Hacker News Board Hacker News, browsed like a classic message board
Live data from Hacker News

Apple M1 Max Geekbench Score

browser.geekbench.com

Post reply on HN
Print view View original

551–560 of 917 posts

Re: Apple M1 Max Geekbench Score

#551

Post by PaulDavisThe1st »

josephg wrote: post #481
PaulDavisThe1st wrote: post #37

Earlier quoted context omitted.

Not when building large software projects.

> Not when building large software projects. While working in rust I am most limited by single core performance. Incremental builds at the moment are like, 300ms compiling and 2 seconds linking. In release mode linking takes 10+ seconds with LTO turned on. The linker is entirely single threaded. Fast cold compiles are nice, but I do that far more rarely than incremental debug builds. And there’s faster linkers (like…

My project [0] is about 600k lines of C++. It takes about 5m40s to build from scratch on a Ryzen Threadripper 2950X, using all 16 cores more or less maxed out. There's no option in C++ for meaningiful incremental compiles. Typically working compiles (i.e. just what is needed given whatever I've just done) are on the order of 5-45 secs, but I've noticed that knowing I can do a full rebuild in a few minutes affects my development decisions in a very positive way. I do 99% of my development work on Linux, even though the program is cross-platform, and so I get to benefit from lld(1).

The same machine does nychthemeral builds that include macOS compiles on a QEMU VM, but given that I'm asleep when that happens, I only care that the night's work is done before I get up.

[0] https://ardour.org/

Re: Apple M1 Max Geekbench Score

#552

Post by sudosysgen »

r00fus wrote: post #416
muro wrote: post #352

Earlier quoted context omitted.

Performance doesn't go down linearly with power. I don't have that card to try, but maybe it would do even better at 15% of power.

There is a floor below which it will go straight to 0 (ie, nonfunctional). The M1 Max is well below that floor at max wattage.

That floor is really, really low when you try to optimize for it. But the issue is that the M1 Max is at 5nm TSMC while the 3080 is on 8nm Samsung, so this isn't actually that impressive.

Re: Apple M1 Max Geekbench Score

#553

Post by onepointsixC »

spacedcowboy wrote: post #317
jjcon wrote: post #284

Earlier quoted context omitted.

> GPU capabilities that blow everything away Compared to previous macs and igpus - an nvidia gpu will still run circles arounnd this thing

Anything that can run rings around it is unlikely to be running on a battery in a laptop, at least for any reasonable length of time.

Not to mention will stretch beyond reason the idea of a "mobile device" with an unwieldily weight and thickness.

Re: Apple M1 Max Geekbench Score

#554

Post by paulmd »

ChrisMarshallNY wrote: post #522
blakesterz wrote: post #100

I just can't figure out what I'm missing on the "M1 is so fast" side of things. For years I worked* on an Ubuntu desktop machine I built myself. Early this year I switched to a brand new M1 mini and this this is slower and less reliable than the thing I built myself that runs Ubuntu. My Ubuntu machine had a few little issues every no and then. My Mini has weird bugs all the time. e.g. Green Screen Crashes when I have…

I thought native M1 Linux was still a work in progress ( disclaimer : I’m not a Linux geek). Even if it’s out there, I suspect it’s not optimized. I think standard distros are running on Rosetta2, which is an (excellent, from all reports) X86 emulator.

linux is actually ported to the M1 at this point and you can run it normally (although of course there are likely still some bugs lurking), however the graphics drivers are still a work in progress although coming along very quickly.

https://www.theregister.com/2021/10/06/asahi_linux_m1_progre...

I'm sure you're right that there are still optimizations to be done, but as you can see from Alyssa's tweet there, it is still a very fast piece of hardware for normal developer usage, as long as you don't fall into one of the scenarios that probably requires further tuning (x86 emulation probably being an obvious one, no idea if they've got Total Store Ordering working in KVM yet but that's obviously a key feature).

mega kudos to the Asahi team, the progress has been really tremendous, it's been just about 6 months since kickoff and they've made progress that some people insisted would take multiple years/was impossible at all.

