Why Intel and AMD don't make chips like the M2 Max and M2 Ultra

[R2DHue]

Also there is a difference in the single core scores of 0.35% between the Pro and the Max and 0.48% multi core. Are you actually trying to make an argument that this isn’t just random variance?

From browser. geekbench. com:

“The data on this chart is gathered from user-submitted Geekbench 6 results from the Geekbench Browser.”

I’m assuming the data was “gathered” from at least more than one “user-submission.”

They don’t specify how many, but it probably gives a more realistic overall use case, depending on many factors, including the temperature of each submitter’s device, any energy compensations, any Machine Learning gathered wrt how each person uses their phone and apps, etc.

If the sample size is large enough, the numbers would be a more statistically significant overall performance representation for the product than one person putting up one set of scores on one phone.

Maybe submit yours and improve the curve…

 

[R2DHue]

I’m not obliged to provide you with multiple benchmarks. I have only one phone. Claiming a significant difference in phone performance after a couple of weeks results, with a difference of less than half a percent, is no way to make an argument.

So hostile…

You were never obliged to do anything. Sheesh!

(Who put the burr….?)

Sorry I asked.

(I won’t do it again.)

 

[APCX]

From browser. geekbench. com:

“The data on this chart is gathered from user-submitted Geekbench 6 results from the Geekbench Browser.”

I’m assuming the data was “gathered” from at least more than one “user-submission.”

They don’t specify how many, but it probably gives a more realistic overall use case, depending on many factors, including the temperature of each submitter’s device, any energy compensations, any Machine Learning gathered wrt how each person uses their phone and apps, etc.

If the sample size is large enough, the numbers would be a more statistically significant overall performance representation for the product than one person putting up one set of scores on one phone.

Maybe submit yours and improve the curve…

Mine was obviously submitted lol.

So to be clear, you think less than half a percent difference in aggregate is enough to justify your claim of a difference between the Pro and Pro Max?

 

[APCX]

So hostile…

You were never obliged to do anything. Sheesh!

(Who put the burr….?)

Sorry I asked.

(I won’t do it again.)

TIL it’s hostile to ask for some evidence for claims that go against the evidence. You asked for benchmarks, I provided one and you got snotty that I hadn’t provided more.

 

[R2DHue]

Mine was obviously submitted lol.

So to be clear, you think less than half a percent difference in aggregate is enough to justify your claim of a difference

”of a difference?”

Yes, that’s kind of the word’s definition.

I just found the scores surprising because I would not expect the 15 Pro Max to perform the same as the 15 Pro, I would expect it to perform appreciably better if there was any difference.

If scores were the same, I wouldn’t have even thought about it.

 

[leman]

I was asking in good faith for knowledge from people with expertise.

I didn’t expect people to get so snotty and to be drawn into a food fight.

I didn't have the impression that you were asking. You were making a bunch of unsubstantiated claims that are obviously wrong to anyone who has spend at least a little bit time looking into these matters. Like A17 having no new cores or the Pro being faster than the Max.

But, where “did I ever get the idea” that there were benchmarks showed the iPhone 15 Pro beating the iPhone 15 Pro Max?

Those results do not show any statistically meaningful difference. Averages/medians don't matter, you also need to look at the spread. If you grab a couple of dozens results from Pro and Max and run basic stats on that you'll see that they are indistinguishable.

 

[R2DHue]

TIL it’s hostile to ask for some evidence for claims that go against the evidence. You asked for benchmarks, I provided one and you got snotty that I hadn’t provided more.

I remember it differently.

We both wrote “friendly reminder.”

If neither is “snotty,” then truce.

 

[R2DHue]

I didn't have the impression that you were asking. You were making a bunch of unsubstantiated claims that are obviously wrong to anyone who has spend at least a little bit time looking into these matters. Like A17 having no new cores or the Pro being faster than the Max.

My fault. My fault.

I didn’t put the question marks in bold.

And I didn’t mean to make you SO offended‼️ 😱

 

[APCX]

I remember it differently.

