Apples new M1 microprocessor

[Marco]

Anywhere with a strong security posture has windows as well due to GP and management infrastructure.

It's a tower of sand, offices are going to go back to dumb terminals and a windows computer is an even more terrible terminal than a desktop environment ... it requires an incredible amount of management infrastructure to deal with just how terrible a terminal it is.

But I said Microsoft would lose the consumer space. Inertia is much less of a factor there ... and all that management infrastructure protecting fools from the insufficiency of Windows at the office fails to protect them at home. System administration for windows is a terrible job and consumers are going to rid themselves of it.

The open PC probably also runs Altium or Zukon which design the Macs 

Until one of them ports.

[bd139]

Yeah the company I just landed a job at today is doing exactly that. Another SaaS provider destined to move an entire meta function onto the web. But it’s a long struggle that change. Windows admin is big money. There are MSPs by the million taking a cut because of the management costs.

I bet we won’t see either on native Mac for at least ten years. Big chunks of rancid win32 don’t move well.

[brucehoult]

Will have native hardware and a 5900X to compare to if anyone ever get some fucking stock in 

I'd definitely be interested in Zen 3 (5900X) numbers!

I'm not sure if it's worth adding the others. Maybe the Xeon .. oh and the graviton you just added while I was composing.

Hmm .. I should note in my results the Zen 2 4500U is in WSL2.

[brucehoult]

I don't normally switch between different PCs depending on the load...   I have one workstation, it does many jobs at the same time by running many VMs, and therefore benefits from many cores.  This is probably a heavier than average workload, but that's why processors with more cores exist.

Sure, and that's why I have a monitor on the Threadripper and use it as my main machine for almost everything, even though I have three other boxes with 15% or 20% better single-core speeds (8650U NUC, Zen 2 Thinkpad, M1 Mac Mini).  20% isn't usually enough to make me care single core, and when something is parallelizable the TR just rocks. And I can run background tasks on it that use a bunch of cores and don't even affect normal use at all.

But I do things with other ISAs including ARM and RISC-V and it's nice to have real hardware that is faster than QEMU on the x86. Having ARM hardware that is JUST AS FAST (or even faster) than the x86 is super.

[brucehoult]

Edit 3: "AArch64 Processor [413fd0c1]" AWS g6g graviton 2 / AWS Linux gcc 7.3.1 (think its centos recompiled)

That's "M6g", yes?  Does /proc/cpuinfo (or otherwise) give a clock speed? Or lscpu when under load?

[bd139]

C6g which is compute optimised apparently. No MHz in cpuinfo. I’ll fire up another instance tomorrow and try lscpu.

[brucehoult]

C6g which is compute optimised apparently. No MHz in cpuinfo. I’ll fire up another instance tomorrow and try lscpu.

Internet says Graviton 2 is 2.5 GHz.  Would be good to have a cross check.

I've added that and the Xeon 8151 to my file. I've put the Xeon as 4.0 GHz as that's the turbo speed.

[Berni]

Those are some impressive numbers there. Certainly faster than i was expecting from it.

Here are some Cinebench R23 scores against some of the other popular mobile CPUs:
https://wccftech.com/intel-and-amd-x86-mobility-cpus-destroy-apples-m1-in-cinebench-r23-benchmark-results/

In terms of latest Ryzen chips one datapoint is a Ryzen 5 5600X with 6 cores doing 3.7GHz base clock and 4.6GHz single core boost clock it gets a scores:
single tread: 1572
multithreaded: 11268

So yes still not living up to the bullshit claim of "Worlds fastest CPU core". Still this is being compared to a desktop CPU with a 65W TDP while the M1 is a mobile CPU with 15W TDP. This is a very impressive amount of speed and looks like it does take the title of "Worlds fastest ARM core" and also "Worlds highest performance/watt CPU core" as i was predicting. But i do admit i was underestimating just how much speed they can get out of it. Good work apple engineers, even if the marketing department sort of buries the achievement under marketing bullshit.

[SilverSolder]

Those are some impressive numbers there. Certainly faster than i was expecting from it.

