dormant vector arithmetic instructions enabled:
Efficiency secret AVX-512 on Alder Lake – The resurrected instruction set in a practical test
E cores disabled and AVX-512 disabled was tested too. They are showing power consumption, performance, and performance/Watt.

 

[Markfw]

Here is the latest calc on primegrid
5950x = 139,886.4 ppd
12700F = 85,669 ppd

The time per unit went up, but is still lower than the 5950x, probably due to the E-cores. And sill 20 threads vs 31. The 5950x is also doing a F@H GPU unit.

But for $325 vs $590 its not bad. However the motherboard was more($216 vs $150), and the memory will not run over 3200(otherwise the 2x penalty running at gear mode 2), while the 5950x will do at least 3600.

So thats $541 vs $740. So if my bath is correct that 189 point/$ for 5950x and 158 points/$ for 12700F. (not including memory, lets call that equal.)

5950x wins either way.

Edit: I upped the 5950x price to newegg at $590, since not everybody lives near a microcenter.

 

[StefanR5R]

In post #35, I commented on a Linux kernel update in which the CPU scheduler's policy was modified to distribute workloads better across "cache clusters" of cores.

Today I saw a news item at Phoronix, reporting on another scheduler change which is related to segmented caches. This change too affects only partially loaded systems, not fully loaded systems. Furthermore, it affects only NUMA systems (and is therefore actually off-topic to an Alder Lake thread). Background: The Linux CPU scheduler tries to balance workloads across NUMA nodes, e.g. take care that both sockets of a 2P computer are similarly loaded. In Linux 5.11 (I think), this balancer was slightly changed to leave an unbalanced load alone if the heavier loaded node was less than 25 % loaded. Empirically, at least computers with one big last level cache per NUMA node profited in many (light) workloads from this modified policy. But it was discovered that computers with several caches per NUMA node don't (notably, AMD dual socket computers). Hence, another scheduler patch is submitted for Linux 5.18 which lets the NUMA balancing act earlier again if there is more than one last level cache per NUMA node. Maybe it's fair to say that the NUMA balancer's goal of making somewhat even use of all available NUMA nodes is extended to the goal of making somewhat even use of all available last level caches.

So this bit is not totally on-topic to Alder Lake. But as CPU topologies are becoming more complex, it's certainly good to know which means exist in the operating systems to use such newer CPUs.

 

[Markfw]

OK, so Now 3 computers have Rosetta tasks. The 2970wx is doing them in 9 to 10 hours, and the 5950x is doing then in 9 to 10 hours, and the 12700F is doing them in 9-10 hours. I don't think this benchmark tells me anything. Ihave to wait until WCG comes back on line...3 more weeks and a few days...

And I was about to post wrong times, the ETA is garbage. Its at 50% in 5 hours, and says 14 hours to go on the 12700F, but 5 hours and 5 hours to go on the other 2 ? And the 12700F is on linux, thew other 2 are in windows.

 

[mmonnin03]

No, Rosetta can't be used for benchmarking. The tasks run for so long and finish. The non-VBox tasks ya could set the estimated run time and would roughly end around then on any PC even though different amounts of work was actually completed. You'd have to compare a good amount of average credit per task between PCs if you wanted to use Rosetta.

 

[Markfw]

OK< I can't find it, but SOMEBODY posted a link to 5.16 linux install kernal on mint. It said something about PPA ????

 

[StefanR5R]

Kernel updates would affect only situations in which few cores are loaded.

 

[StefanR5R]

I want to build a Universe@home BHspin v2 benchmark which shall be dead easy to run and give good comparisons between different computers, or different settings of one computer.

It should also be possible to build a Rosetta@home 4.20 benchmark which is fit to serve these very same purposes.

The only trouble which is holding me back is lack of spare time. Also, porting the little benchmark environments to Windows could be troublesome. Scripting this with bash and let Windows users install cygwin for this would be convenient for me, but not for Windows users. (Or can WSL's bash launch native Windows programs?) — Perhaps python could be a good middle ground between having a somewhat suitable scripting environment and not having too complicated dependencies.

(Man, Linux is so much easier than Windows.)

 

[crashtech]

@StefanR5R , afaict, the two file systems (WSL & Linux) aren't even aware of each other, so it's not even easy to see if a Windows executable will run. I'm betting not, though. Also, when WSL is enabled, Cygwin doesn't work, not for me anyway.

 

[Markfw]

I want to build a Universe@home BHspin v2 benchmark which shall be dead easy to run and give good comparisons between different computers, or different settings of one computer.

It should also be possible to build a Rosetta@home 4.20 benchmark which is fit to serve these very same purposes.

The only trouble which is holding me back is lack of spare time. Also, porting the little benchmark environments to Windows could be troublesome. Scripting this with bash and let Windows users install cygwin for this would be convenient for me, but not for Windows users. (Or can WSL's bash launch native Windows programs?) — Perhaps python could be a good middle ground between having a somewhat suitable scripting environment and not having too complicated dependencies.

(Man, Linux is so much easier than Windows.)

Do linux. I have 18 linux boxes, and 3 windows. And a couple of dual-boot windows/linux.

 

[StefanR5R]

From a little reading on the web it seems that a shell in the original WSL (nowadays also known as WSL 1) is actually able to invoke Windows-native programs (that is, EXE binaries) since Windows 10 Creators Update. WSL 2 on the other hand, being a fully blown virtual machine, obviously cannot do this. — OK, a stack of Windows -> WSL 2 -> Linux -> Wine -> EXE should work, but wouldn't perform quite the same as Windows -> EXE. ;-)

Edit,
some further reading suggests that Microsofts's PowerShell may have all the necessary elements which I have in mind for the benchmark script.