M3 Chip Generation - Discussion Megathread

[APCX]

If I may posit some ideas on Apple Silicon as a whole.....

I believe that the M1, M2 and M3 are all, essentially, the same base design. Sure the core designs have evolved a little, and the GPU has been optimised, and most recently we have added new features. But the core, 'CORE' design concept has vastly remained the same.

- Arm v8.x instructions for the CPU cores.
- The raster engine in the GPU hasn't changed much.

I believe between M1 and M2 they 'fixed' some GPU scaling issues to do with cache (correct me if I'm wrong).

Other than that, what we're seeing here is M1 with two generations of clock speed boosts thanks to process improvements, and more cores, also thanks to process improvements and die space.

I know I'm vastly oversimplifying, but we haven't seen MAJOR changes.

So - my point - I believe that, like Intel, they stick with the same foundational design concepts for a few 'generations' before making a large jump to the next thing. Assuming in our case it will be Arm v9, and new GPU cores.

Maybe M4 will be that new generation. A larger leap in IPC, and jump in single-core performance. That said - no other CPU manufacturer is getting huge IPC improvements these days, and for Intel even shrinking the process will just result in faster clocks and more cores more than likely.

It appears that Apple does need to improve raster performance on the GPU cores for them to be able to compete with the higher end parts of nVidia and AMD. While Apple has never really tried to compete at the top with GPU, there's no reason why they shouldn't want to, given they already do with CPU.

I suppose it depends what constitutes a major change. I guess there are some principles that will guide their decisions going forward which don’t change. Having said that, according to what I’ve read, both the cpu cores and gpu are new.

One of the heads of gpus at Apple writes:
"it has been a long & satisfying journey to build this brand new GPU. I remember as if it were yesterday the first meeting we had to answer the question: "What GPU should we build next?". We had so big ambitions: scary and unbelievable in retrospect. But the end result is a scary fast GPU with never before seen architecture, HW ray tracing & mesh shading."

Gokhan Avkarogullari (@gavkar@mastodon.gamedev.place)

Attached: 1 image it has been a long & satisfying journey to build this brand new GPU. I remember as if it were yesterday the first meeting we had to answer the question: "What GPU should we build next?". We had so big ambitions: scary and unbelievable in retrospect. But the end result is a...

mastodon.gamedev.place

Also both the P cores and E cores are new according to Apple. Some confirmation of this from here:

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

 

[APCX]

It’s a completely new GPU core design. Not an iteration.

Gokhan Avkarogullari (@gavkar) on X

It's been a long & satisfying journey to build this brand new scary fast GPU: never before seen architecture, HW ray tracing & mesh shading. Congratulations to all teams at Apple for making this great GPU, great M3 SoCs and great MacBook Pro products that Apple introduced today.

x.com

Snap!

 

[jeanlain]

It's certainly a great GPU, but benchmarks have shown that core-for-core, the A17 Pro GPU (on which the M3 is based) doesn't perform better than the A16 GPU for rasterization.
Il will probably prove faster in tasks using mesh shaders though.

EDIT: first GFXBench results confirm it. The gains are marginal compared to M2. https://gfxbench.com/device.jsp?ben...=GPU&hwname=Apple M3&did=118295482&D=Apple M3

For compute, it's the same story. Geekbench reports scores close to 30000, while the M2 scores at about 28000.
We're talking about <10% gains.

M2 brought much greater improvements in these aspects.

 

[scottrichardson]

It's certainly a great GPU, but benchmarks have shown that core-for-core, the A17 Pro GPU (on which the M3 is based) doesn't perform better than the A16 GPU for rasterization.
Il will probably prove faster in tasks using mesh shaders though.

Yes I think this is my point - it's not a huge leap in raster performance and that's sadly what's most important for those of us doing raster graphics work and gaming.

 

[scottrichardson]

I suppose it depends what constitutes a major change. I guess there are some principles that will guide their decisions going forward which don’t change. Having said that, according to what I’ve read, both the cpu cores and gpu are new.

One of the heads of gpus at Apple writes:
"it has been a long & satisfying journey to build this brand new GPU. I remember as if it were yesterday the first meeting we had to answer the question: "What GPU should we build next?". We had so big ambitions: scary and unbelievable in retrospect. But the end result is a scary fast GPU with never before seen architecture, HW ray tracing & mesh shading."

Gokhan Avkarogullari (@gavkar@mastodon.gamedev.place)

Attached: 1 image it has been a long & satisfying journey to build this brand new GPU. I remember as if it were yesterday the first meeting we had to answer the question: "What GPU should we build next?". We had so big ambitions: scary and unbelievable in retrospect. But the end result is a...

mastodon.gamedev.place

Also both the P cores and E cores are new according to Apple. Some confirmation of this from here:

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

I stand corrected, thank you.

