CPU Wars 2022 - Apple Silicon vs. Intel/AMD

[maflynn]

Cinebench R23 is not a good cross-platform benchmark. It is better to use Geekbench 5.

Like it or not, Cinebench is a defacto standard for cross platform benchmarking. I like everyone else will keep using it a point of comparison. You can make a case that this app behaves like many other apps for the Mac so its presenting real world-ish results, since many other apps fall into the same category of not being fully optimized.

 

[pshufd]

Like it or not, Cinebench is a defacto standard for cross platform benchmarking. I like everyone else will keep using it a point of comparison. You can make a case that this app behaves like many other apps for the Mac so its presenting real world-ish results, since many other apps fall into the same category of not being fully optimized.

I like Geekbench 5 because their database contains a lot of old equipment and it's helpful so that people with really old stuff, say 2007 - 2014, to compare with Apple Silicon or Windows systems. Cinebench has the benefit of a sustained workload which can test some things in devices that are thermally constrained.

 

[maflynn]

Cinebench has the benefit of a sustained workload which can test some things in devices that are thermally constrained.

I like how it works, what it provides, and its something I rely on to compare my computers, so its something that is in my personal toolbox.

Using this to run in a looped manner as you mention is also a must have, as it tests a computer's ability to work under a sustained load, and provides a benchmark number to measure that performance against others.

 

[name99]

I picked up on the LLVM commit that lists FeatureHCX for the A16, so definitely looking forward to seeing new hypervisor/vm stuff in macOS at some point. I'm curious, though-- why do you think there'll be new page sizes and cache protocol?

GitHub - name99-org/AArch64-Explore

Contribute to name99-org/AArch64-Explore development by creating an account on GitHub.

github.com

The relevant volume is volume 2, starting at page 207 (look for "large page support").
This will give you the patents and how I interpret them and their relevance.

 

[name99]

I picked up on the LLVM commit that lists FeatureHCX for the A16, so definitely looking forward to seeing new hypervisor/vm stuff in macOS at some point. I'm curious, though-- why do you think there'll be new page sizes and cache protocol?

Going forward (eg the HCX stuff) it's unclear to me the extent to which Apple and ARM will diverge.

Someone who cares about security more than me can explain whether the new indirect page permissions stuff, https://community.arm.com/arm-commu...rs-blog/posts/arm-a-profile-architecture-2022
is in any way (meaningful or not) different from APRR/SPRR (discussed in volume 3, pg 40). That seems to be very much ARM building on Apple work.

OTOH we have Apple saying nothing and apparently doing nothing about SVE. I can't tell if this is just because SVE has been delayed to the serious new desktop chip (as opposed to the energy-improving while keeping the architecture the same A15/M2 and A16) or if Apple feels the basic concept of SVE is flawed (not just at the minor level of dumb choices like allowing 3*128b or 5*128b wide registers, but as a basic abstraction).
The alternative would be something like Macroscalar, which Apple worked on for 10 years or so, and which idea has been used or at least suggested in a few different places (Cray, a proposal for POWER, Mitch Alsup's My 66000). Mitch Alsup describes this succinctly as a design that vectorizes *loops* rather than *instructions*. I have some discussion of the issues in volume 4 at the very start (around pg 4 where I discuss the nature of Instruction Fetch) though mostly that, like all the volumes, is about what Apple ships in the M1, not what it might one day ship.

 

[senttoschool]

Like it or not, Cinebench is a defacto standard for cross platform benchmarking. I like everyone else will keep using it a point of comparison. You can make a case that this app behaves like many other apps for the Mac so its presenting real world-ish results, since many other apps fall into the same category of not being fully optimized.

I don't like it. It's been stated countless times that it's unoptimized and it's a poor indicator for CPU performance for both ARM and x86 CPUs. A good benchmark is consistent and can predict the performance for other applications. Cinebench is neither.

The reason why it's popular is because AMD pushed it hard in their marketing during Zen and Zen2 launches. Cinebench disproportionally favored Zen's architecture.

It's a niche software in a niche. It's a straight up terrible benchmark for everything except Cinemark 4D.

Please read what I wrote here on Reddit. Anandtech's Andrei F who now works at Nuvia (Qualcomm) agreed with me on every single point.

https://www.reddit.com/r/hardware/comments/pitid6

 

[altaic]

Like it or not, Cinebench is a defacto standard for cross platform benchmarking.

Please define “defacto” so we can be on the same page.

 

[altaic]

GitHub - name99-org/AArch64-Explore

Contribute to name99-org/AArch64-Explore development by creating an account on GitHub.

github.com

The relevant volume is volume 2, starting at page 207 (look for "large page support").
This will give you the patents and how I interpret them and their relevance.

Have you tracked Olivier Giroux at all? Was at Nvidia, now Apple. He’s got some interesting patents about concurrency and memory coherency.

I’d last saw your v0.7 doc, so I’m glad you kept going. What do you think about a wiki on GitHub? Maybe collaborate with Asahi?

