Apple M1 PCIe Driver Leads The PCI Changes For Linux 5.16

[Developer12]

Doesn't really make sense to me.

You have a vendor that keeps pissing on your efforts by actively denying any information.
I mean, I'd be fine if they was like "Yeah, we won't write anything but we'll give you all hardware documentation."
But they're not, so why aid them in any way? Why make their hardware run anything else?

As I see it, this is actually aiding Apple and costing more man-hours than writing code for documented hardware.
Much like any efforts to write Nvidia drivers is actually helping Nvidia out.

But all be like "Apple, shiny shit, but Nvidia is the plague"?

Until such a day that GNU decides to start running a chip fab, vendors will always be part of the equation.

Out of all the vendors around, apple's M1 chips are actually quite open and trustable, far more so than anything intel has ever made. Take it from Asahi themselves [1], the people actually analyzing them. If you want this openness to continue or dare I say increase then you have to incentivize it with your wallet.

There is no way around this. Not with petitions. Not with boycots. Not with licences. Companies don't care about or handily work around all of the above.

People (particularly the more extreme in free software) turn their noses up at anyone and everything, but at the end of the day that moves NOTHING forwards. Old thinkpads might be fine for the most zealous, but A) every thinkpad ever made (Stallman's T420 included) suffers all the problems in [1] (plus other issues [2]) and B) it ensures better products will never be made. Those people are out of the market and encourage others to leave it as well. All future products will be targeted at those who remain.

There are companies striving for better (eg. also Raptor Computing Systems) or in the case of apple, allowing it, and if that's going to both continue and expand then it needs to be supported. That includes *anything* more open than the industry average. Apple's M1, Raptor's POWER9, both are acceptable.

[1] https://twitter.com/marcan42/status/1444246840728174597

https://twitter.com/marcan42/status/1444298711652126726

https://twitter.com/marcan42/status/1444299067022987273

https://twitter.com/marcan42/status/1444299228197507080

[2] Closed-source EC firmware, a proprietary intel processor with factory-burned microcode (even if left unpatched; then it's just buggy too), kilobytes of on-die 8051 running thermal and power management, plus the 8051(s) running the trackpad, south bridge, sdcard reader, etc., and an i486 system in the iGPU doing scheduling and power management. Purity doesn't exist in old machines.

[coder]

an i486 system in the iGPU doing scheduling and power management.

Really? That's wild! It'd be amusing to run Linux on that.

[uid313]

Just curious where you are getting that opinion - maybe because a lot of the published comparisons are pitting mobile M1 against desktop x86 ? If you look at articles comparing M1 with mobile x86 the differences are quite small and AFAICS mostly a function of M1 being on 5nm while the x86 parts are on 7nm:

https://www.anandtech.com/show/16252...le-m1-tested/5

If you look at the first page of the same article you'll see that M1 power draw is estimated at somewhere around 20W, so fitting between 15W and 35W x86 seems about right.

My understanding is that Apple mobile M1 for iMac and Macbook beats everything that Intel and AMD have have on laptops, everything low-tier and mid-tier that Intel and AMD have on the desktop, and performs almost as good as Intel's and AMD's high-end stuff but at 10% of the power.

Disappointing in hardware specs or open source contributions?

Apple went with bigger, more efficient cores because nobody needs a 16-core laptop, but they sure care about battery life. Intel and AMD are both designing smaller cores that can clock high for gaming and can be packed somewhat densely for many-core workstation & server CPUs. Not to take away from Apple's prowess and achievements, but they're playing a different sport than Intel/AMD - like cricket vs. baseball. That's why their approach differs so much. And the fact that they're not weighed down by x86 also helps massively.

In terms of sociopathic behavior, no one is worse than Apple. For that reason alone, I won't touch any of their stuff.


You put them in the same bin, but AMD is still well ahead on perf/W. It's really Intel who has juicing the power consumption of their CPUs beyond all reasonable proportion, in order to try and stay competitive. Although AMD did play that game back in the Bulldozer era.


You're only seeing reviews of their top-spec K-series part. I want to see how well their non-K i7 and i5 compares with AMD. That should be more revealing of how they're really doing on the efficiency front.

In terms of hardware. x86 just seems uninteresting. Intel and AMD with their high-powered desktop CPUs that almost gets beaten by Apples low-power stuff for laptops.

Yeah, Apple are sociopathic, but their M1 hardware seems really great.

Intel Alder Lake is 250 W or so, the 12900K model pretty much requires water cooling. In reviews I read, it consumes very much power and is very difficult to cool.

Maybe AMD have more energy efficient stuff than Intel, and but even their stuff seems very inefficient compared to the stuff from M1.

[bridgman]

My understanding is that Apple mobile M1 for iMac and Macbook beats everything that Intel and AMD have have on laptops, everything low-tier and mid-tier that Intel and AMD have on the desktop, and performs almost as good as Intel's and AMD's high-end stuff but at 10% of the power.

That sounds like a single-thread performance statement, which makes sense since right now the M1 seems to be the widest & deepest CPU core around. A lot of the reviews are focusing on single-thread performance because that is where the M1 surprised reviewers.

If you don't limit to single thread then the situation changes considerably. As an example, the 15W 4800U (x86) has similar multi thread performance while drawing a bit less power. Unfortunately most of the articles are comparing M1 against desktop x86 and relatively few are comparing against what I would consider comparable (ie mobile) x86.

The Anandtech review was one of the more relevant ones IMO (note that Speedometer is single threaded):

https://www.anandtech.com/show/16252...le-m1-tested/2

The first graph (single threaded) probably comes pretty close to your understanding, while the second graph (multi threaded) comes pretty close to mine.

[uid313]

That sounds like a single-thread performance statement, which makes sense since right now the M1 seems to be the widest & deepest CPU core around. A lot of the reviews are focusing on single-thread performance because that is where the M1 surprised reviewers.

