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

[Basic75]

The complexity of multicore machines is that it can be tough to divide up certain tasks. Some things depend on the result of another operation, as a result, you may not see a speedup from having multiple cores/cpus. There are other issues as well.

Overall, my point would be that neither is acting or simulating the other.

When I used the word "simulate" I simply meant that 1 big processor core can do the job of 2 half as fast ones by switching between two threads at a sufficiently rapid pace. On the other hand the 2 half as fast ones will simply be half as fast as the big processor core when only a single thread is available to run (because 1 of the 2 will be idle.)

And yes, it's the complexity, and sometimes impossibility, of writing such multi-threaded code that leads to the second situation, fewer threads than cores, being quite common.

 

[bobcomer]

Intel's approach since Alder Lake makes a lot of sense to me, even in configurations with 2 or 4 performance cores and 8 area-efficient cores:

The performance cores can take care of the not very multi-threaded applications, and when you have many threads, from one or more applications, then all cores can contribute.

It makes a lot of sense to me too.

That's one of the things that excited me about AS to begin with.

 

[falainber]

No, we are increasing the number of cores because we hit a limit with single-core performance. While we still get small improvements in single core, it also costs increasingly more power. Multiple cores can side-step this problem.

That's all correct. More transistors mean more cores. Comparing Apple design to other designs we are dealing with marginal differences though. So, we might be dealing with deliberate trade-off decisions. As I understand, Apple performance core on N3 is larger than Intel performance core on Intel 7. That's a lot more transistors used for marginal improvement in single core performance. Given the prevalence of multithreaded tasks in performance critical domain, using these transistors for extra cores might be the right decision. Apple priorities might be slightly different because they use the same core for phones and "workstations". For them, the trade-off may be influenced more by the phone priorities (for example, javascript in the browsers is still not very parallelized). For AMD and Intel typical loads and priorities are different which might explain why " Intel and AMD don't make chips like the M2".

 

[leman]

It makes a lot of sense to me too.

That's one of the things that excited me about AS to begin with.

AS doesn't use this approach though. Apple's E-cores are primarily there to handle background or auxiliary tasks while using very little power. You can use them to contribute to demanding tasks to a certain degree, but that's not their speciality. They are energy-efficiency cores, not throughput cores like Intel's.

Besides, Apple doesn't need this kind of design (at least not yet), because Apple's P-cores already use less power than Intel's E-cores while providing almost as much performance as Intel's P-cores.

 

[leman]

As I understand, Apple performance core on N3 is larger than Intel performance core on Intel 7.

No, Apple's performance core (sans cache) on A17 Pro is 2.2mm2 (A14/A15 was around 2.6mm2). Intel's P-core is 5.37mm2. Intel's E-core is 1.7mm2. Basically, Apple's P-core is only 0.5mm2 larger than Intel's E-core. Intel's P-core is 2.5x larger than Apple's P-core. Apple's E-core is super tiny at 0.6mm2.

Also, we have die shots of Intel's upcoming Meteor Lake (new Intel 4 process). The core looks very similar to Golden Cove but is only 25% smaller (5.3mm2). So even with Intel's new process their cores are twice as large than Apple's cores.

And to be complete, AMD cores (Zen4) are very similar in size to Apple's 5nm ones. Compact version of Zen4 (Zen4c) is about 40% smaller.

Sources:

https://twitter.com/i/web/status/1486850331355095040

Meteor Lake Die Shot and Architecture Analysis – Why Is Intel 4 Only A 40% Area Reduction Versus Intel 7?

Die shot analysis shows 40% area reduction on identical structures from Intel 7 to Intel 4. Analysis on Redwood Cove and Crestmont architecture, Foveros Omni, Intel's use of TSMC N3 in the GPU, and SOC/IO chiplets

www.semianalysis.com

(analysis taken from Twitter based on TechInsights die shots, I don't have the original link unfortunately)

 

[R2DHue]

View attachment 2292187
Just a fun screen grab for those harping on intel and amd. While M series are great it's not like the other guys are sitting in the back seat crying. It comes down to use case. Apple made hardware that can do things now that hardware in the past couldn't but you have to be honest with yourself if you think other systems can't do it.

