Apple Confirms Their Future Desktops + Laptops Will Use In-House CPUs

[starshipeleven]

I personally experienced that situation, and I remember a friend of mine buying a dual G5 Macintosh just 8 months before the switch and after one years throws everything in the garbage because Adobe stopped the support on G4/5 series.

Yeah, but the application he installed was still working right?

Because a lot of people on Mac tend to have the same mindset of those that keep Win7 or WinXP and "disable updates", it's the people that think Mac have no malware and all that jazz.

All modern Apple devices are more or less designed to die in 3-4 years anyway (apart from overheating there are also many reasons, depends on the model, see videos by Louis Rossman), there is no need to change CPU arch to force people to upgrade.

[AmericanLocomotive]

and now add the cost of hiring hundreds of people to design your CPUs - I would say that it may be cheaper in the long run.

keep in mind that the decision to do this would have been 3-4 years ago before AMD brought competition back and made Intel actually do good things.

That's an important point. Decisions made today in the semiconductor industry won't manifest themselves for 4-5 years. An AMD executive back in the 90s was famous for saying that "Making semi-conductors is like playing Russian Roulette, except you don't find out for 4 years after you pulled the trigger if you blew your head off".

The decision to make these ARM chips would have been made years ago. Apple likely saw Intel's stagnation in regards to performance and efficiency, and AMD's Renoir was nowhere in sight, so the decision probably seemed like an easy one to make.

With the current competition in the x86 world, I'm not so sure Apple made the right choice, but at this point there is too much momentum for them to stop.

[Vistaus]

Bye bye Boot Camp.

You're right. Windows ARM doesn't exist, so Boot Camp will go away. Oh wait...

[Danielsan]

Yeah, but the application he installed was still working right?

Because a lot of people on Mac tend to have the same mindset of those that keep Win7 or WinXP and "disable updates", it's the people that think Mac have no malware and all that jazz.

All modern Apple devices are more or less designed to die in 3-4 years anyway (apart from overheating there are also many reasons, depends on the model, see videos by Louis Rossman), there is no need to change CPU arch to force people to upgrade.

Yes you could use the Mac but you were progressively losing interoperability with the newer versions of Apple ecosystem software. From a certain extent also Linux is horrible at retro-compatibility but distro are constantly updating everything hence we do not notice this and appimage is useful to keep older version of a certain application.

[hyperchaotic]

There is nothing "eco" about the Apple company or its products. Let us count the ways... planned product obsolescence, non replaceable batteries, poor repairability, lack of support for 3rd party or DIY repair, frivolous fancy packaging, economic support of oppressive communist regimes with abysmal human rights records (China), sweatshop slave labor conditions in their offshore factories. Did I miss any? I'd wager that Apple is responsible for an outsize portion of global e-waste and human suffering.

You're right about a lot of those things. However my 6 year old iPhone 5S runs the latest iOS and it got faster when it was installed as they had paid special attention to older devices, it's very difficult to find other phones with that kind of support lifetime. As for the other things, the problem here is that it's incredibly difficult getting electronics that isn't made in sweatshops (If you have found some ethically sourced devices for yourself that's incredibly rare). Apple is not alone in this, though they've made more improvements with suppliers like Foxconn (Apple don't do their own manufacturing) than most due to the exposure.

I guess what I'm trying to say is that they do have problems, but while we're staring at Apple hating on them, all other companies supplying competing products (from computers/phones to kitchen utensils) enjoy the benefits of having the same and worse business practices while everybody's attention remains solely on Apple. The truth is that the western world thrives on the cheap manufacturing overseas for mostly everything.

[starshipeleven]

appimage is useful to keep older version of a certain application.

*On Debian/Ubuntu

[oiaohm]

So what? For a desktop, or even a server under most circumstances, I wouldn't really care that much that the rpi4 has a tiny size. When a single SSD uses more power than the rpi4 CPU, the power savings are going to be limited by Amdahl's law, essentially. I do care about single threaded performance though on the desktop. Similarly for server latency (single threaded performance matters hugely as well).