We both wrote “friendly reminder.”

If neither is “snotty,” then truce.

I know mine was meant in good faith, I don’t know that about yours. It seemed retaliatory. We aren’t at war, we are discussing computers.

 

[Analog Kid]

In this case I am intentionally absolutist because (and I thought it was clear from the context) with "one task" I meant one thread. And in that case I hope that it's obvious that it doesn't matter how many smaller cores you have, a single larger one will be better because the small ones can't "gang up."

You introduce time slicing to simulate several cores on one. You can also computation slice to simulate one core on several. That's my point. To absolutely say you can't work the other way is wrong.

It's better to acknowledge that truth while still holding to the also correct broader view that for most general purpose tasks fewer, faster cores is better.

 

[R2DHue]

I know mine was meant in good faith, I don’t know that about yours. It seemed retaliatory. We aren’t at war, we are discussing computers.

And I know mine was meant in good faith, I don’t know that about yours.

(That tends to be the case for the person writing it.)

But now that you’ve told me, I’ll take your word for it — in good faith.

(Let’s not fight anymore… LOVE!)

😍

 

[APCX]

And I know mine was meant in good faith, I don’t know that about yours.

(That tends to be the case for the person writing it.)

But now that you’ve told me, I’ll take your word for it — in good faith.

(Let’s not fight anymore… LOVE!)

😍

Ok no problem.

 

[Basic75]

You can also computation slice to simulate one core on several.

Which processor and operating system makes multiple cores work on a single thread?

 

[Basic75]

L3 cache can either mean CPU-only cache (shared by multiple CPU cores) or it can mean the cache shared by various processors of a SoC (e.g. CPU, GPU, Ml accelerator etc.). Apple has very large shared L2 cache that incorporates the role of the traditional CPU L3 cache. Apple also has a large shared cache at an SoC level that is shared by all processors.

Quite right, there are three levels of cache between all CPU cores and RAM on all recent Apple Silicon chips, though it is fair to point out that Apple's caches do have some differences to what AMD and Intel are doing, and that these differences could be considered downsides.

For example on the M2 the 8MB LLC is smaller than the P-core cluster's 16MB L2. And on the M2 Pro and Max you have 32MB of L2, however that's split into 2 chunks of 16MB, one per 4-core P-core cluster, meaning that no single core can use more than 16MB of L2.

 

[leman]

Quite right, there are three levels of cache between all CPU cores and RAM on all recent Apple Silicon chips, though it is fair to point out that Apple's caches do have some differences to what AMD and Intel are doing, and that these differences could be considered downsides.

For example on the M2 the 8MB LLC is smaller than the P-core cluster's 16MB L2. And on the M2 Pro and Max you have 32MB of L2, however that's split into 2 chunks of 16MB, one per 4-core P-core cluster, meaning that no single core can use more than 16MB of L2.

The effective L2 size is smaller, since a single core cannot access all of it's clusters's cache (I don't remember how much, I think it can only see 3/4 of the cache or something like that)? But I would hardly consider that a downside, given the fact that these caches are still much larger than on any other CPU.

There are some real disadvantages to Apple's design, most notably higher latency compared to mainstream x86 implementations. Also, Apples L1 cache bandwidth is considerably lower. However, overall, these design choices work out well for them, and likely contribute to the lower power usage of their CPUs.

 

[quarkysg]

For example on the M2 the 8MB LLC is smaller than the P-core cluster's 16MB L2. And on the M2 Pro and Max you have 32MB of L2, however that's split into 2 chunks of 16MB, one per 4-core P-core cluster, meaning that no single core can use more than 16MB of L2.

Why not? My understanding is that the entire SLC is available to any cores.

 

[Basic75]

Why not? My understanding is that the entire SLC is available to any cores.

SLC yes, L2 no.

Edit: I called it LLC for Last Level Cache, you wrote SLC for System Level Cache. In this case it's the same thing.

 

[quarkysg]

SLC yes, L2 no.

