Hacker News Reader

Comment by s3cur3 | Hacker News Reader

It's not just "ARM is fast," though—Apple's ARM chips in particular beat the pants off their ARM competitors. (See benchmarks of any iPhone compared to flagship Android devices.) I don't think you should expect similar levels of performance from a run-of-the-mill ARM chip on the desktop (though I'm happy to be shown benchmarks to the contrary).

View on HN

Probably because there's basically no competition in the ARM space, especially in non-mobile. I'd like to see what amd and intel can do with ARM before saying apple will continue to be king.

Qualcomm, Samsung, Cavium/Marvell, and nVidia don't count? Not to mention ARM's own cores...

They don't count in the PC space. For desktop PCs there's just the Mac mini and various hacker boards. For laptops there's the new MacBooks and the mediocre Qualcomm 8cx/SQ2 and nothing else.

There's Exynos.

Exynos is dead.

No, there aren't any PCs running Exynos.

That's true if you consider Chromebooks not to be PC-class machines.

For other ARM manufacturers, I wonder how much of this is a software problem.

I mean, what would the market be for a super-fast ARM laptop/desktop machine? It would run Windows RT, I guess?

How much better than Intel/AMD would the ARM machine need to be to justify buying it and running that? It seems like the M1's "magic" is not just the chip, but very much also the Rosetta thing.

Windows RT doesn't exist anymore. Now its Windows ARM, which is basically 100% fully featured Windows with the exception that x86 apps have to be run in emulation and x86-64 apps don't run at all (yet).

> For other ARM manufacturers, I wonder how much of this is a software problem.

IMO, one of the key innovations Apple made in the M1 is a modification to the memory model so that it is much more efficient to translate x86 code into ARM code.

No one said "in the PC space" here.

There's lots of credible ARM work in both the mobile and the server space. Generally, a lot of it would work OK in the PC space, but for the lack of software support. Apple has the ability to leapfrog that particular problem.

Still, Apple seems to have done pretty well compared to both the other mobile and server ARM vendors with their PC offering.

This whole article and thread are about the desktop (not server) space.

I do agree that Apple's push to ARM for laptops will have a dramatic impact on the server space in a year or two. Intel is in for a world of hurt unless they can pull a rabbit out of their hat.

Dunno, seems more like now. Amazon has arm instances, there's a bunch of arm systems in the server space already and the world's fastest HPC cluster is already arm.

Samsung stopped Exynos development a year ago. Cavium is basically dead (yes, bought out by Marvell). Qualcomm is interested only in mobile SoCs at mobile SoCs volumes where they can also sell radios, and Nvidia was traditionally not capable of delivering anything at promised power consumption.

Are you sure about Exynos? They announced the 1080 just a couple of weeks ago and the 2100 is expected next year.

This is from wikipedia (under "Exynos"), so apply the appropriate amount of salt:

On October 1, 2019, rumors emerged that Samsung had laid off their custom CPU teams at SARC.[21][22][23] On November 1, 2019, Samsung filed a WARN letter with the Texas Workforce Commission, notifying of upcoming layoffs of their SARC CPU team and termination of their custom CPU development.[24] SARC and ACL will still continue development of custom SoC, AI, and GPU.[25]

You misunderstood what that means. Samsung is continuing to develop Exynos SoCs using stock Arm CPU cores and AMD GPU cores. This will require fewer people hence the layoff.

Exactly; they do not have their cores anymore, they are going to use ARM's Cortex. So their SoC will be always average at best, by definition. They won't be able to make anything exceptional.

HiSilicon's Kirin chips use Cortex microarchitecture and are very competitive. I didn't follow recent releases, but I remember that the Kirin 970 was way ahead of the Snapdragon and Exynos chips at the time, both in terms of performance and power efficiency.

Huawei often releases new Kirin SoC on October with Mate series meanwhile Qualcomm reseases flagship Snapdragon 8xx on December, so Kirin outperform latest Snapdragon that released on last year.

Samsung stopped developing their custom cores. They will still continue making Exynos, but with reference design ARM cores.

Yes, and their SoCs will be as average as Cortex reference implementation can be. They won't be able to make M1-like SoC.

MediaTek 6nm chips next year https://www.gizmochina.com/2020/09/10/mediatek-to-announce-6...

And Fujutsu as well, I think they are making the fastests chips at the moment, or at least the fastest systems with ARM.

Apple's ARM SoCs are definitely ahead, but it's not like the competition is hugely far behind. Top of the line Android phones and tablets perform well enough for most desktop tasks IMO. It's mostly the software and I/O that makes them unusable as such.

The numbers say otherwise.