Yes you are right that server chips latency is important. But not all servers have need for high IPC single threaded.

https://www.anandtech.com/show/15578/cloud-clash-amazon-graviton2-arm-against-intel-and-amd/5

Compared to a mobile Cortex-A76 such as in the Kirin 990 (which is the best A76 implementation out there), the resulting IPC is 32% better for the Graviton2 in SPECint2006, and 10% better for SPECfp2006.

This here shows you in single threads A76 the prior generation IPC. Graviton2 IPC is fairly much the same in single threaded as a Zen2 core from AMD. Remember Haswell intel that you are saying has 3x faster IPC than A72 is 40% slower in IPC than a Zen2 core. So by A76 arm cores have fairly much caught up and passed to Intel Haswell.

A78 has a 20% uplift and the X1 has a 30% uplift roughly. Single threaded of a X1 as long as ARM is telling the truth will roughly be inline with current Zen2 and the A78 single threaded will roughly be inline with Zen+ AMD for single threaded.

X1 is barely released, and there are no performance numbers other than manufacturer claims. BTW, when a manufacturer says "up to x percent", it's only indicative of the upper range of performance gains (you're generally guaranteed to never exceed them ...)

Except this is arm we are talking about the X1 is most likely not barely released its most likely one of the server core design that will be branded X1. Like how the Kirin990 was in fact released before the Cortex-A76 was announced. Arm if you spend enough time going over the slides comparing the released performance data to existing third party made arm chips you can normally find where the reference design is already in production before the reference design is even announced by 6 to 12 months under a different name.

So Arm is not like Intel and AMD in this regard the benchmarks ARM use are third party independents on existing chip in production the press release for the new design just does not tell you who that is or what the chip design the person was testing thought they had. Yes funny right ARM new design is not new.

Arm does basically like double blind trials on their designs. So the person benchmarking does not know its the new ARM reference design because it was released under a different chip name. Heck this has some advantages for arm they can release like 4 different designs from 4 different makers have independent parties review them all and take the winner. Arm development model is way different to Intel or AMD.

Arm model sees the silicon design really produced put in hands of third party reviewers before that design comes a Arm reference model that is why Arm slides are way more trustworthy on numbers than Intel/AMD as they are not some in house made numbers they are showing.

As for 384 threads - desktop workloads simply don't use them.

This depends on the desktop. Content creation desktop workloads like Blender the more cores the better. Blender will happy using 384 threads.


Arm chip to chip uplifts are quite massive.

A72-A73 +30% IPC performance.
A73-A75 +30% IPC performance
A75-A76 +30% IPC performance.
A76-A78 projected +20% performance
A76-X1 projected is +30% performance.

Remember all those prior ones turned out to be real and exact on the money Arm is not like Intel/AMD, Arm case what is on the slides turns out to be real over and over again. . Good IPC uplift with intel or amd is 15 percent between generations. Generations after A78 and X1 we can expect Arm and x86 to be close on IPC.


Please note that only 2.1x performance increase in IPC between A72 vs A76 not the 3x different you were seeing. Raspberry Pi 4 has another problem its been built for cost it has less than 1/3 the memory bandwidth of a Haskell in the worst case for the Haskell. A single DDR4 dimm give the Haskell chip 14933 MB/s of memory bandwidth to play with. Raspberry Pi4 at best only has 4500MB/s at theoretical best. A72 is not known for good memory bandwidth the server versions of Arm A72 perform a lot better than 4 core A72 on the Raspberry PI 4 due to a decent memory controller. So over 1x performance loss in your compare between raspberry PI4 and Haskell x86 chip is in fact lack of memory bandwidth that is partly caused by the way the A72 fills it caches as well resulting in stalls if it has lack of memory bandwidth. A72 design really does not like being staved with lack of memory bandwidth and the RPI4 does a really good job of staving it but again the RPI4 is built for cost not performance.

A72-A78 is 2.52x IPC uplift. and this gets you to Zen+ level IPC.
A72-X1 is 2.72 IPC uplift and that gets you to Zen 2 level IPC.

