Apple Announces Its New M2 Processor

[t.s.]

Due to temporary migration issues, my work issued me both an Intel Mac and an M1 Mac. My experience is that M1 is vastly superior. Docker images build faster, tests run faster, Google Meets doesn't make the computer hot and the battery last at least twice as long. I'd also mention neither the builds nor Meets make the fans spin on the M1, but ermmmm, that would be cheating.
That said, the Intel one seems to be a lesser model. I'm not an expert in Macs, but both are MacBook Pros 13", the M1 is 2020 and has 16GB of RAM and a 512GB SSD and the Intel 2019 with 8GB of RAM and 128GB SSD. So, it very well just be the extra RAM that makes the difference, I can't really tell (and the disk makes a huge difference in that I don't need to rebuild images as often because the disk gets full :^) ).
Also I don't run x86 binaries, so I have no idea how those behave.

Why, of course it'll behave like that.
- 4 Core base 1.4GHz Turbo 3.9GHz (hot-hot-hot, throttle) vs 8-10 Core optimized core.
- 14nm+++ vs 5nm
- standard vs very high memory bandwidth
- usual soldered mem, ssd vs almost everyhting packed on the SoC (especially mem)

[t.s.]

..My experience is that M1 is vastly superior..

Oh, and almost forgot: Expensive elegant, performance, and good battery life machine, which is a break-go-to-trash garbage, no contest! Soldered wifi and RAM? Still acceptable. But soldered SSD?! Soldered SSD!! FFS!!!

[sinepgib]

Oh, and almost forgot: Expensive elegant, performance, and good battery life machine, which is a break-go-to-trash garbage, no contest! Soldered wifi and RAM? Still acceptable. But soldered SSD?! Soldered SSD!! FFS!!!

Note I'm not in favour of soldered anything really. But that's moving the goal post. Somebody asked if anyone here actually used an M1 and asked for their experience as user. My experience as user was that. Who cares why and how if that's not what's being asked? I didn't say it was because of the design or whatever. So chill my dude.
Soldered SSD, I agree, is the most absolute nonsense I've read since this shitty era of non-upgradable non-repairable hardware. I can accept it when we talk about a SoC and it actually has a positive effect in terms of performance, but the SSD, come on? Disks are also, in my experience, the piece that fails first, you must be able to switch that one. RAM shouldn't be soldered either IMO, and sadly most end-user products nowadays do that shit. And they no doubt will copy the soldered SSD at some point.

[Anux]

- usual soldered mem, ssd vs almost everyhting packed on the SoC (especially mem)

On those MACs SSD and RAM are soldered to the mainboard, its not included in the SOC. Only the controllers for RAM and SSD are on the SOC.

[ravyne]

This post argues that's not the case and that the reason M1 manages to do so well is a fundamental limitation in x86 ISA chips' ability to parallelize instruction decode for a CISC architecture with variable width opcodes. (In addition to other things like "SoCs have certain advantages and there aren't any x86 SoCs in that market segment yet".)

Fixed-width ISA makes it much easier to keep the instruction ROB full, which is why Apple Silicon ROB is so deep. In turn, that makes more independent instructions available for scheduling, and that, in turn, is why Apple Silicon can keep their wide execution engine busy. Hence, their high IPC.

Intel/AMD have SMT/hyper-threading precisely because they need to consider the extra thread(s) in order to find enough independent instructions to keep their ROB full enough to feed their execution engine. In large part, that's because they need more than one instruction decoder (logically, if not physically, not sure) to keep up. This is why an Intel or AMD CPU with SMT disabled typically has ~20% less instruction throughput, and as much as ~30% -- it's basically impossible to build (and keep busy) a wide and fast X64 (or any wildly variable-length instruction set) processor without SMT.

Intel is building a predictor into their current instruction decoder to cope. The predictor is a gain, but mispredicts introduce greater unpredictability -- that's gotten them nearer to Apple's ROB size, but they'll hit a wall with that too. Meanwhile, the predictor burns some power, consumes some area, and has become a permanent fixture of their future cores -- and doesn't contribute directly to the business of computation. These things may be negligable on their own, but it becomes a death by a thousand cuts. The sheer size and variable width of the ISA is contagious, it affects all kinds of engineering decisions.

Fixed-width or minimally variable-width ISAs just don't have that problem, or it's downstream effects. They can spend more of their power, area, and engineering resources on other problems (and certainly there are other problems in CPU architecture than these).

[brucethemoose]

What's wrong with being stuck on macOS aside from gaming? Which I see them making a huge impact this year since their architecture is similar between all their products. They even announced MetalFX Upscaling. They're way ahead of everyone else when it comes to ARM based apps and seems like they're doing fine with hardware. Unlike Windows/Linux all their apps are 64-bit, their HDR implementation is practically perfect.