Here are some Cinebench R23 scores against some of the other popular mobile CPUs:
https://wccftech.com/intel-and-amd-x86-mobility-cpus-destroy-apples-m1-in-cinebench-r23-benchmark-results/

In terms of latest Ryzen chips one datapoint is a Ryzen 5 5600X with 6 cores doing 3.7GHz base clock and 4.6GHz single core boost clock it gets a scores:
single tread: 1572
multithreaded: 11268

So yes still not living up to the bullshit claim of "Worlds fastest CPU core". Still this is being compared to a desktop CPU with a 65W TDP while the M1 is a mobile CPU with 15W TDP. This is a very impressive amount of speed and looks like it does take the title of "Worlds fastest ARM core" and also "Worlds highest performance/watt CPU core" as i was predicting. But i do admit i was underestimating just how much speed they can get out of it. Good work apple engineers, even if the marketing department sort of buries the achievement under marketing bullshit.

Engineering team makes it to Mars -  marketing claims they reached Pluto.   It was ever thus...

[Cerebus]

Engineering team makes it to Mars -  marketing claims they reached Pluto.   It was ever thus...

[brucehoult]

I got arm64 Ubuntu running in a VM on the M1 Mac Mini.

Just off the live CD image so far. I've done an install onto another virtual disk but can't get that to boot :-(

[brucehoult]

Apparently this coordinated release is now officially a shit show. There is no virtualisation support at API level so software devs are shot. On top of that macOS now calls home to check binary signing on app startup. The cluster that handled this fell over yesterday and locked up huge numbers of people’s kit entirely. This drew attention to the fact that it can be used for surveillance. Also iMessage end to end encryption is apparently no longer appearing in documentation. On top of that people are braking NDAs and the performance is definitely exaggerated outside of synthetic benchmarks.

There are no NDAs now.

In real use performance is as good as the best Windows or Linux x86 machines. No problems. Even emulated x86 apps feel absolutely normal and snappy.

As for VMs, that's just wrong. I've been running arm64 Ubuntu in a VM on mine. Performance is just a hair from native -- I guarantee you've never seen such arm64 Linux performance elsewhere.

[brucehoult]

Many of these benchmarks are multithreaded benchmarks.  Which is great if your workload is multithreaded but so much is not multithreadable.

I have a 2019 Ryzen 3800X and its single-thread performance was only 50% faster than my 2010 Core i5.  But multithread almost 600% faster.

It looks like Apple are using big.LITTLE which is a common ARM architecture but it doesn't make a true octo-core CPU as only the big or LITTLE core in a core set can be active (presumably because they have some shared resources.)  Maybe Apple has unlocked that for specific workloads that aren't data dependent.

All 8 cores operate at the same time if there are more than 4 active threads.

To a first approximation the small cores seem pretty much as useful (or not) as hyperthreading on Intel/AMD i.e. it's like 4 cores plus HT.

I also note Apple are shy about the clock rate of their CPU - now this isn't a useful benchmark between different architectures but might indicate how far behind Apple are.  It takes a lot of knowledge to be able to get an x86 up to 4GHz.  I'd be surprised if they're cracking 2GHz.  Which will mean M1 will be exclusively for silly laptops, no one will use it in a proper workstation.

I don't really know the true clock rate. Some things report 3.2 GHz. In any case it performs like any i7 or Zen with a 4 GHz turbo, so maybe 3.2 is just the base speed, or maybe that's the 8-wide dispatch at work. I dunno.

I plan to do some serially-dependent benchmark to figure out the true clock speed e.g. a a few billion adds in a row with each depending on the previous one.

[bd139]

Excellent news all round. So do I buy a 5900X and a 3070 RTX or a Mac Mini 

My new employer has ordered me an i9 dell laptop and I get a hefty AWS budget every month to piss up the wall so I don't need the Ryzen at home now.

[brucehoult]

On benchmarks it's a little more disingenuous on Apple's part as well. It reports around the same as a Ryzen 5950X single threaded on Geekbench, not the most reliable synthetic benchmark either. But that's a pretty crap comparison because as mentioned the M1 only has 4 equal fast cores (versus 16 on the 5950X) and 4 shitty ones. The Ryzen has 16 equal cores, much larger unified caches and also when you take TDP into account, the M1 efficiency per core is still only 40% lower than the Ryzen and AMD are making reliable 20% gains per generation at the moment on performance at the same TDP and apparently are already testing parts on TSMC 5nm process.