 

[APCX]

It's certainly a great GPU, but benchmarks have shown that core-for-core, the A17 Pro GPU (on which the M3 is based) doesn't perform better than the A16 GPU for rasterization.
Il will probably prove faster in tasks using mesh shaders though.

I’d be interested in any benchmarks you have for this please.

 

[APCX]

It's certainly a great GPU, but benchmarks have shown that core-for-core, the A17 Pro GPU (on which the M3 is based) doesn't perform better than the A16 GPU for rasterization.
Il will probably prove faster in tasks using mesh shaders though.

EDIT: first GFXBench results confirm it. The gains are marginal compared to M2. https://gfxbench.com/device.jsp?benchmark=gfx50&os=OS X&api=metal&cpu-arch=ARM&hwtype=GPU&hwname=Apple M3&did=118295482&D=Apple M3

For compute, it's the same story. Geekbench reports scores close to 30000, while the M2 scores at about 28000.
We're talking about <10% gains.

M2 brought much greater improvements in these aspects.

Hmmm gfxbench hasn’t been updated for a year and opencl is abandoned by Apple. Not saying you’re wrong, but I don’t think these scores prove it

 

[jeanlain]

Hmmm gfxbench hasn’t been updated for a year and opencl is abandoned by Apple. Not saying you’re wrong, but I don’t think these scores prove it

GFXBench certainly doesn't use advanced rasterization techniques that could make the M3 shine.
As for compute, I don't expect the difference between openCL and Metal to be larger for the M3. We'll see it very soon.

 

[diamond.g]

Hmmm gfxbench hasn’t been updated for a year and opencl is abandoned by Apple. Not saying you’re wrong, but I don’t think these scores prove it

Maybe GravityMark with a +2 LOD bias would be a good alternative?

 

[wonderings]

Any idea on how a base M3 stacks up against an M1 Max? Looking at a new iMac to compliment my M1 Max MacBook Pro and wondering how it stacks up

 

[mbouds]

Hi guys, I am a little bit confused as to whether M3 is built on N3E or not. MacRumors buyers guide states its built on N3E and so does wikipedia (the most accurate of resources lol). I find it hard to believe they only used N3B for A17 Pro and have already moved away from it, especially when the M chips are usually built of the year old mobile chips.

Can anyone give me a definitive answer either way? It won't make a difference to my purchase decision... I already bought an M3 Max 16 core

 

[Confused-User]

I believe that the M1, M2 and M3 are all, essentially, the same base design. Sure the core designs have evolved a little, and the GPU has been optimised, and most recently we have added new features. But the core, 'CORE' design concept has vastly remained the same.
[...]
Other than that, what we're seeing here is M1 with two generations of clock speed boosts thanks to process improvements, and more cores, also thanks to process improvements and die space.

I know I'm vastly oversimplifying, but we haven't seen MAJOR changes.

This is really not right.

People keep talking about "just" "some clock speed boosts" but they don't appreciate that there is no "just" about it. It's a MAJOR (to use your term and uppercaseness) effort to make the chip scale up that much in frequency. Being on a new process node makes that possible, but you also need to redo the logic.

 

[ChrisA]

I have doubts the 35 TOPS is 16bit.


You are realy talking about the company with a $26 billion R&D budget?

Yes, All those people need to be kept busy. You can't give them "peaky" workloads where they work 60+ hours a week for months, and then you send them all home for months. Using a staggered product release schedule keeps them all buzzy with a more or less uniform workload.

I have the same doubts, My guess is that 35 TOPS is int-8. That is pretty much what other compting TPUs are getting, 35 TOPS for int-8

 

[Homy]

M3 Max doesn't beat M2 Ultra in multi-core.

Highest single-core so far

 

[JordanNZ]

M3 Max doesn't beat M2 Ultra in multi-core.

View attachment 2305854 View attachment 2305855

Highest single-core so far

View attachment 2305856

Let’s see how it’s looking after a couple of days for background tasks to finish. The fact it’s that close to the studio is insane.

 

[APCX]

Let’s see how it’s looking after a couple of days for background tasks to finish. The fact it’s that close to the studio is insane.

Indeed. Impressive gains.

 

[bcortens]

Its a fair comparison as both chips have the same objective to compute and one did it in less time.

To quote GB:
"...n Geekbench 6, one workload is used and all the cores work together on that one shared objective. It is still true that the more cores you have, the quicker it will complete. However, there is now interaction between the cores."