The numbers show that the other chips are slower, but they don't show that they are "unusable".

x86 chips aren’t “unusable” either, and if these other ARM chips wanted to compete in the desktop space they’d be just as - or more - “hot and/or slow, clunky, noisy”. ISA just doesn’t matter half as much as microarchitecture.

And those x86 are only still usable because M1 hasn’t been out long enough for the new software, that needs the speed to be “useable”, to be written.

5nm process would help a lot.

Put it this way: if the new Macs had the performance of the fastest Qualcomm ARM cores available today, we wouldn’t be having this conversation now.

To be clear I'm not talking benchmarks, I have no idea what the raw numbers are. I'm talking about my personal experience using ARM SoCs. Specifically I've worked with Tegra chip and they're powerhouses for instance, I could definitely see myself do my day-to-day development on such a chip.

When iPhone came out in 2007, Android was still an OS for BlackBerry like devices. It didn't take long for Android to discover touch screens once Apple made it clear that they were dramatically better. Similarly, it's not like other ARM chip manufacturers are going to have to go through all of the same R&D expense to figure out how to make chips that are comparably fast--they just need to copy what Apple did without infringing on any particular patents. Even if they did have to spend the same R&D expense to get there, Apple has already derisked it. So they have a small fraction of Apple's expense and none of the risk, so I would fully expect to see M1 competitive chips come out in a few years' time. This isn't to say that they'll catch up with Apple completely; only that they'll be able to recoup a significant amount of the gap in a relatively short period of time such that there will be a similar breakthrough in the PC market (this is effectively what TFA is predicting as well, if that lends me any credibility).

If it were so easy, why are all of the other ARM manufacturers years behind?

In my opinion, Apple "derisked" high-performance ARM chips with the A9-ish era five years ago. Other companies have had a long time to spend catching up to Apple mobile chips, and they haven't.

That's a completely reasonable challenge, and I might be wrong. My very amateur speculation is that, in the mobile space, Apple got incrementally better over time. Rank-and-file Android users didn't really notice the performance gap and thus didn't demand faster chips. M1 changed expectations for the laptop market virtually overnight, and rank-and-file users will expect M1 performance, battery life, etc. I think this represents an enormous market opportunity for PC manufacturers and thus chip suppliers that didn't really exist circa A9.

My key assumption is that there's some pareto principle at play--80% of the M1's performance gains over the next best generic ARM chip are due to a handful of key innovations while the remaining 20% come from a long tail of incremental improvements. This would mean that generic manufacturers could copy those handful of key innovations (modulo IP rights, which might be another fatal flaw in my speculation) and recoup 80% of that performance gap in just a few years. If that assumption is bad and the performance is entirely an accumulation of lots of minor innovations over time, then I suspect it will take chip manufacturers quite a lot longer to catch up.

There's more to it than chip design though. There is the OS work to take advantage of the chip's new features. And designing new chip features to accommodate sw bottlenecks. Apple has very close collaboration between software and hardware engineering, while in the competitive space, those two things are not under the same roof. Makes it tougher.

True, and I don’t expect generic manufacturers to completely close the gap between them and Apple, but I do expect there will be enough to copy that we’ll still see a step change in PC performance.

Qualcomm has been unable to compete with Apple's ARM phone processors for years. It's gotten to the point that the only processor that comes close to a new Apple SOC on benchmarks is Apple's previous SOC. Why do you expect the PC market will be different than the phone market?

Because now Qualcomm just use (or modifying a bit, like cache capacity) Cortex-A CPU core design. Both Intel and AMD designs own core.

> it's not like other ARM chip manufacturers are going to have to go through all of the same R&D expense to figure out how to make chips that are comparably fast--they just need to copy what Apple did without infringing on any particular patents

Very handwaivy but the devil is in the details of that "all they have to do". Copying success is not that straight forward or else AMD would have had a response to Conroe in 2007, Intel for Zen in 2018, and Qualcomm for A-chips around 2014.

> When iPhone came out in 2007, Android was still an OS for BlackBerry like devices.

When iPhone came out in 2007, Android did not even exist. [1]

[1] https://en.wikipedia.org/wiki/Android_(operating_system)

How is that any true? From the article:

> An early prototype had a close resemblance to a BlackBerry phone, with no touchscreen and a physical QWERTY keyboard, but the arrival of 2007's Apple iPhone meant that Android "had to go back to the drawing board".

Your own source contradicts you.

I'm not sure how. iPhone was available in June 2007. Android went into public beta in November 2007, and reached 1.0 and was first released in a commercial device in September 2008. [1]

I guess Android, or what morphed into Android, was probably being developed, in a pre-beta state, when iPhone was announced in January 2007. Okay. It sounded to me like someone was saying that Android was released on Blackberry-like phones at the time iPhone came out. That was only thing i was objecting to. Sorry, a small point I almost didn't even bother to make, and probably not helpful to anybody, anyway.