 

[maflynn]

I don't like it

And I do

Please define “defacto”

Perhaps I misused the word, but my intention is commonly used by most if not all reviewers, regardless of platform.

 

[altaic]

Perhaps I misused the word, but my intention is commonly used by most if not all reviewers, regardless of platform.

The discussion is about comparing platforms/architectures. Common use of a biased metric is irrelevant. Reviewers’ use even more so.

 

[maflynn]

The discussion is about comparing platforms/architectures. Common use of a biased metric is irrelevant. Reviewers’ use even more so.

See here's where I disagree because I'm of the opinion that the cinebench functions like many existing apps for the Mac. Can it be optimized for the M1? Sure, but the same can be said for many other apps and that turns into a situation where it correctly provides metrics on how the Mac will largely perform in relation to other platforms.

To me, its like Apple stating their GPU is faster then the RTX 3090, when that's sort of misleading in real world usage. I think cinebench is realisticlly showing real world performance.

You can disagree with me, that's fine, this is my opinion, and its fine that you have your opinion, there's no reason to force my thoughts on you and vice versa. I respect what you had to say but I'll continue to use cinebench as a means of comparision.

Peace

 

[altaic]

See here's where I disagree because I'm of the opinion that the cinebench functions like many existing apps for the Mac. Can it be optimized for the M1? Sure, but the same can be said for many other apps and that turns into a situation where it correctly provides metrics on how the Mac will largely perform in relation to other platforms.

To me, its like Apple stating their GPU is faster then the RTX 3090, when that's sort of misleading in real world usage. I think cinebench is realisticlly showing real world performance.

You can disagree with me, that's fine, this is my opinion, and its fine that you have your opinion, there's no reason to force my thoughts on you and vice versa. I respect what you had to say but I'll continue to use cinebench as a means of comparision.

Peace

View attachment 2088564

I’m completely fine with you pointing out that that graph is basically marketing BS, but that doesn’t make cinebench a legit benchmark to compare architectures— regardless of clickbait. Two wrongs don’t make a right, right?

 

[leman]

And I do

You are perfectly within your rights to like or not like anything you want but your personal preference it doesn’t change the fact that it’s a bad benchmark. Already the fact that its results do not correlate with industry standard benchmarks should make it clear.

I think cinebench is realisticlly showing real world performance.

Except it doesn’t. Cinebench is supposed to estimate Cinema4Ds rendering performance yet you will most likely use an GPU accelerated renderer in practice. At most what CB23 scores will tell you is the expected performance for Blender CPU based renders, and even that is under a big question mark since blender continues being optimized.

So the reality is that you cherry pick one particular (albeit reasonably popular) software library where Apple happens to underperform and use that to make statements about its performance as a whole. How is that a constructive approach? There is a good reason why industry uses a wide range of workloads and tests to estimate performance. I mean, it’s akin to me taking R (a popular statistics suite) or compiler benchmarks - both software Apple Silicon happens to be very good at and claim that x86 is slow.

 

[MacCheetah3]

I can help… Being a strong advocate of any benchmark is silly.

There is a good reason why industry uses a wide range of workloads and tests to estimate performance.

And test suites such as PCMark attempt to more realistically mimic a user’s variety of tasks. But let's face it, even their test gauges a fraction of probable workloads.

The only way it really works is when you cater to a specific audience. Even then, for example, organizations such as GamersNexus appear to use per-game integrated benchmarks or replicated sample play rather than tools like Time Spy or Heaven/Valley/Superposition. Another example is how/why PugetBench is an automation add-on (e.g., plugin) for actual apps.

Basically, there’s no valid way to gauge the general performance of a system. Doing so is fine for entertainment, however, not reliable in practical comparison.

 

[name99]

Have you tracked Olivier Giroux at all? Was at Nvidia, now Apple. He’s got some interesting patents about concurrency and memory coherency.

I’d last saw your v0.7 doc, so I’m glad you kept going. What do you think about a wiki on GitHub? Maybe collaborate with Asahi?

I'm out of this game now; I have other projects I want to work on.
As I said, other people are welcome to pick up where I left off - at the end of Vol 3 is a long list of suggestions for where people can start.

What Asahi are doing is interesting, but Hector blocked me from following him on Twitter (the new generation are even less tolerant of any disagreement about *anything* than prior generations...) so, yeah, not much collaboration gonna happen there!

 

[playtech1]

I think when looking at Intel and AMD and comparing them with Apple you need to bear in mind two things:

- Apple sells complete computers, not just the chip, so comparing Apple Silicon to Intel/AMD is not an apples to apples(!) comparison

- High performance desktop computers are a niche market for PCs these days (how many desktop computers do you see in a store?), with even desk-based staff getting laptops. You can tell this from how Intel and AMD focus so much on gaming benchmarks - it's like there isn't much of an interest in the top model outside of gamers. The lack of desktop focus is even more true for Apple than other manufacturers, given that Macs don't end up on typical corporate desks.

All this means that desktop CPU comparisons are not really all that meaningful - Apple is playing a different game and Apple Silicon chips are attuned to that.