Just depends on what software you need. I known some stuff I use just isnt tested on OSX.

[Developer12]

You do both realize that you can't yet daily drive Linux on an M1 Mac, right? Running a few synthetic benchmarks through a primitive interface isn't exactly an Apples to apples (pun intended) comparison.
This was the same sort of thinking back in the days where games ran faster in WINE vs Windows, because WINE simply lacked the rendering capabilities. It essentially was running games at a lower detail level.
Strip down MacOS to just a command line (which last time I checked, is actually possible - not sure if it still is), I'm sure the benchmarks would turn out roughly the same.
Run Linux with GNOME or KDE with compositing effects on and benchmark graphical programs like Chrome or a game and Linux isn't going to have such an obvious lead anymore.

I'm not favoring Apple here, I'm just saying that when all things are equal, they do a damn good job optimizing form and function. We can whine about their closed nature all day but they know what they're doing. Apple might not have the most optimal solution for each individual program but their forced homogeneity allows more complex software to run more efficiently. In the modern world, everything is layered with abstraction. Apple is effectively removing some of these layers.


It's loosely based on a BSD kernel. Some here would argue BSD is even faster than Linux.

Yes, by all means continue to tell me about how I can't daily drive this primitive interface: https://regmedia.co.uk/2021/08/23/asahi-gnome.jpg
Go do some research before you open your mouth.

You can run a full copy of debian on the M1, today, and the only thing that would affect benchmarks is the lack of power management. Yet the benchmarks largely come out the same between macOS and linux.

[Developer12]

Must be why x86 chips, despite having less transistor density (because built on inferior node) are still faster than ARM right? Or why the fastest supercomputer is x86 based huh?

I don't think performance means what you think it does. If you mention power efficiency please never touch the internet again.

Apple fanboys are more delusional than clowns eating Russian propaganda.

1. The nodes aren't that different. Only one generation.
2. The dies aren't even close to the same size. x86 dies are waaaaaay bigger.

No, I mean literal performance. Instructions completed per second. Even if you go all the way back the the pentium Pro and a Sun Ultra 5, the pentium has twice as many transistors to reach even close to the same level of performance. The difference between them, or between modern x86 chips and the M1, is that you burn too many transistors on pipeline control and instruction decoding and instruction caching when implementing the x86 ISA.

But sure, fewer transistors can get you more power efficiency. But to be honest? Power efficiency is intel's biggest problem right now even in performance. The heat dissipation of Alder lake is seriously knee-capping the clock speeds it can hit and sustain.

[Developer12]

X86 costs in efficiency due to that it is old architecture is 5% efficiency according to some Intel engineer. Internally processor is RISC like anyway, and only bonus is that you need to decode CISC into RISC.

In fact 6xxx series mobile CPUs for AMD trade blows very well in efficiency vs M1. Losing just a litle in ST scores but actually winning vs M1 on 8 core configuration per watt.

By far biggest impact is Apple's memory configuration (RAM chips soldered very close to CPU) so apple pays smaller price for something not in cache, probably less complicated I/O part of die because of closed platform.

And lastly number of transistors ... oh boy.

12900k is around ~~10bln transitors (no official data but it is likely to be overestimate).
RTX 3090 is 28.3 bln transistors (official data).

M1 ultra is 114 bln transistor. In nutshell in size of one m1 ultra silicone you can get 3 times 12900k + rtx 3090.

Now yes M1 ultra is faster in multithreaded workloads then 12900k and consumes less power. But we talk here about insane transitor diffrence. 12900k spending 5 times number of transitors on just E cores would bring huge efficiency improvments if they were tuned towards let's say 3GHz. And in single threaded 12900k wins.

At least in CPU war Apple is competitive. But when we get to rtx 3090 ... oh boy . In blender/V-ray etc. (and i am talking about CUDA/Vulkan vs Metal) we talk about... 500% performance diffrence for rtx 3090. In fact RTX 3090 stock, is more power efficient per work done on entire computer consumption then M1 ultra. And RTX 3090 is seen as "The inefficient one that when you drop TDP to 75% you still retain 96% of performance".

>according to some intel engineer
>"internally a risc processor anyway"

Maybe don't get your figures and taglines from the guys getting creamed.

[Developer12]

Seriously I very much doubt your claim here. What performance? Raw number or perf/watt? Also in comparison with what exactly? I hope you understand you can't compare 2 CPU preciselly when they are on different processes right? And also you don't know their transistor counts...

I'm speaking in general. Even as early as the pentium pro and the Sun Ultra 5 the pentium needed double the transistors for it's instruction decoding and control. It's "risc-like-core" didn't save it that embarrassment.