Apple M2 vs. AMD Rembrandt vs. Intel Alder Lake Linux Benchmarks

[Swifty Rust]

> I should note here that for our Chromium compile, we benchmarked compiling the Windows version on Windows, and the Mac version on macOS
Okaaayy, did they at least use the same compiler and settings? They are really scarce on info about test setup there.

Edit: I read that LLVM has less optimizations for ARM, that would explain it beeing much faster. A cross compile on either plattform would give an indication.

Results are correct. Cross-compile time is nearly identical. Cope with that

[Dukenukemx]

I never said many times. I said "exceed"



Yes, I noticed it while I was working, but I did not keep the timings I made.

A quick search found the same website that you have been quoting the whole time. If you took more time to search for your beloved benchmarks instead of bickering, you would have seen the M1-pro obliterates everything else when it comes to code compile, which is what I said (from own experience)

image.png

https://www.techspot.com/review/2357-apple-m1-pro/

Thank you. That made my eyes hurt. Ok fellas, lets look and see here. This is a benchmark that compiles Chromium, but is this for itself or for other platforms? Says so in the article.

While the M1 Pro so far has traded blows with the best x86 CPUs in rendering and encoding tests, Apple Silicon has a massive advantage in code compilation. I should note here that for our Chromium compile, we benchmarked compiling the Windows version on Windows, and the Mac version on macOS, using Google's recommended process for each OS. Unfortunately, I tried but couldn't get the MacBook Pro to compile the Windows version, even though this is possible on Intel Macs. So this benchmark wasn't as apples-to-apples as I'd like, but it's still a real-world use case in my opinion.

So basically the Macbook Pro could only compile code for itself, which doesn't make this a fair benchmark. Also the M1 Pro is a 26 core CPU, and would be faster than the M2. The Ryzen 9 5900HX the next fastest CPU to the M1 Pro is an 8 core. Core i9 1190hk is 8 core, Ryzen 7 5800H is 8 core, you're starting to see a theme here fellas. Also this is an old benchmark as much faster and more capable CPU's have come out since, from both AMD and Intel. Also we were talking about the M2 not the M1 Pro, which has like 18 more cores.

You Apple guys can't just go picking out benchmarks that tailors to your argument.
Apple Guy: M1 is more efficient, just look.
PC Guy: But the 6800U is pretty good too.
Apple Guy: That's not fair, you're using a fan.
PC Guy: Ok but M2 still loses even with a fan.
Apple Guy: Look at the M1 Pro it's better.
PC Guy: But is has like 18 more cores.

[Dukenukemx]

the fool of all fools is speaking...

Careful, your ad hominum is showing.

it makes perfect sense to test the m2 with a fan and without a fane.

It lost in most cases to the Ryzen 7 6850U, and sometimes to Intel. In some cases very badly. So I'm sure you want it with a fan. You know you could attribute it to the fact that this is a Linux distro in early stages.

if i would need to buy a laptop and i would have these intel and amd and apple devices to choose for the hardware alone i would choose the apple one.
the only fact what would stop me to do so is the bad linux support. and the intel and amd ones have better linux support.

By the time Linux on M1/M2 gets 3D acceleration working, AMD and Intel would have faster and more efficient CPU's out.

but if apple ever change mind and supports linux i am sure many people will buy a apple laptop

Linux guys sure but not the majority of people. We have trouble getting x86 guys onto Linux.

[yump]

Not for the ones I have/had. If you know how to force cTDP down from the Linux please share. I mean actually set the SoC power limit, which is usually done by notebook manufacturer in the firmware. I'm not talking about CPUfreq governors etc.

It's probably different on AMD, and it might be locked out on your chip/firmware, but the interface for Intel is in /sys/class/powercap, and you can change it by writing to the *_power_limit_uw control files.

$ grep . /sys/class/powercap/intel-rapl:0/*
/sys/class/powercap/intel-rapl:0/constraint_0_max_power_uw:84000000
/sys/class/powercap/intel-rapl:0/constraint_0_name:long_term
/sys/class/powercap/intel-rapl:0/constraint_0_power_limit_uw:110000000
/sys/class/powercap/intel-rapl:0/constraint_0_time_window_us:93696
/sys/class/powercap/intel-rapl:0/constraint_1_max_power_uw:0
/sys/class/powercap/intel-rapl:0/constraint_1_name:short_term
/sys/class/powercap/intel-rapl:0/constraint_1_power_limit_uw:120000000
/sys/class/powercap/intel-rapl:0/constraint_1_time_window_us:2440
grep: /sys/class/powercap/intel-rapl:0/device: Is a directory
/sys/class/powercap/intel-rapl:0/enabled:1
grep: /sys/class/powercap/intel-rapl:0/energy_uj: Permission denied
grep: /sys/class/powercap/intel-rapl:0/intel-rapl:0:0: Is a directory
grep: /sys/class/powercap/intel-rapl:0/intel-rapl:0:1: Is a directory
grep: /sys/class/powercap/intel-rapl:0/intel-rapl:0:2: Is a directory
/sys/class/powercap/intel-rapl:0/max_energy_range_uj:262143328850
/sys/class/powercap/intel-rapl:0/name:package-0
grep: /sys/class/powercap/intel-rapl:0/power: Is a directory
grep: /sys/class/powercap/intel-rapl:0/subsystem: Is a directory

On my machine, only the long-term limit seems to be functional, which is why I have the time window set to only 0.1s. I have a python script doing fan control and reactive thermal throttling.