The M1 isn't competing against the 5950x, it's competing against the four core 3100, 3300X, 4300G, 4350G, 3700U and to a large extent against the 6 core non HT such as the 4500U also.

There will be other chips with more cores later.

Also the M1 is clocking 3.2GHz base frequency. And it is 100% thermal throttled so like hell it'll be able to sustain the workloads they say it will.

I'm seeing *zero* signs of thermal throttling on my M1 Mac Mini, even after hours. It doesn't get either warm or noisy.

I get the feeling what is going to happen is this opens up a whole new benchmarking dick swinging competition rather than looking at the actual gains which so far apparently to really be a loss of privacy, decrease in overall flexibility, increase in cost and general bad story for the end user. Your computer is turned from a general purpose computer into an Apple Toaster.

I really don't know where you're getting any of this from.

I'm compiling and running my own code (and open source code) as usual. Everything runs great, whether compiled for arm64 or amd64. I'm running the standard Ubuntu arm64 ISO in a VM, with performance just a percent or two below native.

[brucehoult]

Excellent news all round. So do I buy a 5900X and a 3070 RTX or a Mac Mini 

My new employer has ordered me an i9 dell laptop and I get a hefty AWS budget every month to piss up the wall so I don't need the Ryzen at home now.

It's obviously not going to compete with a 12 core 4.8 GHz 5900X. You'd better get that or wait six or twelve months.

[tooki]

[...]
Many Mac users work on technical fields, especially on ISPs, and they rely heavily on virtualization of Intel based Linux and FreeBSD systems. Those users also need a lot of memory, more than 16 GB.

So these machines won't work for them.

[...]

So, can we think of these things as decently performing but closed appliances, suitable for not-too-challenging applications by "aspirational" users that are not too price sensitive?   

No, they're quite high-performing (for most tasks) semi-closed appliances suitable for lots and lots of challenging applications for both consumers and professionals who are willing to spend a little bit more up-front on a computer in exchange for reduced hassle during its lifespan.

I thought this thread (including the Anandtech link earlier) has established that the new Apple SoC is primarily efficient, rather than being the ultimate in absolute performance - i.e. the Intel and AMD processors it is being compared with are able to draw much more power if needed, at which point they blow the new Apple chip away? 

I'm glad there are choices available from Apple, Microsoft, and the open source community.   As soon as you give up on the idea that one PC has to do everything, life becomes much simpler (but more cluttered)! 

Apple has at no time claimed it’s the most powerful CPU out there. But what they are claiming is that it beats the fastest laptop CPUs while using a fraction of the power: https://www.apple.com/mac/m1/

They expressly say that it outperforms the (unspecified) laptop chip at all power levels. The footnote says “Comparison made against latest‑generation high‑performance notebooks commercially available at the time of testing.”

So yes, it’s incredibly efficient. But it’s not efficiency at the expense of performance, it’s efficiency leveraged to provide outstanding performance since it’s not constantly running into thermal limits or battery life considerations, like (for example) my 2012 i7 MacBook Air.

[brucehoult]

For Apple to swoop in out of nowhere and makes a ARM that beats both is pretty much science fiction.

The M1 beats the previous Intel chips Apple was shipping.

It doesn't beat the very latest Intel and AMD chips (with comparable core count), or at least not by much, and wins some benchmarks and loses some -- but not by a lot either way.

The M1 is *competitive* with the latest Intel and AMD 4 core + HT or 6 core without HT offerings. For example it's faster single-threaded and competitive multi-threaded with the 6 core Zen 2 4500U in my Thinkpad E14 Gen 2.

[tooki]

I thought this thread (including the Anandtech link earlier) has established that the new Apple SoC is primarily efficient, rather than being the ultimate in absolute performance - i.e. the Intel and AMD processors it is being compared with are able to draw much more power if needed, at which point they blow the new Apple chip away? 

Nope.

The M1 is right up there, core for core, against the biggest baddest hot running desktop 125W i9 10900K, 105W AMD Ryzen 9 5950X etc.

It's got fewer cores, of course.