[1] https://en.wikipedia.org/wiki/Android_version_history

Do they? Because i have yet to see any numbers that aren't clearly the result of a massively streamlined SoC. Yes the ARM chip is great, but the RAM is blazing, tiny, and more close to the ARM chip than anything out there for x86, right?

I'll be interested to see if Apple can keep this up with 128GiB of RAM and a beefy GPU.

So far to me M1 is great, ARM is great, but comparing a tiny amount of RAM soldered onto the CPU compared to flexible and huge RAM potential devices seems dishonest. Unless Apple manages to squeeze 128GiB of RAM on their chip, at least.

I want to see what my Blender workstation looks like on Apple's Architecture. Where i purpose built a huge CPU, GPU, and RAM for Blender rendering and large scenes. I suspect Apple will start seeing huge falloffs in the SoC design under larger workloads.

Apple is dominant in making phones though. And this Laptop is designed like a Phone. So yea, it's amazing for that.

Architectural analysis of the chip [0] suggests it is a massive improvement as a general purpose CPU, not just because of Apple's vertical integration and streamlining.

The magnitude of extra ROB entries, having an 8x wide instruction decoder, that massive L1I cache, and such fat cores with lots of ALU and other op units while simultaneously keeping over a 3GHz clock is a phenomenal architectural improvement overall.

Apple may not need to add more RAM locally. Other silicon designers like IBM with POWER10 are moving to serial memory with massive "L4" caches, seen previously with their Centaur memory controllers. I can see Apple using OMI or a similar serial memory bus to add extra memory, using 16/32/etc GB of PoP RAM for caching reads and writes to slower off-chip memory. I expect Intel and AMD will eventually move to a similar memory model, at least in server grade hardware.

[0] https://www.anandtech.com/show/16226/apple-silicon-m1-a14-de...

Yea and to be clear i'm not saying it's not awesome. I'm just saying that a lot of things - like real world benchmarks of programs/etc - massively benefit from the proximity and speed of the RAM. Proximity seems to be a very difficult beast to scale. Works great in small scale, horrible in large scale.

Sidenote, even if Apple rolled out 128GiB tomorrow i probably couldn't afford it. As the OP says:

> including automobile-priced Mac Pros

That's rough.

Hypothetically, with these new memory architectures, you could get the best of both worlds. Best case, you get a huge speed up if everything fits in the high-speed RAM, worst case, you fall back to the "old" RAM.

Side note, while Apple charges ridiculous prices for their upgrades, Mac Pros are actually some of the more ""reasonably priced"" computers comparatively. I've looked at HP and Lenovo's equiv. workstations, and you pretty much get screwed by everyone at that scale.

My desktop can take up to dual 128 GB sticks but one stick costs over $1,000, so I'm stuck with my 32 GB stick :(

... OTOH, it's pretty insane you can get 128GB of RAM for $1,000 nowadays. I remember when 8GB was a lot and cost several hundred. Time and miniaturization go on I guess...

> RAM soldered onto the CPU

It's just very close to the CPU, on the same package. It's not on the same die or anything. There are laptops that already have a shorter physical path to RAM than the M1 MacBooks, stacking the RAM right next to the CPU is common.

So it's just a matter of time before >16GB is available with an M1. These are just entry level machines in the first round.

> It's just very close to the CPU, on the same package. It's not on the same die or anything.

Yea, i was just simplifying. Though in SoC i think that's a perfectly fine simplification lol.

> So it's just a matter of time before >16GB is available with an M1. These are just entry level machines in the first round.

Think there will be > 32GiB?

Almost certainly, although maybe not immediately.

Let the Pro devices come out next year.

I know I'm probably older than most people here, but the notion that 16 GiB is "a tiny amount of RAM" is tremendously amusing to someone who paid $740 in 1986 (with a developer's discount) for 2 "MB" of RAM (we didn't have MiB back then)...

To be fair, we sort of still don't have MiB even now.

It depends on what you're comparing: A14 on phones have a Multi Core score which would be quite comparable to Snapdragon 888 (extrapolating a 20% boost over SD865) that was just announced all while A14 being clocked higher (3.0ghz 4 core vs 2.84 x 1, 3 x 2.4) comparing performance cores. Single thread OTOH Apple wins comfortably but I would suspect X1 to reduce the lead a bit. As for AI performance Apple claimed 11 TOPs, while Qualcomm claims 25 TOPs

A14 has two performance cores.

Reply on news.ycombinator.com