Apple CPU Gains Grind to a Halt and the Future Looks Dim(semianalysis.com)
semianalysis.com
Apple CPU Gains Grind to a Halt and the Future Looks Dim
https://semianalysis.com/apple-cpu-gains-grind-to-a-halt-and-the-future-looks-dim-as-the-cpu-engineer-exodus-to-nuvia-and-rivos-impact-starts-to-bleed-in/
6 comments
> Apple advantage was in mips-per-watt. AMD wipes the floor with them in performance department
Are we talking single or multi-core? I don't think they are far in single-thread performance.
Are we talking single or multi-core? I don't think they are far in single-thread performance.
No one should care about single core performance these days. It just feels so deceptive. Workloads that are heavy are usually parallelized to use the chip fully. If performance is important, the portability and power efficiency aspect also becomes a bit contrived.
In broad strokes, a computationally heavy task that would be set up during the evening and be ready the next morning on a high end desktop wouldn't be ready on an M1 for a week.
In broad strokes, a computationally heavy task that would be set up during the evening and be ready the next morning on a high end desktop wouldn't be ready on an M1 for a week.
> No one should care about single core performance these days.
Javascript is still single core bound, and that's often a bottleneck for consumer computer use.
Javascript is still single core bound, and that's often a bottleneck for consumer computer use.
Can't it start multiple threads to take advantage of multiple cores?
I'm not sure how well browsers on MacOS utilizes cores, or if there is an artificial restriction there. At the very least it will have a separate thread for each tab.
Though, if I'm not mistaken, I do believe that Web Workers is supported on all major browsers, including Safari. This allows the main JS thread (the one per tab) to spin up additional separate threads for other workloads. This page seems to confirm this: https://caniuse.com/webworkers
Here is an example to try out: https://codepen.io/prosetech/embed/qeqGQR? (source: [1]). When I tried it in chrome (default settings on linux) with a search range of 1 to 5,000,000, it took roughly 16 seconds and overall CPU usage was around 12%. Feel free to try it on an M1 (I'm genuinely curious, since comparable benchmarks are really hard to come by). This doesn't use the GPU at all, and is already CPU resource restricted, so I would expect it to fare well.
I'm sure the browser will handle some scheduling to limit 100% usage regardless of what the OS allows it.
So, it is I said before. That it seems a bit dishonest to exclusively focus on single core performance, as if that matters more than what the CPU can do when fully utilized. And, when fully utilized, the M1 is not remarkable in the slightest. It was a seriously misleading and dishonest marketing campaign. The only exception is, power consumption, which is pretty good. I find the idea of needing a lot of computing power, while also being limited by battery life, to be a contrived scenario. Wouldn't you be plugged in? Even the trains I've been to in recent memory have had power outlets.
With all that said, I think it's silly to consider javascript performance to be all that relevant. I hope this doesn't come across as a gatekeeping point of view. But when I talk about computing power, I don't really think about whether a web-page loads in 100 ms or 110 ms. But rather of machine learning tasks, photoshop filters. A software compilation. A blender render. Whether these things take 14 hours, or 1 hour. Which is what you can expect to be the difference in performance when fully utilizing a top end desktop GPU+CPU vs the M1. A practical difference of a "see you tomorrow morning", vs "see you in a week".
[1]: https://medium.com/young-coder/a-simple-introduction-to-web-...
Though, if I'm not mistaken, I do believe that Web Workers is supported on all major browsers, including Safari. This allows the main JS thread (the one per tab) to spin up additional separate threads for other workloads. This page seems to confirm this: https://caniuse.com/webworkers
Here is an example to try out: https://codepen.io/prosetech/embed/qeqGQR? (source: [1]). When I tried it in chrome (default settings on linux) with a search range of 1 to 5,000,000, it took roughly 16 seconds and overall CPU usage was around 12%. Feel free to try it on an M1 (I'm genuinely curious, since comparable benchmarks are really hard to come by). This doesn't use the GPU at all, and is already CPU resource restricted, so I would expect it to fare well.
I'm sure the browser will handle some scheduling to limit 100% usage regardless of what the OS allows it.
So, it is I said before. That it seems a bit dishonest to exclusively focus on single core performance, as if that matters more than what the CPU can do when fully utilized. And, when fully utilized, the M1 is not remarkable in the slightest. It was a seriously misleading and dishonest marketing campaign. The only exception is, power consumption, which is pretty good. I find the idea of needing a lot of computing power, while also being limited by battery life, to be a contrived scenario. Wouldn't you be plugged in? Even the trains I've been to in recent memory have had power outlets.