Ah ... apologies. Got it.

 

[leman]

If the sample size is large enough, the numbers would be a more statistically significant overall performance representation for the product than one person putting up one set of scores on one phone.

You cannot talk about statistical significance without looking at the spread. For example, here are two numbers: 2883 and 2896, the difference is even larger than what GB6 averages show. And yet I generated both of these numbers from the same normal distribution with mean 2890 and standard deviation of 5.

In fact, I think it's very problematic that usual benchmarks report averaged values/point estimates instead of the actual result spread. It would be very easy to do these things right.

 

[Gudi]

You can not buy a (new) Mac with an Intel CPU or (any) PC with Apple Silicon. So where is the competition? What does it matter who has the better single core performance? If you can’t pull the CPU out of its socket and replace it with the other, all those bragging rights are moot.

 

[APCX]

You can not buy a (new) Mac with an Intel CPU or (any) PC with Apple Silicon. So where is the competition? What does it matter who has the better single core performance? If you can’t pull the CPU out of its socket and replace it with the other, all those bragging rights are moot.

What does that even mean? Are you saying performance doesn’t matter because Apple doesn’t use an Intel/Amd cpu?

I don’t know how to process that. You would really have hated the 90s when Sparc, Alpha, MIPS, PowerPC and X86 were around and creating great competition.

Is there no competition between Apple and Samsung? Can’t pull the cpu out of your phone, or generally out of your laptop. What a bizarre opinion.

Competition is between devices, for tasks you perform on those devices.

 

[MayaUser]

@leman so, until now, whats your opinion on the future M3 based on what we know about A17 pro
Can, based on the inside envelope, active pasive /cooling, to see M3 family different clock speeds especially on the P-cores for the M3 pro and M3 Max and Ultra devices with active cooling? Or just increase P-cores?
Thank you

 

[leman]

@leman so, until now, whats your opinion on the future M3 based on what we know about A17 pro
Can, based on the inside envelope, active pasive /cooling, to see M3 family different clock speeds especially on the P-cores for the M3 pro and M3 Max and Ultra devices with active cooling? Or just increase P-cores?
Thank you

I hope that we will see higher clocks on some of the actively cooled devices (like desktops, that have the thermal capacity). Will we see it? Who knows. For the MacBook Air I’d expect to see a clock frequency around 4Ghz. It will run a bit hotter than M2 on signs-core bursts but nothing the chassis can’t handle.

Btw, I am starting to think that we are underestimating the IPC improvements of the A17. The IPC calculations are based on the peak clock frequency, but in my test it is only maintained for few seconds. Five-ten seconds into the test the frequency stabilizes around 3.5-3.6Ghz. In contrast, A16 appears to maintain the peak clock for much longer. We don’t have the data to look at this in detail, but some napkin math from results when A17 is run with ice packs etc. make me think that the IPC increase might be close to 7%.

 

[deconstruct60]

Yes, the 68000 was a 16bit implementation (bus, alu) of a full 32bit architecture. However it only had 8 data registers (and 8 address registers.)

Sorry about that. Was thinking '32-bit registers' and wrote '32 registers' . 32 registers would be been a huge deal for 1980 era system that was affordable with the transistor budget restrictions those days.

In terms of register pressure it is further away from the PowerPC, , but still not eyeball deep in 'complicated' instructions. It isn't extreme RISC adherent but it isn't quite CICS either.

 

[Analog Kid]

Which processor and operating system makes multiple cores work on a single thread?

Automatic parallelization tool - Wikipedia

en.wikipedia.org

You said "can't simulate", so I don't see a reason to expand the scope of what should be a trivial discussion. It's unreasonable to limit the point to existing operating systems, compilers, etc. When we needed to make one processor look like many, we built the tools to do it. GPU programming and shaders are typically many cores working on a single workload in a way that, even if barred from async execution, would appear as a single host thread calling a peripheral.

But the details of how aren't really important. It is possible, it is often necessary, it is just not preferred.