Apple M1 ARM Performance With A 2020 Mac Mini

[marvin42]

Michael, maybe you could try virtualized Linux next time: https://mobile.twitter.com/KhaosT/st...36063990190085

[deppman]

I've never been a fan of Apple software, and for 25 years i was enjoying building PCs. I'm shocked what Apple has just delivered. I'd really like them to not be evil, and to not only open the platform to Linux, but actively support Linux, Windows and other 3rd party developers to quickly onboard software to new architecture. Maybe it is time to let go fun of tinkering with hardware, a bit sad for me, but at the end it is performance/efficiency which matters. How comfortable the workflow is. I'm wondering about gaming community, does Apple have any strategy for gaming market?

ARM performance for general-purpose computing is much more likely to come from Nvidia or Qualcomm as discussed below.

II think we are all in shock, probably even Apple is to some extent. They developed Wunderwaffe, hopefully they have a wisdom on how to use it to make the computing market better, not only to monetize on what they did. In 10 years from now the world could be running on Apple Silicon, and it takes great responsibility from Apple not to abuse their monopoly. Anyway big thanks to engineers and brave souls at Apple who had given us something amazing after decade of stagnation on x86 market.

The "brave souls" also include those from Qualcomm, Nvidia, HP, SGI, Sun, HP, and others that have pioneered RISC architecture. Apple is just one of many of these pioneers. Qualcomm and Nvidia have been producing impressive ARM SoCs for years with fantastic power, GPU, and CPU performance like the QC 865 or Nvidia Xavier. The M1 is just another step in this evolution.

Apple hasn't given anyone anything, and unless they change their ways, they aren't going to either. Unless I missed something, they will be pushing their M1 products much more toward iPhone-like closed appliances than general-purpose computing. If you looking to "upgrade" your walled garden to a padded cell, my guess is Apple is ready for you.

[bridgman]

And you know what amazes me most? That Apple had developed it for years, Intel and AMD engineers must knew about it. And they did nothing. Did they really think they can make money on 1-5% improvement per year? They were lazy, conservative, thinking they can sit in comfort zone with mocked competition. They deserve this fate.

We were closer to 19% per year, weren't we ? Much higher if you include increased core counts.

[F.Ultra]

No way, IMHO he is right. If you take a look to anandtech benchmarks, emulated software runs roughly ~75% the performance of native via Rosetta 2 (source: https://www.anandtech.com/show/16252...le-m1-tested/6), considering there is a software emulation mechanism in the middle, this is just huge result

Rosetta is not an emulator, it's a translator. 75% for that is good but not huge.

[Space Beer]

I am quite sure AMD or Intel could make similar chip with simalr peformanse. As already mentioned, they had some interesting designs before (HSA, eDRAM, HBM, big.little arch, even x86-ARM hybid)

[Space Beer]

The difference is, they don't have market for such product. They can not force OEMs to implement specific design, they can not force OS manufacturers to make special version optimized for such chip, they can not force deveoplers to use all the features os such chip and accelerators. Intel could throw some more money for some optimizations (like the did with some Adobe programs), but that's it. Or make Linux distro that not so many people use (Clear Linux)

Apple products are designed for specific users groups. PCs are designed for everyone. Which also means no one in particular. Until OEMs realize this, work with AMD/Intel/nVidia to make semi-custom designs, work with MSFT/Canonical/Red Hat... to make semi-custom OS, work with Autodesk/Adobe/Dassault/JetBrains/Oracle... to optmize their products for such devices, and actually make products for specific user groups, they cant only hope that people who would buy Mac, also need some features Apple is not interested to add

[artivision]

Fore those who cannot count cores, Apple's benchmarks come from 4 cores and not 8. Big out of order cores cannot be used along with small in order cores. Even if they did, those small 4 cores together will just par on big core. Apple's processor is the best commercial processor SOC by far.

[geearf]

The key difference here is the microarchitecture of M1 which is incredibly fast. The simpler Arm ISA has fixed-length RISC instructions which allows 8-wide decode. This graph shows the steady progression of Apple's microarchitectures, showing clearly x86 has run out of steam.

No, both static and dynamic translation always add overheads. Compilers keep improving, but they are already very good, and there isn't that much opportunity to do better. Also many improvements help all ISAs, so it wouldn't close the gap.

I am sorry if I misunderstand, but I thought the post I was replying to said there was already translation from CISC to RISC, happening directly on the CPU. Was that wrong?