Re: Apple M1 Max Geekbench Score

#555

Post by sudosysgen »

Closi wrote: post #308
yazaddaruvala wrote: post #275

Earlier quoted context omitted.

Its not clear to me that Apple will make a desktop-class processor. The unit economics likely don't make sense for them. All of Apple's innovation seems to be towards better and cheaper AR/VR hardware. Desktop-class processors would be a distraction for them. And with all of the top cloud players building custom silicon these days, there is little room for Apple to sell CPUs to the server market even if they were inc…

IMO it's obvious that there will need to be a desktop version - and all the rumours are pointing towards a Mac Pro release with silly specs - i.e. an SOC with something like 40/64+ cores. Why would Apple want to give up their portion of the high-power desktop market to Windows? What's the alternative? That they release another Mac Pro with intel, despite their stated intention to move everything away from x86, or tha…

Well, those are rumors assuming a chiplet architecture, which Apple has never tried before and would require very significant modifications to the layout.

Simply quadrupling the die isn't really feasible, the price increases exponentially.

Re: Apple M1 Max Geekbench Score

#556

Post by zbrozek »

grishka wrote: post #485
ksec wrote: post #167

I have seen far too many people making comments on MacPro "Pro" Chip. A hypothetical 32 Core CPU and 64 Core GPU is about the max Apple could make in terms of die size reticle limit without going chiplet and TDP limit without some exotic cooling solution. Which means you cant have some imaginary 64 Core CPU and 128 Core GPU. We can now finally have a Mac Cube, where the vast majority of the Cube will simply be a heat…

> about the max Apple could make in terms of die size reticle limit without going chiplet I probably have no idea what I'm talking about as I'm a software guy. But — I remember that company that made that crazy ML accelerator where an entire silicon wafer is one chip. How'd they do that? Why can't Apple and others do the same/similar?

Cerebras

I'd bet it's the plumbing. Getting data in and out, supplying something around 10-100 kiloamps, attaching it to anything without it ripping itself off via thermal expansion, getting the heat out - all of that sounds miserably difficult.

Like sure you can do some outrageous and expensive things to make it all work in small quantities and for huge sums of money. But you can't build a mass-market laptop that way.

I'd guess that the upper limit with present-day and near-future packaging technology for commodity hardware is ~1000 mm^2.

Re: Apple M1 Max Geekbench Score

#557

Post by steelframe »

jandrese wrote: post #336
jacurtis wrote: post #254

Earlier quoted context omitted.

I have both, and use both everyday. I use an M1 MacBook Pro (16Gb Ram) for personal projects and as my standard home/travel computer. It is amazing and fast. I use a Lenovo Carbon X1 Laptop with similar specs (i5, 16Gb Ram, m.2 ssd) for work that runs RHEL 8 (Red Hat Enterprise Linux). It's insanely fast and stable. The overhead to run RHEL is so small it would blow your mind at the performance you get from almost no…

Downside of RHEL is the package repo is anemic and out of date. Sometimes horribly out of date. It's hardly uncommon to run into some issue with an application and then look it up online and find out that the fix was applied 8 versions after the one that's in the repo. Worse is when you start grabbing code off of Git and the configure script bombs out because it wants a library two versions ahead of the one in the re…

I have had luck running Debian Stable with apps from Snap when the older .deb from the repo doesn't cut it. I haven't found that setup at all complicated.

