Intel Core i5 12600K / Core i9 12900K "Alder Lake" Linux Performance

[name99]

The M1 can only run a single p-core at 3.2GHz if all three other p-cores are in deep sleep mode.

The moment you spin up a 2nd p-core they both drop to 3.1GHz.

3 or 4 p-cores active throttles all of them down to 3.0GHz.

Do you know where you saw this? Is it established from OS tables?
And do know the equivalent numbers for the M1P/M1M?

[smitty3268]

Do you know where you saw this? Is it established from OS tables?
And do know the equivalent numbers for the M1P/M1M?

marcan (of asahi linux) figured that out while he was bringing up cpu reclocking support.

You can find it on his twitter.

The CPU's seem to have that built into them (or else the firmware that is initializing them on boot), and simply reject being put into higher frequencies by the OS if multiple cores are running.

Not sure about the M1P/M1M, but I suspect it is identical, just with 2 different 4-core clusters instead of 1.

Edit: yep, found this which seems to confirm the same behavior :

https://www.anandtech.com/show/17024...ormance-review

The CPU cores clock up to 3228MHz peak, however vary in frequency depending on how many cores are active within a cluster, clocking down to 3132 at 2, and 3036 MHz at 3 and 4 cores active. I say “per cluster”, because the 8 performance cores in the M1 Pro and M1 Max are indeed consisting of two 4-core clusters, both with their own 12MB L2 caches, and each being able to clock their CPUs independently from each other, so it’s actually possible to have four active cores in one cluster at 3036MHz and one active core in the other cluster running at 3.23GHz.

This is all on the minor side of turboing compared to x86, and it's fair to question how much effort Apple has put into that. I don't believe there are any favored cores or anything more aggressive, but it's essentially a 3GHz CPU that can boost up to 3.2GHz. Percentage-wise on x86 that would translate to about a 300MHz boost.

[name99]

marcan (of asahi linux) figured that out while he was bringing up cpu reclocking support.

You can find it on his twitter.

The CPU's seem to have that built into them (or else the firmware that is initializing them on boot), and simply reject being put into higher frequencies by the OS if multiple cores are running.

Not sure about the M1P/M1M, but I suspect it is identical, just with 2 different 4-core clusters instead of 1.

Edit: yep, found this which seems to confirm the same behavior :

https://www.anandtech.com/show/17024...ormance-review



This is all on the minor side of turboing compared to x86, and it's fair to question how much effort Apple has put into that. I don't believe there are any favored cores or anything more aggressive, but it's essentially a 3GHz CPU that can boost up to 3.2GHz. Percentage-wise on x86 that would translate to about a 300MHz boost.

Thanks for the update! Nice to see these details filled in.

[numacross]

No it's not, the only "optimization" is that it used the performance governor by default and the compile target is Skylake AVX-512.

To run Clear on Alder Lake Michael will have to disable the E-cores, thereby enabling AVX-512. Don't know if his motherboard allows for that.

Clear Linux doesn't require AVX-512, and the minimum requirements are stated in the documentation:

• Instruction Set:
  • 64-bit
• Instruction Set Extensions:
  • Supplemental Streaming SIMD Extension 3 (SSSE3)
  • Intel® Streaming SIMD Extensions 4.1 (Intel® SSE 4.1)
  • Intel® Streaming SIMD Extensions 4.2 (Intel® SSE 4.2)
  • Carry-less Multiplication (PCLMUL)

The distribution can use AVX-512 if available by having multiple versions of binaries available. And the default compile flags are: march=westmere and mtune=haswell.

[gnartz]

Michael what is the current status of the vulnerability mitigations? Is it still helpful to disable the on boot time to maximize performance?

[blackshard]

You wrote: "it just not feasible". The truth is: Software applications on heterogenous x86 CPUs are feasible.

Again, never said that heterogenous cores are not feasible in x86...
Not "feasible" was meant to be not a sensible path to follow for the user question about statically allocating threads/processes to particular cores if avx-512 instruction is detected by loader.

You also wrote: "The only way to do that is catching the IRQ fault during runtime ...". The truth is: It isn't the only way.

The only way to do the thing the user I was answering wants to do.

You wrote as well: "Yet you don't know if that instruction will ever be executed". The truth is: There do exist methods to mitigate this issue to a large extent.

Yeah of course, you spend more time in heuristics evaluating decision trees than in real code.

In summary: You are mostly trying to find ways how to make it fail. Please try to stop thinking in terms of simple source code patches with very low complexity. A proper solution to the problem of efficient use of AVX-512, with E-cores enabled, is complex. Intel disabled AVX-512 on Alder Lake because they fear software complexity, especially Intel's marketing department, and because Intel believes that the general public today (year 2021) would fail to understand the point of a heterogeneous x86 CPU.

