M4 Rumors (requests for).

Also most actually new microarchitectures are every 2 years or so in the industry, not every year like Apple was doing at the beginning. And, as everyone has already said, Apple had a lot low hanging fruit in those years. As did their fabrication partners too I might add. The progress Apple made in the 2010s was insane but that was as much a product of where they were in their CPU development cycle as who was there (not to take away anything from their team who did exceptional work). Expecting that level of progress to continue unabated forever was/is a little naive. I don’t care what level of geniuses they have or don’t have on the team.

I don't think Apple was ever doing a new u-arch every year, but they were iterating different parts of the architecture at different times. E.g. one year it could be some update to the cache hierarchy, the next year it could be a tweak to the execution units etc...
 
I’m sure you’re correct.

Slightly OT but I’ve been seeing rumours that Zen 5 has a huge increase in IPC. The estimates say between 20%-40%. Seems hard to believe.
Wow. Even if it’s on the lower that’ll be amazingly good. If true it’ll be really interesting to see how they managed it.

I don't think Apple was ever doing a new u-arch every year, but they were iterating different parts of the architecture at different times. E.g. one year it could be some update to the cache hierarchy, the next year it could be a tweak to the execution units etc...

My memory was the first few years it was but I might be mistaken. Regardless the performance gains almost every year were pretty exceptional. But as you said a 350 pt increase in a benchmark looks a lot more impressive percentage-wise at 1000 pts than 3000pts.
 
Last edited:
Wow. Even if it’s on the lower that’ll be amazingly good. If true it’ll be really interesting to see how they managed it.
It would be but as I said I”m pretty suspicious. I didn’t keep a very close eye on previous AMD generations, but I could swear that every generation, AMD fans were saying that “Zen 2/3/4 has massive improvements”. They definitely seem good, but I”m not sure the predictions were true. The other thing is the only benchmark I saw quoted is…drum roll…everyone’s favourite benchmark…Cinebench R23!
 
It would be but as I said I”m pretty suspicious. I didn’t keep a very close eye on previous AMD generations, but I could swear that every generation, AMD fans were saying that “Zen 2/3/4 has massive improvements”. They definitely seem good, but I”m not sure the predictions were true. The other thing is the only benchmark I saw quoted is…drum roll…everyone’s favourite benchmark…Cinebench R23!
I think Zen 3’s were actually as impressive as the rumors suggested. It was 19% - especially impressive as it was on the same node so they couldn’t use the new fabrication process to create advantages in their new design. But again Zen 2 was no where near Intel’s IPC. Zen 3 was about hitting parity. Zen 4 was another 13% IPC increase. So still nice. I don’t know what the rumors were though.
 
I think Zen 3’s were actually as impressive as the rumors suggested. It was 19% - especially impressive as it was on the same node so they couldn’t use the new fabrication process to create advantages in their new design. But again Zen 2 was no where near Intel’s IPC. Zen 3 was about hitting parity. Zen 4 was another 13% IPC increase. So still nice. I don’t know what the rumors were though.
Ahh fair play then. They must have been quite far behind then? Even now they aren’t as fast as Intel’s 13th gen or the M3.
 
Slightly OT but I’ve been seeing rumours that Zen 5 has a huge increase in IPC. The estimates say between 20%-40%. Seems hard to believe.

There is an alleged leak showing Zen5 part's single-core performance at 2800 points in CB23. If true, a very good result for AMD. Of course, CB23 is a one-sided benchmark that is very sensitive to L1 and SIMD bandwidth. It is not clear that performance wins in CB23 would translate well to other applications.
 
Ahh fair play then. They must have been quite far behind then? Even now they aren’t as fast as Intel’s 13th gen or the M3.
Very very far behind before Zen 3. I think in single core they aren’t as fast Intel but I believe that’s more Intel pushing their wattage up higher rather than IPC. Both AMD and Intel operate pretty far out of the “knee” but Intel is extreme even if they’ve dialed it back a bit in their latest generation.
 
It would be remiss of me not to point out of course that there is still room for curveballs in the M3 generation! We haven’t finished it yet. I mean probably not, but you never know …
 
Very very far behind before Zen 3. I think in single core they aren’t as fast Intel but I believe that’s more Intel pushing their wattage up higher rather than IPC. Both AMD and Intel operate pretty far out of the “knee” but Intel is extreme even if they’ve dialed it back a bit in their latest generation.
I have a 14700KF (8P+12E) to mess around with (morbid curiosity!).
The package uses >40W with one core loaded at peak boost (5.6GHz).
It hits the 250W boost power limit without every core loaded.
If I disable all 12 “””efficient””” cores, the remaining 8 P cores will use ~210W at ~4.8GHz.
It’s a hilarious thing.
 