With all that said, I think it's silly to consider javascript performance to be all that relevant. I hope this doesn't come across as a gatekeeping point of view. But when I talk about computing power, I don't really think about whether a web-page loads in 100 ms or 110 ms. But rather of machine learning tasks, photoshop filters. A software compilation. A blender render. Whether these things take 14 hours, or 1 hour. Which is what you can expect to be the difference in performance when fully utilizing a top end desktop GPU+CPU vs the M1. A practical difference of a "see you tomorrow morning", vs "see you in a week".
[1]: https://medium.com/young-coder/a-simple-introduction-to-web-...
Both
the architecture was mostly created by ARM
What do you think this means?
What do you think this means?
The article claims 100's of Apple's CPU engineers have left to join other start-ups. And thus Apple's ARM cpu development is stuck. If true, this is bad news for Apple as they have bet their entire future on ARM cpus. It'll certainly be interesting to see if people are still attracted to the next generation of ARM based Apple laptops and desktops if they aren't as good as Intel / AMD CPUs, given the current limitations of Apple ARM cpus (non-upgradeable SoC that fully supports only one OS). (But I find it hard to believe that Apple doesn't have some contingency plan for this - like I said, Apple has bet their future on ARM processors, and the tightly closed systems that they can build with it is every Apple CEO's wet dream).
This is exactly the kind of risk people warned about.
You cannot do everything in-house and think you can always be better than every other company in the world. One bad year in one department and your competitors will eat your market share.
You cannot do everything in-house and think you can always be better than every other company in the world. One bad year in one department and your competitors will eat your market share.
No, this is exactly the opposite. If you lock in your users in an ecosystem that's compatible with exactly nothing outside, they will happily eat your marginal updates for years. Even if problems would endure, Apple can plan calmly, as the switch would be much more costly for its customers.
I'd say it's too soon to call the updates marginal. This is a process improvement with some, small, architectural improvement. You can't pull an M1-level improvement on every generation.
They managed to do that the first time for about 20 years, then had to be rescued.
Now they have enough money to do that more than just 20 years, but nothing lasts forever.
Now they have enough money to do that more than just 20 years, but nothing lasts forever.
With the cash Apple is holding, I have zero doubts they will be perfectly fine.
>this is bad news for Apple as they have bet their entire future on ARM cpus
I'm sure Qualcomm & Co. would happily help them out if they're in a bind with their own design.
I'm sure Qualcomm & Co. would happily help them out if they're in a bind with their own design.
I meant their own ARM cpus. If they have to turn to someone else they lose the price advantage, in which case it doesn't make sense to run their own CPU division to manufacture CPUs.
> non-upgradeable SoC
That doesn't seem to be slowing down sales of laptops and minis, or phones. Memory is tightly packed, but that's for SoC size - the processor tile talks to the memory via a DDR4 bus on the SoC board and can easily talk to memory across the motherboard. I'd bet on that for future ARM-based MacPro's and upgradable iMacs.
Making a socketed M1/M1X/M2/M2X is a matter of packaging, IF Apple wants it: they are not in the general CPU market. There may be some market in upgrading CPUs in the MacPro's, but I doubt it'd justify the investment.
> that fully supports only one OS
That's not really a problem for Apple - they chose to base all their offerings on a portable OS for a reason. The current system we call MacOS was ported to run on 68K, x86 (under NeXT for 32-bit, under Apple 64-bit), HP-PA, SPARC, PPC, ARM (both 32 - as iOS - and 64-bit).
All in all, the roadmap for beefier M's is clear - a wider backend, ISA extensions for common workloads (one of the reasons the M1 performs so well), and a larger reorder buffer (another reason the M1 performs so well) to keep the wider backend well fed.
That doesn't seem to be slowing down sales of laptops and minis, or phones. Memory is tightly packed, but that's for SoC size - the processor tile talks to the memory via a DDR4 bus on the SoC board and can easily talk to memory across the motherboard. I'd bet on that for future ARM-based MacPro's and upgradable iMacs.
Making a socketed M1/M1X/M2/M2X is a matter of packaging, IF Apple wants it: they are not in the general CPU market. There may be some market in upgrading CPUs in the MacPro's, but I doubt it'd justify the investment.
> that fully supports only one OS
That's not really a problem for Apple - they chose to base all their offerings on a portable OS for a reason. The current system we call MacOS was ported to run on 68K, x86 (under NeXT for 32-bit, under Apple 64-bit), HP-PA, SPARC, PPC, ARM (both 32 - as iOS - and 64-bit).