So in the same sense that a 4 cylinder engine is just as powerful as a V-8, cylinder per cylinder?

There is far, far, FAR more to an engine’s performance than the number of cylinders.

Similarly, there is far, far, FAR more to CPU performance than the number of cores and the clock speed.

[tooki]

For Apple to swoop in out of nowhere and makes a ARM that beats both is pretty much science fiction.

The M1 beats the previous Intel chips Apple was shipping.

It doesn't beat the very latest Intel and AMD chips (with comparable core count), or at least not by much, and wins some benchmarks and loses some -- but not by a lot either way.

The M1 is *competitive* with the latest Intel and AMD 4 core + HT or 6 core without HT offerings. For example it's faster single-threaded and competitive multi-threaded with the 6 core Zen 2 4500U in my Thinkpad E14 Gen 2.

That’s not what Apple is saying with regards to laptops, where they’re expressly comparing it to “latest‑generation high‑performance notebooks commercially available at the time of testing.”

[brucehoult]

Geekbench is a load of shit for sure. The key thing though is measured compute performance. It appears that the answers are now in with Cinebench coming in...

I have no idea why people obsess over Cinebench. Rendering video is something I use my computer to do .. basically never.

Linux kernels builds are at least somewhat relevant to what I do with a computer that is performance-critical -- I watch a lot of youtube (or VLC) but once you're not dropping frames extra performance is irrelevant.

[bd139]

I do a hell of a lot of transcoding on mine.

As for my earlier requirement, the compute requirements of my desktop are now very low thanks to new role. 5900X would be a pointless purchase for me now.

[brucehoult]

The only remaining question is - what are the downsides?  Are there any?

Many Mac users work on technical fields, especially on ISPs, and they rely heavily on virtualization of Intel based Linux and FreeBSD systems. Those users also need a lot of memory, more than 16 GB.

So these machines won't work for them.

Virtualization of arm64 OSes works fine. I'm doing it. Look back a couple of posts. I'm running arm64 Ubuntu on my M1.

Emulation of x86 works fine, at mostly 70 to 80% of native speed. I've seen cases of up to 95% of native e.g. my primes benchmark.

Of course some people need more than 16 GB of RAM. So wait a few months.

Personally, my big Threadripper has 128 GB, but that's mainly so each of the 32 cores has 4 GB to work with. On an effectively 4 core machine 16 GB is fine for most purposes.

[Cerebus]

I have no idea why people obsess over Cinebench. Rendering video is something I use my computer to do .. basically never.

Linux kernels builds are at least somewhat relevant to what I do with a computer that is performance-critical...

I think the general public would put those two the other way around, possibly in exactly the same terms. 

[tooki]

Geekbench is a load of shit for sure. The key thing though is measured compute performance. It appears that the answers are now in with Cinebench coming in...

I have no idea why people obsess over Cinebench. Rendering video is something I use my computer to do .. basically never.

Linux kernels builds are at least somewhat relevant to what I do with a computer that is performance-critical -- I watch a lot of youtube (or VLC) but once you're not dropping frames extra performance is irrelevant.

Probably because video encoding is one of the most demanding CPU tasks. When not hardware-accelerated, it’s incredibly taxing on a system, as it involves gargantuan amounts of data and processing. (We use compressed formats that have staggering compression ratios. The raw data the encoder needs is huge, and the amount of processing needed to achieve high compression is insane.)

The typical things most people do most of the time (web browsing, watching videos, creating documents, etc) simply aren’t demanding enough to tease out meaningful performance comparisons.

I remember back in the early 90s, when computer benchmarks ran massive scripted suites of real-world software, that one common benchmark item was how long it took a spreadsheet (originally Lotus 1-2-3, later Excel) to recalculate a spreadsheet with a couple of thousand cells. (Not lines. Cells.) They had to increase the size of the spreadsheets to keep up with performance improvements, before eventually abandoning that test because all computers got fast enough that recalculating a spreadsheet became instant, so they just made it automatic as we know it today. (IIRC, we got to the point in those tests that they couldn’t even increase the test spreadsheets, as they’d reached the limits at the time of how many columns and rows you could have; that is, limits written into the code, not computing resource limits.)