Thank you!

[deppman]

Have you been reading this thread? Sure the M1 doesn't win everything, but it's drawing comparisons between it and crazy more expensive chips that use many times more power. Personally my interest is in floating point. Take a look at the Anandtech chart reposted at the bottom of page 13 on this thread. The M1 got 38.71. Keep in mind the M1 has a GPU, neural processor, and image processing acceleration and consumes around 10 watts, possibly as high as 20 when actively cooled. The ryzen 3600 has no GPU, consumes way more power and gets 33 at FP. The intel I7-10700k has no GPU, is a high end chip and gets 34.98. The ryzen 7 3700x, one of the best desktop chips for the money, out from 6 months ago finally pulls ahead, barely at 40.20 by 3.8%.

I've read lots of the thread and can discern hyperbole from fact as well. As a system integrator and veteran user of Risc Unix workstations and Linux ARM deployments using Nvidia Jetson boards, I probably have a little more insight than the average bear.

First, the M1 doesn't win a lot in the areas that really count. It doesn't appear you can add an eGPU, and even if it could, Apple refuses to sign Nvidia drivers, so the industry standard and usually much faster CUDA, OptiX, and cuDNN libraries are off the table. You can't expand the RAM to 72GB system r 64GB + 8GB dGPU like other pro workstations - you're stuck to 22% of that. It runs the abomination that is MacOS. Lock-in is likely to only going to get worse, as "Microsoft has to ... bring to license that technology for users to run on these Macs". That means one nice, tight, closed ship.

Second, the CPU performance is impressive but not earth-shattering. The i7-10875H CPU is available in the same price envelope, but it beats the M1 multi-core by around 20%, although the M1 does better by around 20% in single-core, so let's call it a draw -- and that's a 10th gen part. In real-world use, the power deficit of the x86 is far less than Apple purports: at idle it runs less than 6W (M1 runs 4W) and boosts to around 20W for most mobile desktop tasks (I know, I've optimized for the chip extensively). It's only when you're doing heavy compute where power draw gets dramatic and you're likely going to be plugged into the mains for that on either platform. Assuming you can even do that with the M1 since with only 16GB of total shared RAM you're not going to be able to do much 8K video editing at all.

The iGPU is impressive for an iGPU, around 20% slower than a GTX 1650 (don't be fooled by the ice storm outlier). But if you need more power you can't go there - there's no eGPU, Nvidia, or AMD option. It is still 500% slower than a mobile RTX 2080 for deep learning and gets destroyed by an RTX 2060. The M1 WILL draw far greater than "10-20W of power" for this performance as CPU alone use exceeds 20W. <del>Expect GPU heavy loads to run around 40W and 60W with CPU use</del>. EDIT/ It turns out my GPU estimates were way off and that total power draw is limited to perhaps 32W. That is indeed excellent results for the GPU. /EDIT

These still are great numbers, just not the "walk on water" numbers Apple would have its fans believe.

In summary, these are the reasons many Mac products are still on Intel. I'm by no means an Intel or Apple fanboy. But the reality is that much of the improvements are evolutionary, and there are lots of regressions as well. If you insist on Blue Bubbles and MacOS to impress your friends, then it might be a good upgrade. Personally, I'm hoping this will spark more open ARM initiatives from the likes of Nvidia, who have ironically support Linux on Jetson for 7 years now.

[duby229]

The key difference here is the microarchitecture of M1 which is incredibly fast. The simpler Arm ISA has fixed-length RISC instructions which allows 8-wide decode. This graph shows the steady progression of Apple's microarchitectures, showing clearly x86 has run out of steam.

No, both static and dynamic translation always add overheads. Compilers keep improving, but they are already very good, and there isn't that much opportunity to do better. Also many improvements help all ISAs, so it wouldn't close the gap.

OK, so nobody really liked AMD's CMT architecture that they implemented in Bulldozer and its derivatives, it was designed to be 4-wide decode with two integer units on each of two pipelines. But the important thing to note is that architecture is fully capable of 8-wide decode with 4 integer units on each of two pipelines. I don't think it's a stretch of the imagination to assume that todays 7nm process is far superior to the 32nm process Bulldozer was fabricated on. A modern CMT architecture could definitely be the solution.

I still have faith in it, it is by far -the- most scalable x86 architecture ever conceived. x86 definitely hasn't run out of steam, it just needs to be re-imagined using lessons learned from BD -and- Zen.