I have a 14700KF (8P+12E) to mess around with (morbid curiosity!).
The package uses >40W with one core loaded at peak boost (5.6GHz).
It hits the 250W boost power limit without every core loaded.
If I disable all 12 “””efficient””” cores, the remaining 8 P cores will use ~210W at ~4.8GHz.
It’s a hilarious thing.
I feel like I should specify that when I said Intel dialed it back in the latest generation I was referring to their latest laptop chips … not the desktops! Yikes.
 
While improving IPC is tough, Apple may be able to instead (or additionally) squeeze out gains by improving the performance of its various coprocessors (neural engine, etc.).

And according to this Feb 15, 2024 article by Roman Loyola of MacWorld, China's Economic Daily* reports that will be a focus of the M4.

[*Amusingly owned by the Central Propaganda Dept. of the Chinese Communist Party, and for which Loyola gives the unfortunate abbreviation "ED".]



The China Daily article adds this...I don't know its significance. (Translation by Google Translate):

"In addition to increasing its investment in TSMC, Apple has also secured a large amount of advanced packaging production capacity from TSMC. According to the industry, Apple is still mainly placing orders with TSMC for 2.5D advanced packaging processes such as InFO and CoWoS. This year, there is an opportunity to push the demand for advanced packaging to the most expensive and difficult 3D architecture SoIC advanced packaging, which means that TSMC will simultaneously hold Apple's wafers. OEM for advanced processes and advanced packaging and other large orders."
 
Last edited:
While improving IPC is tough, Apple may be able to squeeze out additional gains in the performance of its various coprocessors (neural engine, etc.).

And according to this Feb 15, 2024 article by Roman Layola of MacWorld, China's Economic Daily* reports that will be a focus of the M4.

[*Amusingly owned by the Central Propaganda Dept. of the Chinese Communist Party, and given the unfortunate abbreviation "ED".]



The China Daily article adds this...I don't know its significance. (Translation by Google Translate):

"In addition to increasing its investment in TSMC, Apple has also secured a large amount of advanced packaging production capacity from TSMC. According to the industry, Apple is still mainly placing orders with TSMC for 2.5D advanced packaging processes such as InFO and CoWoS. This year, there is an opportunity to push the demand for advanced packaging to the most expensive and difficult 3D architecture SoIC advanced packaging, which means that TSMC will simultaneously hold Apple's wafers. OEM for advanced processes and advanced packaging and other large orders."

 
Is this about the packaging tech mentioned in the Economic Daily article? If so, why isn't TSMC named as a co-defendant? Harder to sue b/c it's based in Taiwan?
 
Odd … wouldn’t that be leveled at TSMC more than Apple? Then again I’m not entirely sure what is being patented … the article is awfully vague.

Is this about the packaging tech mentioned in the Economic Daily article? If so, why isn't TSMC named as a co-defendant? Harder to sue b/c it's based in Taiwan?

Seems I am not alone in my confusion! 🙃
 
Is this about the packaging tech mentioned in the Economic Daily article? If so, why isn't TSMC named as a co-defendant? Harder to sue b/c it's based in Taiwan?

Odd … wouldn’t that be leveled at TSMC more than Apple? Then again I’m not entirely sure what is being patented … the article is awfully vague.



Seems I am not alone in my confusion! 🙃

Rule number 1: I don’t talk about fight club.
 
With respect to the rumors that there will be a strong focus on the neural engine it’s interesting to me that the last time the neural engine got a bump in the A series of chips, it was marketed as no improvement on the M series. May already be considered good enough there and not pursued further for now. It’s only used for inference as I understand it not training so good enough may be good enough and thus leave a bigger transistor budget for order things
 
A few months ago there was a flurry of patents regarding neural engine. If I remember correctly it was something about having two hardware units, each specialized for different type of computation, and an engine that coordinates both. Also a bunch of patents for hardware based merge sorts.

And last week there were some patents about faster computation of special FP operations like exp and log, which could be used either in the GPU or in the NPU.
 
A few months ago there was a flurry of patents regarding neural engine. If I remember correctly it was something about having two hardware units, each specialized for different type of computation, and an engine that coordinates both. Also a bunch of patents for hardware based merge sorts.

And last week there were some patents about faster computation of special FP operations like exp and log, which could be used either in the GPU or in the NPU.
May I ask which site you use to search for patents? Do you just search for ‘Apple’ or are there specific engineers you search for?
 
Back
Top