A Look At Alder Lake Against Older CPUs Going Back To Sandy Bridge, Kaveri

[phoronix]

Phoronix: A Look At Alder Lake Against Older CPUs Going Back To Sandy Bridge, Kaveri

After recently carrying out the Core i5 12400 Linux benchmarks against other modern processors, for curiosity I ran some benchmarks on some older hardware going back to AMD Kaveri and Intel Sandy Bridge for a look at how the modern Ubuntu performance compares across all those systems...

https://www.phoronix.com/scan.php?page=news_item&px=ADL-To-SNB-Kaveri-Benchmarks

[birdie]

I don't understand how to read these graphs.

In far too many tests, CPUs like Core i5 9400F are at the top of the list for "Perf Per Thread" - this can't be true, e.g. ZSTD Compression, Core i5 8400 is the fastest?

In far too many tests ADL CPUs are in the middle or at the bottom of "Perf Per Thread" or "Per Core" - this can't be true.

Lastly, would be nice to see which tests are multithreaded and which are single threaded.

In some tests, e.g. Blender 3.0, 12900K is more than 10 times (!) faster than 2500K. That's insane. "4 cores vs 16 cores (24 threads)" doesn't sound like it could lead to such a massive difference.

Would be great if you checked why some Web Tests are failing on some CPUs:
Core i5 2500K: The test quit with a non-zero exit status. The test quit with a non-zero exit status. The test quit with a non-zero exit status. E: ERROR: test_untitled_test_case (__main__.UntitledTestCase)
Core i7 4770K: The test quit with a non-zero exit status. The test quit with a non-zero exit status. The test quit with a non-zero exit status. E: ERROR: test_untitled_test_case (__main__.UntitledTestCase)
A10-7850K: The test quit with a non-zero exit status. The test quit with a non-zero exit status. The test quit with a non-zero exit status. E: ERROR: test_untitled_test_case (__main__.UntitledTestCase)

Could be that the distro you're using has applications compiled with the instructions these CPUs don't support.

Geometric Mean Of All Test Results looks great for ADL which isn't even supported by the Linux kernel properly. A 24 thread part beats a 32 thread part hands down.

[chuckula]

Thanks Michael!! I know that most reviews focus on pitting the newest & shiniest parts from competitors against each other, but in the real world very few people are changing hardware that fast.

These numbers provide much better context for those of us with old hardware to see how much of a performance uplift we can expect!

[uid313]

iGPU tests would be interesting too.

[mlau]

o

I

In some tests, e.g. Blender 3.0, 12900K is more than 10 times (!) faster than 2500K. That's insane. "4 cores vs 16 cores (24 threads)" doesn't sound like it could lead to such a massive difference.

ADL has AVX2 plus better overall ipc and turbo, so it's not that unlikely.

A 24 thread part beats a 32 thread part hands down.

Well, it's a wider design, and uses more power, so it'd be really disappointing if it didn't beat the 5950x. I noticed though that the amd-pstate does improve turbo on 5950x quite a bit, so it'd be great if Michael could repeat the test with linux-5.17 some time in the future.

[OrionHadur]

I don't understand how to read these graphs.

In far too many tests, CPUs like Core i5 9400F are at the top of the list for "Perf Per Thread" - this can't be true, e.g. ZSTD Compression, Core i5 8400 is the fastest?

In far too many tests ADL CPUs are in the middle or at the bottom of "Perf Per Thread" or "Per Core" - this can't be true.

Lastly, would be nice to see which tests are multithreaded and which are single threaded.

In some tests, e.g. Blender 3.0, 12900K is more than 10 times (!) faster than 2500K. That's insane. "4 cores vs 16 cores (24 threads)" doesn't sound like it could lead to such a massive difference.

Would be great if you checked why some Web Tests are failing on some CPUs:
Core i5 2500K: The test quit with a non-zero exit status. The test quit with a non-zero exit status. The test quit with a non-zero exit status. E: ERROR: test_untitled_test_case (__main__.UntitledTestCase)
Core i7 4770K: The test quit with a non-zero exit status. The test quit with a non-zero exit status. The test quit with a non-zero exit status. E: ERROR: test_untitled_test_case (__main__.UntitledTestCase)
A10-7850K: The test quit with a non-zero exit status. The test quit with a non-zero exit status. The test quit with a non-zero exit status. E: ERROR: test_untitled_test_case (__main__.UntitledTestCase)

Could be that the distro you're using has applications compiled with the instructions these CPUs don't support.

Geometric Mean Of All Test Results looks great for ADL which isn't even supported by the Linux kernel properly. A 24 thread part beats a 32 thread part hands down.

I dont understand that 'Power consumption monitor' graph either . THet heck are those lines and stripes,lol . Dude, Michael,or whats your name, how about you put down some regular human beings - readable graphs, like X= x%, AMD=x% , intel=this %, etc. ? >P .
Not all are native english people,and not all are the sooper masters of the holy CLI of linux (Tho I'm pretty familiar with it)

[ezst036]

Thanks Michael!! I know that most reviews focus on pitting the newest & shiniest parts from competitors against each other, but in the real world very few people are changing hardware that fast.

These numbers provide much better context for those of us with old hardware to see how much of a performance uplift we can expect!

This.

[bavay]

This might be a silly question, but are these results per core or per cpu? i.e. the 16 cores Core i9 12900K is ~3 times more powerful than the 4 cores Core i7 4770K because of its greater IPC or because it has 4 times more cores?

[bug77]

I don't understand how to read these graphs.

In far too many tests, CPUs like Core i5 9400F are at the top of the list for "Perf Per Thread" - this can't be true, e.g. ZSTD Compression, Core i5 8400 is the fastest?

In far too many tests ADL CPUs are in the middle or at the bottom of "Perf Per Thread" or "Per Core" - this can't be true.

Lastly, would be nice to see which tests are multithreaded and which are single threaded.

In some tests, e.g. Blender 3.0, 12900K is more than 10 times (!) faster than 2500K. That's insane. "4 cores vs 16 cores (24 threads)" doesn't sound like it could lead to such a massive difference.

Would be great if you checked why some Web Tests are failing on some CPUs:
Core i5 2500K: The test quit with a non-zero exit status. The test quit with a non-zero exit status. The test quit with a non-zero exit status. E: ERROR: test_untitled_test_case (__main__.UntitledTestCase)
Core i7 4770K: The test quit with a non-zero exit status. The test quit with a non-zero exit status. The test quit with a non-zero exit status. E: ERROR: test_untitled_test_case (__main__.UntitledTestCase)
A10-7850K: The test quit with a non-zero exit status. The test quit with a non-zero exit status. The test quit with a non-zero exit status. E: ERROR: test_untitled_test_case (__main__.UntitledTestCase)

Could be that the distro you're using has applications compiled with the instructions these CPUs don't support.

Geometric Mean Of All Test Results looks great for ADL which isn't even supported by the Linux kernel properly. A 24 thread part beats a 32 thread part hands down.

In the absence of support for ITD, Linux scheduling is as blind as Win10's: in tests that don't multithread well, work intensive jobs may end up on an E core. That's how ADL results can be all over the place.
Normally, you want to disable the E cores until ITD support lands, but for benchmarking that's not really an option.

[chuckula]

This might be a silly question, but are these results per core or per cpu? i.e. the 16 cores Core i9 12900K is ~3 times more powerful than the 4 cores Core i7 4770K because of its greater IPC or because it has 4 times more cores?

It's aggregate and the best way to figure out per-core performance is to look at single-threaded benchmarks in the detailed breakdown.