Re: Apple M1 Max Geekbench Score

#558

Post by lanza »

PragmaticPulp wrote: post #245
blakesterz wrote: post #100

I just can't figure out what I'm missing on the "M1 is so fast" side of things. For years I worked* on an Ubuntu desktop machine I built myself. Early this year I switched to a brand new M1 mini and this this is slower and less reliable than the thing I built myself that runs Ubuntu. My Ubuntu machine had a few little issues every no and then. My Mini has weird bugs all the time. e.g. Green Screen Crashes when I have…

> I just can't figure out what I'm missing on the "M1 is so fast" side of things. Two reasons: 1. M1 is a super fast laptop chip. It provides mid-range desktop performance in a laptop form factor with mostly fanless operation. No matter how you look at it, that's impressive. 2. Apple really dragged their feet on updating the old Intel Macs before the transition. People in the Mac world (excluding hackintosh) were stu…

You're missing out on the fact that Apple didn't release a 12980hk or 5980hx competitor. These are 30 watt chips that trounce the competition's 65 watt (e.g. the 12980hk and 5980hx) and beyond chips.

Hell, this Geekbench is faster than a desktop 125 watt 11900k. It's faster than a desktop 105 watt 5800x.

Apple intentionally played to the competition here. They know AMD/Intel reach some performance level X and released CPUs that perform no greater than X * 1.2. They know they are in the lead since they are paying TSMC for first dibs on 5nm, but they didn't blow their load on their first generation products.

Intel will release Alder Lake and catch up, AMD will reach Zen4 and catch up and Apple will just reach into their pocket and pull out a "oh here's a 45 watt 4nm CPU with two years of microarch upgrades" and the 2022 MBP 16 will have Geekbench scores of ~2200 and ~17000.

There's a de facto industry leader in process technology today -- TSMC. Apple is the only one willing to pay the premium. They also have a much newer microarch design (circa 2006ish) vs AMD and Intel's early 90s designs. That's a 10-20% advantage (very rough ballpark estimate). The also are on arm which is another 10-20% advantage for the frontend.

The big deal here is that this isn't going to change until Intel's process technology catches up. And, hell, I bet at that point Apple will be willing to pay enough to take first dibs there as well.

AMD will never catch up since we know they don't care to compete against Apple laptops and thus won't pay the premium for TSMC's latest art. Intel might not even care enough and let Apple have mobile/laptop market first dibs on their latest node if Apple is willing not to touch the server market. Whether or not they'd agree on the workstation MacPro vs 2 slot Xeon workstation market would be interesting.

It might be a long time before it makes sense to buy a non-Apple laptop.

Re: Apple M1 Max Geekbench Score

#559

Post by tcmart14 »

Master_Odin wrote: post #502
johncolanduoni wrote: post #187

Earlier quoted context omitted.

But if you run x86 images, it will use qemu’s software emulation inside the Linux VM (which is quite slow).

This has been the big slowdown for my org as a number of packages we rely on in docker on debian do not yet have an arm64 version, so have to emulate x86 and take the huge performance hit.

Yea, don't get the benefit of the hypervisor when constrained to x86 Vms.

Re: Apple M1 Max Geekbench Score

#560

Post by paulmd »

throwawaylinux wrote: post #538
paulmd wrote: post #529

Earlier quoted context omitted.

also bear in mind that AMD's standards for these "challenges" have always involved some "funny math", like their previous 25x20 goal, where they considered a 5.02x average gain in performance (10x in CB R15 and 2.5x in 3D Mark 11) at iso-power (same TDP) to be a "32x efficiency gain" because they divided it by idle power or some shit like that. But 5x average performance gain at the same TDP doesn't mean you do 32x a…

Wow that's stupid, I didn't look that closely. So it's really a 5x perf/watt improvement. I assume it will be the same deal for this, around 5-6x perf/watt improvement. Which does make more sense, FP16 should already be pretty well optimized on GPUs today so 30x would be a huge stretch or else require specific fixed function units.

it's an odd coincidence (there's no reason this number would be related, there's no idle power factor here or anything) but 5x also happens to be about the expected gain from NVIDIA's tensor core implementation in real-world code afaik. Sure they advertise a much higher number but that's a microbenchmark looking at just that specific bit of the code and not the program as a whole.

it's possible that the implication here is similar, that AMD does a tensor accelerator or something and they hit "30x" but you end up with similar speedups to NVIDIA's tensor accelerator implementation.

Post reply on HN