If you don't limit to single thread then the situation changes considerably. As an example, the 15W 4800U (x86) has similar multi thread performance while drawing a bit less power. Unfortunately most of the articles are comparing M1 against desktop x86 and relatively few are comparing against what I would consider comparable (ie mobile) x86.

The Anandtech review was one of the more relevant ones IMO (note that Speedometer is single threaded):

https://www.anandtech.com/show/16252...le-m1-tested/2

The first graph (single threaded) probably comes pretty close to your understanding, while the second graph (multi threaded) comes pretty close to mine.

And that third graph of that benchmarks shows the iPhone 12 Pro outperforming AMD's top-of-the-line high-end Ryzen 9 5950X. Consuming probably 1/100 of power of what the Ryzen does?

I wasn't aware of that the 15 W 4800U had similar multi-theading performance while drawing less power tho, but that was a nice revelation.

I wish Intel or AMD would make a ARM-v9 or RISC-V CPU, maybe it would be better than their x86. It seems x86 needs 100x the power than that of Apple's CPU in order to stay competitive (in single-thread performance).

[Developer12]

I wish Intel or AMD would make a ARM-v9 or RISC-V CPU, maybe it would be better than their x86. It seems x86 needs 100x the power than that of Apple's CPU in order to stay competitive (in single-thread performance).

With Alder Lake it looks like this going to only get worse. Perf-per-watt is absolutely terrible compared with anything from AMD, and that's with a DDR5 advantage.

[bridgman]

And that third graph of that benchmarks shows the iPhone 12 Pro outperforming AMD's top-of-the-line high-end Ryzen 9 5950X. Consuming probably 1/100 of power of what the Ryzen does?

Yes to the outperforming, but again that is a single thread benchmark. Even if you ignore the the difference between desktop and mobile power profiles - it's only a slight oversimplification to say that power goes up with the cube of the clock frequency - a 5950X running 1 of 32 threads is hardly going to be drawing anything like rated power. What the benchmark shows is two things:

#1 - wider deeper cores can give higher single thread performance if you can afford the R&D and the silicon area

#2 - Apple can afford the R&D and TSMC 5nm helps a lot with affording the silicon area

I wish Intel or AMD would make a ARM-v9 or RISC-V CPU, maybe it would be better than their x86. It seems x86 needs 100x the power than that of Apple's CPU in order to stay competitive (in single-thread performance).

Nothing to do with X86 vs ARM, it's about how wide and deep the execution part of the core is assuming the rest of the chip can keep up, eg the cache hierarchy and memory subsystem. The M1 core is wider and deeper than even Alder Lake:

M1 - 630 deep (out of order window)
Alder Lake - 512
Tiger Lake - 352
Zen3 - 256
Zen2 - 224

M1 also appears to be wider (more execution units) than any other CPU at the moment, although that is harder to quantify because of (a) different mixes of ALU, branch and AGU, (b) different degrees of instruction fusion in earlier stages. None of this has anything to do with ARM vs x86.

Single thread performance correlates more with width than depth, but you need significant increases in depth (instruction window) to find enough instructions that can be executed in parallel and take advantage of a small increase in width. Effective width is just a lot harder to quantify. Note that Zen3 was quite a bit wider than Zen2 although not much deeper.

Just to be clear, nobody is saying that the M1 is not a good CPU design - it's at least a half-generation ahead in terms of the usual progression of core width & depth - but you need to compare mobile x86 with mobile M1 to get an accurate perspective on power/performance differences. Comparing desktop and mobile CPUs then concluding that ISA is the key difference is not a particularly sound methodology.

it would be interesting to see M1 compared against 5800 rather than 4800 for mobile x86, since the mobile x86 parts used in the Anandtech article were Zen2. Not complaining - M1 came out a bit before Zen3 mobile and most of the articles were written during that window - but would be a more relevant comparison in terms of current production.

These are exciting times for CPUs. I haven't seen anything like this for ~20 years.

[coder]

If you don't limit to single thread then the situation changes considerably. As an example, the 15W 4800U (x86) has similar multi thread performance while drawing a bit less power. Unfortunately most of the articles are comparing M1 against desktop x86 and relatively few are comparing against what I would consider comparable (ie mobile) x86.

The Anandtech review was one of the more relevant ones IMO (note that Speedometer is single threaded):

https://www.anandtech.com/show/16252...le-m1-tested/2

The first graph (single threaded) probably comes pretty close to your understanding, while the second graph (multi threaded) comes pretty close to mine.

From the same article, different page:



This M1 is a 4+4 design, meaning it has 4 performance and 4 efficiency cores. It's telling to look how well just the 4 performance cores stack up against the 4800U, still beating it in FP. And then the full 4+4 configuration beats 4900HS at FP, while using less power.

So, I think there's a reasonable case to be made that Apple's P-cores are both faster and more efficient than Zen2.

[coder]

Moving on, we now have the M1 Max benchmarks (https://www.anandtech.com/show/17024...mance-review/5).

Even using just the 8 P-cores, the Max demolishes the 5980HS, which is almost a fair comparison, except that the Max occasionally goes above 35 W of package power, in these benchmarks. Now, if the 5980HS' package power doesn't include DRAM, then I think we can say it's still a fair comparison.

So, you're not wrong to point out that comparing it with an APU reduces the efficiency gap, but it's still substantial.

[coder]

With Alder Lake it looks like this going to only get worse. Perf-per-watt is absolutely terrible compared with anything from AMD, and that's with a DDR5 advantage.

We still need to see how their lower-power models compare. What's true for all CPUs is that efficiency falls off a cliff, when you juice it for max performance. If you keep it in a reasonable power envelope, it might still perform competitively.