Apple customers and Apple Inc have Gerard Williams to thank for the fact that Apple’s multi-year lead in Silicon has evaporated.

And probably why the A17 (and likely the M3) still use the Avalanche/Blizzard CPU cores which debuted in the A15 in 2021.

In fairness, IMHO, Apple should have gone back to their “X” naming scheme and named the iPhone 15 Pro and Pro Max’s SoC the A16X.
˙

 

[APCX]

Apple customers and Apple Inc have Gerard Williams to thank for the fact that Apple’s multi-year lead in Silicon has evaporated.
˙

It seems to have condensed into a significant lead if the leaks about the S24 are true.

 

[R2DHue]

No, Apple's performance core (sans cache) on A17 Pro is 2.2mm2 (A14/A15 was around 2.6mm2). Intel's P-core is 5.37mm2. Intel's E-core is 1.7mm2. Basically, Apple's P-core is only 0.5mm2 larger than Intel's E-core. Intel's P-core is 2.5x larger than Apple's P-core. Apple's E-core is super tiny at 0.6mm2.

Also, we have die shots of Intel's upcoming Meteor Lake (new Intel 4 process). The core looks very similar to Golden Cove but is only 25% smaller (5.3mm2). So even with Intel's new process their cores are twice as large than Apple's cores.

And to be complete, AMD cores (Zen4) are very similar in size to Apple's 5nm ones. Compact version of Zen4 (Zen4c) is about 40% smaller.

Sources:

https://twitter.com/i/web/status/1486850331355095040

Meteor Lake Die Shot and Architecture Analysis – Why Is Intel 4 Only A 40% Area Reduction Versus Intel 7?

Die shot analysis shows 40% area reduction on identical structures from Intel 7 to Intel 4. Analysis on Redwood Cove and Crestmont architecture, Foveros Omni, Intel's use of TSMC N3 in the GPU, and SOC/IO chiplets

www.semianalysis.com

(analysis taken from Twitter based on TechInsights die shots, I don't have the original link unfortunately)

Can anyone with the proper expertise explain why:

1. The iPhone 15 Pro outperforms — to a meaningful degree — the iPhone 15 Pro Max in all benchmarks except Metal?

2. Why most ARM SoC makers use L3 Cache but Apple never does? (SRAM too expensive?)
˙

 

[APCX]

OT. I am convinced there is significant bot activity in this forum.

 

[R2DHue]

It seems to have condensed into a significant lead if the leaks about the S24 are true.

Shame… 🙁

It’s been reported that he took a bunch of key Apple Silicon executives and engineers — and by some accounts, proprietary Apple IP — with him to found Nuvia — which he sold to Qualcomm… 😠

 

[APCX]

Shame… 🙁

It’s been reported that he took a bunch of key Apple Silicon executives and engineers — and by some accounts, proprietary Apple IP — with him to found Nuvia — which he sold to Qualcomm… 😠

Right…and Qualcomm seems to have done nothing with it. I’m not sure Gerard is as indispensable as people fear. To be clear, Im saying Apple has a significant lead.

 

[leman]

Apple customers and Apple Inc have Gerard Williams to thank for the fact that Apple’s multi-year lead in Silicon has evaporated.

It hasn't evaporated in the slightest. True, ARM is catching up — early results suggest that Cortex X4 has comparable IPC to A14/A15. However, X4 is a considerably wider architecture, so it's a bit strange that these results are not higher?

And probably why the A17 (and likely the M3) still use the Avalanche/Blizzard CPU cores which debuted in the A15 in 2021.

No I doesn't? Where did you hear that nonsense? The new cores are significantly different.

1. The iPhone 15 Pro outperforms — to a meaningful degree — the iPhone 15 Pro Max in all benchmarks except Metal?

This statement is false, so there is nothing to explain.

2. Why most ARM SoC makers use L3 Cache but Apple never does? (SRAM too expensive?)

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.

To put this in perspective: 17 Pro has 16MB shared L2 cache between the two CPU cores, as well as 24MB of SoC cache. Those cache sizes are comparable to desktop CPUs that have many more cores.