GB6 is not a good test for embarrassingly parallel workloads because, as they acknowledge in its design, it is trying to simulate workloads that have threads with dependencies (IIRC but I may be wrong). As such it isn’t a good test of theoretical maximum multi core performance on independent workloads, GB5 is better for that.

I think GB7 should have both a GB5 like independent test AND the GB6 style test.

 

[sunny5]

Apple M3 chips appears to be a hybrid of the A17 Pro, A16 Bionic designs

The new M3 chips are here and there’s a lot to digest. As Apple doesn’t make a point of revealing the design origins of its chips, it can be an interesting exercise to put the pieces of the puzzle together.

www.notebookcheck.net

Rumors suggest M3 might be based on A16, not A17 Pro according to leaked Geekbench 6 scores because of how much M3 improved.

 

[scottrichardson]

3,151 single core GP6 benchmark score for M3 max is pretty awesome.

 

[JordanNZ]

Apple M3 chips appears to be a hybrid of the A17 Pro, A16 Bionic designs

The new M3 chips are here and there’s a lot to digest. As Apple doesn’t make a point of revealing the design origins of its chips, it can be an interesting exercise to put the pieces of the puzzle together.

www.notebookcheck.net

Rumors suggest M3 might be based on A16, not A17 Pro according to leaked Geekbench 6 scores because of how much M3 improved.

That’s an article with a lot of words, and very little substance. They’re basing the entire thing off of the NPU.

 

[sv8]

A18 could be on N3P. I'd even say it's likely. The A16 was on N4P, with TSMC giving the same exact volume production estimate ("2H 2022" for N4P and now "2H 2024" for N3P).

That doesn't entirely quash your suggestion, but it does remove an element.

I think a lot of the N3B hand-wringing (I'm not saying you're doing that, or really any of the regulars in this thread) is of the jumping-to-conclusions variety. N3E exists because of FinFlex transistor technology. It's a far more flexible node than N3B, by design. I think it's unlikely any M3 UltraFusion (which, remember, is not standing still -- TSMC has made investments and advances in "fusion" technology as well) silicon would be on N3E, mostly because there is no real advantage in that for Apple. The flexibility of N3E doesn't matter if you already have what you need in N3B.

The importance of N3E and N3P (nobody knows what N3S will be, or even if it will happen) is they are the last FinFET nodes. They are the height of that transistor technology. They will be around for a very long time. N2 and beyond, with "Nanosheet" (gate all around) transistors is next. Apple will stay on the cutting edge, so I'll eat my hat if the M4 presentation at WWDC in 2025 doesn't feature the word "Nanosheet" prominently, but an A18 on N3P could be a long-term stalwart in the iPhone and iPad lineups.

I see. There is a possibility that mass production of N3P will be completed in time for A18 in 2024.

N3E mass production is targeted for H2 2023, but actually it will be in 2024, so I thought N3P mass production was targeted for H2 2024, so it would actually be in 2025.

I would also like to hear an answer from Confused-user as to whether he considers N3E to be a node only for iPhones and iPads, but I'm glad to hear an interesting answer from tenthousandthings.

thank you!

 

[sunny5]

That’s an article with a lot of words, and very little substance. They’re basing the entire thing off of the NPU.

But CPU seems fishy for sure as M3 is two generation ahead compared to M2 and yet, the performance gain is little.

Beside, the single core performance might be the result of high clock speed. M3 has 4.05Ghz while M2 has 3.48Ghz which is 16.38% more and Geekbench 6 results also showing almost 16% difference in terms of single core performance.

 

[bcortens]

But CPU seems fishy for sure as M3 is two generation ahead compared to M2 and yet, the performance gain is little.

View attachment 2305895
Beside, the single core performance might be the result of high clock speed. M3 has 4.05Ghz while M2 has 3.48Ghz which is 16.38% more and Geekbench 6 results also showing almost 16% difference in terms of single core performance.

Ummm you are aware you’re comparing the base m2 pro vs regular m3 right?

Also, almost 16%? The picture you include clearly shows, on the picture, that the improvement is 16.7%. So, more than 16%…

 

[APCX]

Beside, the single core performance might be the result of high clock speed.

Unlike Intel/AMD/Qualcomm who achieve their performance through wishing wells and magical pixies.

 

[scottrichardson]

Comparing M3 Max to M2 Max on Geekbench 6 OpenCL test:

My M2 Max 38 Core = 87,992
Best M3 Max Geekbench 6 Result = 93,627

6.4% faster?