Reality A76 is roughly in IPC first generation Zen that is well head of Haskell as long as it not memory staved.

vladpetric really putting a x86 Haskell. up against a RPI 4 if Haskell could not win there was something wrong. You really missed how badly the lack of bandwidth memory tied RPI4 performance behind back.

RPI4 soc chip is designed for cost and power effectiveness yes less memory chip supported equals less idle power so more power effective in lots of cases. Server A72 chips are designed for performance so you see full blown memory controller in the class Haskell or better x86 chips have. Arm chips with full blow memory controllers perform way different to their embedded forms even that they are meant to be the same reference design. Its a surprise to lot of people that embedded chips are losing 1x or more performance due to horrible memory setups this explains why a lot of embedded chips remain on A72 and have not moved to A76 crippled memory bandwidth the A76 is not going to perform well might as well use cheaper silicon fabrication.

Arm server chip in a desktop workstation in lot of ways could be quite decent yes we do see workstations made by different arm server vendors. These workstations are most likely going to be way cheaper than the apple workstations with more cores so more suitable for content create than the apple ones will be as long as you can get the software.

[Michael_S]

All modern Apple devices are more or less designed to die in 3-4 years anyway (apart from overheating there are also many reasons, depends on the model, see videos by Louis Rossman), there is no need to change CPU arch to force people to upgrade.

I don't own any Apple products, but to be fair to Apple they have the least evil planned obsolescence life cycle for smartphones. All the Android vendors, and Microsoft when they were in the smartphone market, were terrible about supporting devices past two years. This is one area where Apple does right by consumers.

I'm hoping the PinePhone or something like it will be an adequate smart phone for real use (navigation, voice calls, texts, occasional mobile web browsing) soon. If not, when my current Android phone turns into a steaming pile of garbage two years after I got it - or sooner, which has been the trend - I might just buy an iPhone.

[starshipeleven]

I don't own any Apple products, but to be fair to Apple they have the least evil planned obsolescence life cycle for smartphones. All the Android vendors, and Microsoft when they were in the smartphone market, were terrible about supporting devices past two years. This is one area where Apple does right by consumers.

Software support is quite frankly irrelevant for mobile. Most users don't give a flying fuck about software support, as long as the device and the apps still work, all is fine.
Normal 100-200 euro Android phones will easily last around 4+ years before the battery is dead and the device becomes permanently tethered to a USB power source, and that's how much time most people will still use the same phone.

If not, when my current Android phone turns into a steaming pile of garbage two years after I got it - or sooner, which has been the trend

Did you buy a Samsung? They need to be fully factory reset every now and then. Apart from that I don't understand how it can become a "pile of garbage"

I might just buy an iPhone.

For the price of a single iPhone you can buy between 3 and 4 normal Android phones, and if you are choosing decent ones like for example the Nokia ones that are supported by Google directly as they are Android One devices so you get full 3 years of security patches you can get a lot more mileage for your buck.

[vladpetric]

vladpetric really putting a x86 Haskell. up against a RPI 4 if Haskell could not win there was something wrong. You really missed how badly the lack of bandwidth memory tied RPI4 performance behind back.

You're making a lot of points here, just addressing this one right now: it is pretty rare to saturate the memory bus, even with an RPi4. The reason is that the cache subsystem tends to work really well ... Let's consider a "bad" workload with L1 + L2 cache hit rates of ~95% (which is pretty low FWIW). That means that 1 in 20 loads goes to memory. Further assuming that 33% of all instructions are loads, then 1 in 60 instructions goes to memory. There is a significant cycle penalty with that miss, but 1 in 60 instructions missing won't saturate the memory bus, even on the RPi4.

Now even if you have a workload that misses to memory more than that (say, a database or the kernel), it will still be really hard to saturate the bus from the processor. Why? Because the pipeline, even with dynamic scheduling, will end up stalling a lot. A stalling pipeline will not generate requests to the memory bus at all ...

So why the bandwidth to begin with? On the RPI 4 - DMA primarily. Granted, you can totally have workloads that need a lot of DMAing, but those weren't the kind of workloads I was considering (I'm thinking more of CPU intensive stuff, e.g., SPEC CPU 2006/2017, where IO is trivial).