Intel disabled the avx-512 for their purpose me and you don't know.
To me, it is just marketing. You get is as a bonus, but you have to disable the E-cores; avx-512 is not fused as always stated by marketing. It is just smoke for marketing purposes. We will see if their intent to really disable avx-512 or keep it as a "bonus" for geeky users.
edit: to clarify, Intel historically never gave *anything* as "bonus" that can be enabled by geek users, now I'm quite skeptic that, all of the sudden, they forgot to cripple a huge thing like AVX-512 on a processor it is supposed to not provide them. So my opinion is that they keep it available for power users that are happy to tweak the system and wants at all costs the latest "incredible" technology. My opinion is that the avx-512 presence on these cpus is a marketing gimmick to attract enthusiasts.

Also, the problem is not the "efficient use of avx-512", there is nothing so hard in catching the fault during runtime and moving the thread elsewhere. I don't see anything complex in this. Surely it requires patches to the kernel scheduler and some handler work, but in theory is surely possible, so your assumption about "simple source code patches ..." is wrong.

Your posts confirmed that Intel's marketing department was right: People do freak out, and don't know what to do, when faced with complexity.

Apparently they freak out themselves since, as you stated, they fear the complexity they put in themselves.

[L_A_G]

Say, how exactly is the 'b' important? Are you saying you would have looked the other way if it had been a single Dollar or just a lesser sum? Can you name a sum, which would have been ok? *lol*

Talk about getting hung up on stupid things... That was to emphasize that it wasn't just a faux pas, it was a very big faux pas. You don't get those kinds of fines lightly. That's more than the kinds of fines companies get when they actually kill people.

It is rather common practice in business, and in political systems, to play favours. If this is done with bribes, exclusive contracts, special offers, shares, lawsuits (and settlements), licenses, or a gentlemen's agreement behind closed doors, does not change what it is - it is an attempt to become successful or to remain successful. However, the form in which it is done can matter for politics and how it gets used in the press. They too like to have success, make money and get votes. To make it about an amount and it being billions is pointless. One can equally question why AMD, after they had gained the right to produce x86 CPUs, then went on to buy ATI, which of course cost billions, only to create an even bigger market share for themselves.

There's a big difference between working hard to maintain an advantage by offering better products at a better price and just straight up sabotaging a competitor that does with bribes and intimidation. The latter is very much illegal and what Intel did, hence netting them a fine in the order of magnitude higher than what you get when you kill people. I honestly don't see how you can somehow equate Intel breaking the law to an extent that netted them billions in fines to an entirely legal corporate buyout. It's as if you're trying to make some kind of "might makes right" argument in a similar vein to people who try to justify the abominable conduct of nazi Germany and the even worse conduct of WW2 imperial Japan.

AMD's problem was that they just could not produce a better x86 CPU, while Intel themselves did not create the best CPUs either, only AMD had the better political position (being the underdog in a dominantly left industry plays strong with voters). The moral is that AMD should have produced better CPUs, and only now they do.

At the time this was happening AMD was making objectively superior processors. When this was happening they were making the Netburst-based processors you just lambasted against AMD's widely lauded K8's. What the EU anti-competition authorities found when they investigated Intel and read their internal emails is that they Intel knew their Netburst-based products were inferior and all the illegal practices they initiated were to stop AMD from gaining market share with superior products.

Let that sink in; Intel knew AMD was making superior products and broke the law to stop them from capitalizing on these superior products.

Monsanto however sued countless farmers around the world over their GMO crops, driving them into bankruptcy and suicide. Their chemicals have poisoned countless people. Of course, Monsanto being a big company played favours in business and politics, too, like all other companies, but this is not the issue. The "issue", meaning deaths, Monsanto has caused is obviously a different one and needs no explaining. Then take today's pharma industry and how COVID has created winners and losers in the business. Obviously does hardly anyone care about morals when businesses make billions by saving lifes.

Considering we're talking fines in the same kind of orders of magnitude the two are very much comparable. It's not like Intel could kill people in the same way a chemicals company, but they've shown the exact same kind of lack of morality and willingness to break the law as Monsanto has in their pursuit of illegal practices to dominate markets for maximum profit. The law applying equally to rich and poor alike is one of the fundamentals of civilized countries' legal systems and there are no "might makes right"-exceptions for wealthy corporations.

Maybe you are one of those people who believes in "might makes right" and that the rich and powerful should be above the law, morality and even basic decency. But thankfully no even halfway decent country is set up under those kinds of basically fascist principles.