 

[R2DHue]

Right…and Qualcomm seems to have done nothing with it. I’m not sure Gerard is as indispensable as people fear. To be clear, Im saying Apple has a significant lead.

Doesn’t seem like Qualcomm’s done nothing with it.

Qualcomm had been using stock ARM core designs until now (I believe).

Gerard Williams’ Nuvia instead took Apple’s approach to designing its own proprietary ISA-compatible architecture (Nuvia calling theirs Oryon — which probably just uses a lot of proprietary Apple IP, IMO) and now Qualcomm is using it.

(Wonder where Nuvia got all the IP to design Oryon in such record time… 🤔)

ARM sued, because it gave a license to both Nuvia AND Qualcomm and now Qualcomm wants the pay ARM for only one license.

Apple sued for IP theft but recently either settled or dropped the suit (maybe to fend off antitrust regulators, but Apple was in the right, IMHO).

Qualcomm didn’t buy Gerard Williams’ Nuvia for $1.8 Billion nothin’ I’m sure!
˙

 

[APCX]

Doesn’t seem like Qualcomm’s done nothing with it.

Qualcomm had been using stock ARM core designs until now (I believe).

Gerard Williams’ Nuvia instead took Apple’s approach to designing its own proprietary ISA-compatible architecture (Nuvia calling theirs Oryon — which probably just uses a lot of proprietary Apple IP, IMO) and now Qualcomm is using it.

(Wonder where Nuvia got all the IP to design Oryon in such record time… 🤔)

ARM sued, because it gave a license to both Nuvia AND Qualcomm and now Qualcomm wants the pay ARM for only one license.

Apple sued for IP theft but recently either settled or dropped the suit (maybe to fend off antitrust regulators, but Apple was in the right, IMHO).

Qualcomm didn’t buy Gerard Williams’ Nuvia for $1.8 Billion nothin’ I’m sure!
˙

I’m not sure what I’m supposed to be seeing here, but to be honest with you, that looks like a whole lot of nothing. Quoting an article that is speculating about a processor to be released at the end of 2024, with no specific numbers in terms of performance, so many years after Qualcomm purchased Nuvia, amounts to very little in my eyes.

 

[R2DHue]

It hasn't evaporated in the slightest. True, ARM is catching up…

I rest my case.
˙

No I doesn't? Where did you hear that nonsense? The new cores are significantly different.

Not according to the microscopy.

They‘re significantly different? Really? What are they called now? How is the architecture “significantly different” from Avalanche/Blizzard?

Why an average of only a 10% CPU performance increase over the A16 Bionic then? —With 6 cores instead of 5?
˙

This statement is false, so there is nothing to explain.

Prove it.

Let’s see the benchmarks.
˙

 

[R2DHue]

I’m not sure what I’m supposed to be seeing here, but to be honest with you, that looks like a whole lot of nothing. Quoting an article that is speculating about a processor to be released at the end of 2024, with no specific numbers in terms of performance, so many years after Qualcomm purchased Nuvia, amounts to very little in my eyes.

“so many years”?

Qualcomm’s acquisition of Nuvia was in March, 2021.

(🖐️ One…, two… 🤔)

I’m assuming changing the core architecture of a VLSI SOC, fabbing it and deploying it in a finished product doesn’t happen overnight.
˙

 

[APCX]

I rest my case.
˙


Not according to the microscopy.

They‘re significantly different? Really? What are they called now? How is the architecture “significantly different” from Avalanche/Blizzard?

Why an average of only a 10% CPU performance increase over the A16 Bionic then? —With 6 cores instead of 5?
˙


Prove it.

Let’s see the benchmarks.
˙

Just a friendly reminder that it’s sometimes better to read the threads before wading in.

Having said that, here’s my 15Pro Max benchmark

 

[APCX]

“so many years”?

Qualcomm’s acquisition of Nuvia was in March, 2021.

And they said they’d have a product by the end of 2022. one…two…overdue….for an M1 competitor.

(🖐️ One…, two… 🤔)

I’m assuming changing the core architecture of a VLSI SOC, fabbing it and deploying it in a finished product doesn’t happen overnight.
˙

Probably shouldn’t promise it then.