That said, if you're doing this for battery life, the CPUfreq governor approach is probably better. That is, you should cap the maximum clock frequency with cpupower frequency-set -u 3.6GHz or something like that. Intel's strategy with single-core turbo and unsustainable boost clocks is a big part of why they lose so hard in efficiency. Low-intensity workloads like smooth-scrolling in a web browser cause the CPU to run at some ridiculous voltage. You can see from this graph that Apple has a more sensible approach.

[MadCatX]

That's a good test they even have YouTube watching runtime in it. But the power measurements are only taken for Cinnebench, or did I overlook something?
Keep in mind it's another Laptop and OS than what was tested here.

Cinebench is not the most appropriate benchmark but unfortunately the only one that I could find where they compared apples to apples.

Are you comparing the peak power or what do I get wrong?

The difference between sustained and peak power draw on the 6800U is noteworthy. It shows how much power the x86 chips have to burn through to get to their advertised maximum performance. Actively cooled M2 on the other hand barely seems to throttle down. Time vs. power draw dependence for the entire Cinebench run would be very informative but nobody tested the M2 in that way AFAICT.

[mdedetrich]

So basically the Macbook Pro could only compile code for itself, which doesn't make this a fair benchmark. Also the M1 Pro is a 26 core CPU, and would be faster than the M2. The Ryzen 9 5900HX the next fastest CPU to the M1 Pro is an 8 core. Core i9 1190hk is 8 core, Ryzen 7 5800H is 8 core, you're starting to see a theme here fellas. Also this is an old benchmark as much faster and more capable CPU's have come out since, from both AMD and Intel. Also we were talking about the M2 not the M1 Pro, which has like 18 more cores.

Again wrong, I am compiling Java programs where the output are platform independent Java byte code (i.e. .class files) and my M1 pro is around 3-11 times faster compared to a Thinkad T14s Gen2/Carbon X1 Gen 9.

If you actually have the machine (evidently you dont) you can see how ridiculous the margin is when it comes to compiling code.

[Raka555]

The part that you keep forgetting is that the M1-Pro can do this while on battery, for a whole day long, while remaining cool and noise free !

Also the M1 Pro is a 26 core CPU, and would be faster than the M2. The Ryzen 9 5900HX the next fastest CPU to the M1 Pro is an 8 core. Core i9 1190hk is 8 core, Ryzen 7 5800H is 8 core, you're starting to see a theme here fellas. Also this is an old benchmark as much faster and more capable CPU's have come out since, from both AMD and Intel. Also we were talking about the M2 not the M1 Pro, which has like 18 more cores.

And no, you are wrong again , the M1-Pro has 10 cores (8P and 2E), the rest of the cores are GPU cores.
And even if it had 26 general purpose cores, it is still more power efficient.

So basically the Macbook Pro could only compile code for itself, which doesn't make this a fair benchmark

No, wrong again.
It uses clang as its compiler, so it can do what ever clang is capable of.
I have cross compiled programs for my RPI Zero and bare metal RPI3, just fine.

I'd like to see you cross compile chromium to run on windows, on your linux machine ...
You probably won't even be able to get it to compile natively.

[MadCatX]

Again wrong, I am compiling Java programs where the output are platform independent Java byte code (i.e. .class files) and my M1 pro is around 3-11 times faster compared to a Thinkad T14s Gen2/Carbon X1 Gen 9.

If you actually have the machine (evidently you dont) you can see how ridiculous the margin is when it comes to compiling code.

As I'm sitting here, waiting for my Gen1 to build the kernel I thought I could share some ideas on this:

When I code, I do mostly C++ and some web garbage with Typescript and Webpack. Both workloads are a nasty mixed bag of multithreaded and singlethreaded workloads. The initial stages are very multithreaded whereas the final stage (linking and LTOing with C++, all the packaging and minimization and assets gathering with Webpack) is mostly singlethreaded. An x86 chip whose energy budget is already taxed by the MT stage now needs to deliver good ST performance but it can't do that because it's already too hot. A more complex project with multiple MT -> ST stages will strangle an x86 chip because it basically runs on "energy debt" to have good ST performance. Actively cooled M1/2 does not seem to need to boost way over its sustainable power drain to have good ST performance which *could* be a reason why it's so good at compiling.

[Slartifartblast]

5nm and putting the RAM on the SoC, it's nothing magical. The credit goes to TSMC.

[arQon]

Is compiling programs from source which fully saturates the CPU considered "real work" for you?

What part of "one in a thousand" confused you? Or was it all the words calling M1 "a 'meaningful' moment of progress" and such?

ffs...