All in all, the roadmap for beefier M's is clear - a wider backend, ISA extensions for common workloads (one of the reasons the M1 performs so well), and a larger reorder buffer (another reason the M1 performs so well) to keep the wider backend well fed.
> That doesn't seem to be slowing down sales of laptops and minis
It did and that's why Apple themselves had to backtrack when the Mini with the soldered RAM and SSD failed to sell as much as anticipated. (That's why the next version had swappable RAM, and not soldered).
All things being equal, people will always choose the device that is more easy to repair and upgradeable.
Apple has 2 things going for the M1 devices:
- They are priced cheaper to make the hardware attractive.
- They offer similar performance to AMD CPUs (with the advantage of lower power consumption.)
Both these factors help overcome the obvious deficiency - that the hardware is not upgradeable or repairable and the system software is limited to one OS. If Apple's ARM processors cannot keep up with AMD / Intel / other ARM cpus, both hardware and software limitations become glaring. And Apple is left with only the price advantage. (And we all know that Apple doesn't play that game as it isn't interested in the low margins market).
It did and that's why Apple themselves had to backtrack when the Mini with the soldered RAM and SSD failed to sell as much as anticipated. (That's why the next version had swappable RAM, and not soldered).
All things being equal, people will always choose the device that is more easy to repair and upgradeable.
Apple has 2 things going for the M1 devices:
- They are priced cheaper to make the hardware attractive.
- They offer similar performance to AMD CPUs (with the advantage of lower power consumption.)
Both these factors help overcome the obvious deficiency - that the hardware is not upgradeable or repairable and the system software is limited to one OS. If Apple's ARM processors cannot keep up with AMD / Intel / other ARM cpus, both hardware and software limitations become glaring. And Apple is left with only the price advantage. (And we all know that Apple doesn't play that game as it isn't interested in the low margins market).
> - They offer similar performance to AMD CPUs
Apple doesn't use AMD and there is no supported AMD machine that can run macOS.
I like my AMD-based Linux box, but, if I want macOS, that's not going to solve my needs (it won't, for instance, run Xcode)
Apple doesn't use AMD and there is no supported AMD machine that can run macOS.
I like my AMD-based Linux box, but, if I want macOS, that's not going to solve my needs (it won't, for instance, run Xcode)
Apple's loss, I'd say. AMD Macs would sell like hot cakes.
Grinds to a halt? The M1 was quite possibly the best consumer ARM cpu in the past year and CPU development has just stopped? M1X/M2 is expected later this year, wait until then before we all decide Apple is "doomed." Could it be that the phone is getting more incremental boosts in a time where the Mac/iPad is getting the real improvements?
I think its much more likely that the effects of COVID measures are just now being felt by many corporations both in the supply chain as well as product development.
I think its much more likely that the effects of COVID measures are just now being felt by many corporations both in the supply chain as well as product development.
The M1 is essentially a large version of the A14 chip and uses the same firestorm and icestorm cores. With such unsubstantial improvements in the A15, the M2 should theoretically also have unsubstantial gains because it will be a large version of A15. The article also mentions that some key technical staff jumped ship so that might somehow halt innovation.
How substantial really depends on the workload. It's small in single-thread performance, but it's easy to grow it with more cores and I'd love to see a lot of background processes moving to more icestorm cores to keep the firestorm ones free.
While IPC is probably not a huge bump, the brute force of doubling the System Cache should yield decent improvements in a lot of workloads on the laptop side.
Could it be that rivals are behind these negative articles and trying to drum up negative PR by spinning them and not telling the whole truth?
Could be. But then it should be easy to verify the facts - did 100's of Apple engineers, along with senior ones, leave Apple CPU division? And does the Apple ARM division suffer because of this talent drain? The whole article makes assumptions based on these facts. Perhaps the assumptions may be exaggerated ... but if people working on the M1 left Apple, it would certainly be an issue for Apple.
Its not just the engineers leaving, the newer chips will also be produced on n4 instead of n3.
Appears to be a dupe of https://news.ycombinator.com/item?id=28536048
Found where the Intel PR people are.
It's kinda weird to make such strong claims based on a processor which nobody has used, based on vague numbers in the product announcement for a low-end iPad. Is it that the new iPad mini's A15 is comparable to an A14 because the A15 isn't any more powerful, or is it because Apple decided how powerful they wanted the iPad mini to be, and slowed down the processor to that, to conserve power. Who even knows?
I don't think this is event remotely correct to begin with (Apple advantage was in mips-per-watt. AMD wipes the floor with them in performance department, but also uses more power).
And the author doesn't seem to understand that the architecture was mostly created by ARM.
With a start like that, I am not sure if I should read the rest.