 

[leman]

Not according to the microscopy.

They‘re significantly different? Really? What are they called now? How is the architecture “significantly different” from Avalanche/Blizzard?

Microscopy? What are you talking about?

CPU researchers run microbenchmarks on A17 Pro and demonstrated that is has a wider architecture (more integer units, more branches, more instructions decoded) as well as larger out of order execution windows.

Here is some info from one of the best known experts on Apple CPU and GPU microarchitectures:

Dougall (@dougall@mastodon.social)

Quick first impressions of the new P-core in the A17 Pro: * 9 wide decode/frontend (up from 8), for 9-per-cycle MOVs (register and immediate) and NOPs. * 8 integer units (up from 6), four of which can handle flag operations (up from 3). * 6-per-cycle ADR (up from 4). * Load/store/SIMD throughput...

mastodon.social

Why an average of only a 10% CPU performance increase over the A16 Bionic then? —With 6 cores instead of 5?

They both have 6 cores (2P and 4E). The performance increases depends on the benchmark. Some microbenchmarks show up to 20-30% improvements. Some show improvements consistent with increase in clock frequency.

I have a suspicion that in most of these tests A17 Pro runs lower frequency than its nominal peak (closer to 3.7Ghz than 3.8Ghz), so IPC improvements are likely underestimated.

Prove it.

Let’s see the benchmarks.
˙

I have nothing to prove. You are the one making exceptional claims. Demonstrate your evidence that Pro is faster than Max and I will happily refute it.

 

[R2DHue]

Just a friendly reminder that it’s sometimes better to read the threads before wading in.

Having said that, here’s my 15Pro Max benchmark View attachment 2292757

Just a friendly reminder.

A comparison requires at least two.
˙

 

[R2DHue]

Microscopy? What are you talking about?

CPU researchers run microbenchmarks on A17 Pro and demonstrated that is has a wider architecture (more integer units, more branches, more instructions decoded) as well as larger out of order execution windows.

Here is some info from one of the best known experts on Apple CPU and GPU microarchitectures:

Dougall (@dougall@mastodon.social)

Quick first impressions of the new P-core in the A17 Pro: * 9 wide decode/frontend (up from 8), for 9-per-cycle MOVs (register and immediate) and NOPs. * 8 integer units (up from 6), four of which can handle flag operations (up from 3). * 6-per-cycle ADR (up from 4). * Load/store/SIMD throughput...

mastodon.social

They both have 6 cores (2P and 4E). The performance increases depends on the benchmark. Some microbenchmarks show up to 20-30% improvements. Some show improvements consistent with increase in clock frequency.

I have a suspicion that in most of these tests A17 Pro runs lower frequency than its nominal peak (closer to 3.7Ghz than 3.8Ghz), so IPC improvements are likely underestimated.




I have nothing to prove. You are the one making exceptional claims. Demonstrate your evidence that Pro is faster than Max and I will happily refute it.

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.

(Though maybe I should’ve expected it. People with knowledge and expertise can often get self righteous and condescending to those asking for knowledge.)

I think it sucks what Gerard Williams did and I’m rooting for Apple.

I thought we were on the same team.

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

See:
˙

 

[APCX]

Just a friendly reminder.

A comparison requires at least two.
˙

You asked for benchmarks, I provided one. Posting the aggregate scores at the early stages of the release doesn’t show much.

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?

As said previously, it might be time to go back and read some of the threads on here.

 

[APCX]

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.

(Though maybe I should’ve expected it. People with knowledge and expertise can often get self righteous and condescending to those asking for knowledge.)

I think it sucks what Gerard Williams did and I’m rooting for Apple.

I thought we were on the same team.

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

See:
˙

No one is getting snotty. You aren’t paying attention to the data. You don’t have to give a damn what I say. I’m no one on here. Dooming about Gerard when he has so far produced the square root of zero is unnecessary.

 

[R2DHue]

You asked for benchmarks, I provided one. Posting the aggregate scores at the early stages of the release doesn’t show much.

Exactly…
˙

 

[APCX]

Exactly…
˙

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.