It's very notable that Intel and AMD are launching desktop first with mobile down the road, whereas Apple went with a mobile first approach because that's what they sell! If Intel and AMD were miles ahead then it might be a concern, but we are nowhere near that situation.

 

[cbum]

My take away of this and similar threads is that these benchmarks are mostly useful to compare your own old machine(s) to potential purchases you are considering. You can assess how the benchmark results on your older boxes relate to how the boxes perform in your favorite workflows, and get an impression what the benchmark results on your prospective purchase may mean.

 

[leman]

I think cinebench is realisticlly showing real world performance.

By the way, I was curious about this so I did a test. M1 Max gets around 12K points in CB23 multicore while 12900HK gets around 18K, so Intel is around 50% faster here. And yet in Blender CPU rendering benchmarks the 12900HK gets 234 points while M1 Max gets 219 points, so Intel is just 7% faster...

BTW, my M1 Max only uses around 25W of CPU power running the Blender CPU benchmark. Which again shows that this software fails to properly utilise the hardware.

 

[mi7chy]

Using the logic around here it's not optimized for Intel.

 

[leman]

Using the logic around here it's not optimized for Intel.

Using an Intel-designed library? Are you trying to say that Intel is utterly incompetent? 🤣

 

[Kazgarth]

I think when looking at Intel and AMD and comparing them with Apple you need to bear in mind two things:

- Apple sells complete computers, not just the chip, so comparing Apple Silicon to Intel/AMD is not an apples to apples(!) comparison

- High performance desktop computers are a niche market for PCs these days (how many desktop computers do you see in a store?), with even desk-based staff getting laptops. You can tell this from how Intel and AMD focus so much on gaming benchmarks - it's like there isn't much of an interest in the top model outside of gamers. The lack of desktop focus is even more true for Apple than other manufacturers, given that Macs don't end up on typical corporate desks.

All this means that desktop CPU comparisons are not really all that meaningful - Apple is playing a different game and Apple Silicon chips are attuned to that.

It's very notable that Intel and AMD are launching desktop first with mobile down the road, whereas Apple went with a mobile first approach because that's what they sell! If Intel and AMD were miles ahead then it might be a concern, but we are nowhere near that situation.

Intel changed it's strategy to mobile first. All the upcoming advancement in architecture are launching first for laptops.
Intel 4 (Meteor Lake) cores will launch in Q1/23 for mobile computers, end of 23 for Desktop PCs. Same goes for Intel 3, 20A & 18A.

 

[exoticSpice]

Intel changed it's strategy to mobile first. All the upcoming advancement in architecture are launching first for laptops.
Intel 4 (Meteor Lake) cores will launch in Q1/23 for mobile computers, end of 23 for Desktop PCs. Same goes for Intel 3, 20A & 18A.

MTL is not coming Q1/23 more like Q3/23. Yes Intel is moving to a mobile first because that's where the money is.

 

[Kazgarth]

MTL is not coming Q1/23 more like Q3/23. Yes Intel is moving to a mobile first because that's where the money is.

I've read about the delay rumors, still not confirmed as far as I know.
But even if they shift it to Q3, it only means the desktop version will be delayed further.

 

[ian87w]

I have more faith on AMD than waiting for Intel to actually ship their stuff in sufficient manner without breaking the bank. Raw performance is no longer crucial imo for the general consumer computing. Performance efficiency is the king as it allows more compact devices, fanless devices, lighter devices, etc. Apple is king in performance per watt right now it's not even a joke. The others cannot even give full performance without being plugged in and having noisy fan while the M1 is doing stuff without fans.

That's not saying Intel and AMD couldn't catch up, but so far AMD is the one delivering the goods, while Intel is still pretending that their 11th gen are premium products. My Pavilion Aero 13 is an amazing light laptop, and it has AMD in it, not Intel.

Meanwhile, thanks to the focus on mobile, Apple already has a head start with ML, neural engine, and accelerators within their SoC design. These are the new frontiers that will allow futuristic uses of our computers. raw CPU performance.

The only thing is, will Apple fumbles around as competitors catching up? That's the thing the Apple sometimes does , stumbling on its own due to form over function, or some focus on virtue signalling. Is Apple can maintain their momentum, it would be tough for Intel to catch up unless they have a breakthrough in architectural change.

And then there's the software side.

 

[exoticSpice]

The only thing is, will Apple fumbles around as competitors catching up? That's the thing the Apple sometimes does , stumbling on its own due to form over function, or some focus on virtue signalling. Is Apple can maintain their momentum, it would be tough for Intel to catch up unless they have a breakthrough in architectural change.

Yeah I believe Apple has a clear focus which Macs to thin and which Macs to make properly cooled.

MacBook Air -> Make it as thin as possible and fully silent

Macbook Pro 114"/16" -> Dual Fans

iMac 24" -> Thin and slim

Mac Studio/Pro -> cooling is important and thermals are key


During the 2016-2019 Mac era Apple put thiness a top priority even on Pro models.