Exponential gains of M1

[Fomalhaut]

Could we see an Apple Silicon GPU based on the M1 technology? Or that works different and we need x86?



How do they increase 20% yearly? I thought to increase speed you need to add more transistors, cores, or make them closer by nanometers? What I am asking is what do they do yearly that makes it 20% faster that they couldn't do last year?

We already have an Apple Silicon GPU on the M1, but I assume you meant a discrete GPU of some description. There is nothing that specifically ties dGPUs to x86 chips - ARM chips can (and do) support PCIe and would be able to access third-party GPUs given appropriate drivers.

Apple has various options for providing more powerful GPUs with the Mx chips:

1) Increase the number/capability of the GPU cores on the primary silicon die. This of course makes the die larger increasing the possibility of manufacturing failures, but the M1 is relatively small (c. 120mm^2) compared to larger CPUs and especially GPUs (a GTX 3080 has a die area of 628mm^2)

2) Have a separate GPU die on the same Mx package, connected via a proprietary interconnect. This potentially allows the GPU to be larger, and to have custom SoCs with the same CPU cores but differing levels of GPU capability, but still maintain the unified memory architecture and high performance.

3) Have a completely separate GPU package (similar to current MacBooks and iMacs) that is connected via either PCIe, or some new interconnect. The challenge here is how to make it work with unified memory. I think AMD has some technology that does this, allowing the GPU to directly access system memory, in addition to using its own VRAM.

As for CPU year-on-year performance increases, it depends on several things:
1) Run at higher clock frequency, which is possible with smaller transistors (less power required, so can increase the clock frequency to stay at the same TDP)
2) Better / faster cache and memory management
3) Architectural improvements the CPU execution mechanism (e.g. predictive branching) that increase the possible Instructions-per-clock.
4) Use of Application Specific blocks on the SoC - to speed up particular operations like file compression, encoding, lexical handling. This is a strong area for Apple.

I sure there are many other possibilities. It's a case of getting a few percent marginal gains for each change, that leads to a significant overall improvement.

 

[Fomalhaut]

I completely agree with you and is one of the main reasons why I decided to jump in Gen 1. I think two things are going to happen going forward. The spec bumps/updates will be every 18-24 months, and they will offer gains in both power and efficiency, but certainly nothing monumental (IMHO).

Point being, if you want a new MBA or MBP now, get it now, as it more than likely won't have been worth it to wait an additional 18-24 months for M2, which I think will only be incrementally better.

I agree that performance per core will only get incrementally better, and will probably slow down compared to the past few years as Apple runs out of "low hanging fruit". Apple may find itself in the same position as Intel circa 2015, when year-on-year improvements were single-digit percentages.

Their focus will probably be on improving the custom modules on their SoC to optimize more applications, and of course to offer more CPU and GPU cores.

I bought an M1 (Mini) to experiment with, but am waiting for the next MacBook Pro with an M1X (or whatever), which I hope will offer a big jump in multi-core and GPU performance. An 8+4 core CPU should be 75% faster than the M1 for multi-core workloads, and hopefully a 12-16 core GPU could match the current AMD Radeon Pro 5600 or even 5700 found in the MBP16 and iMac.

 

[s66]

Exponential gains are really, really hard to maintain by definition.
The computer industry is not capable of maintaining exponential growth for long enough to be important. Nor is any other industry.

But Apple has shown a really neat track-record of linear growth in performance in their A series.

The big jump from Intel to the M1 is also the goal of Apple. It's them showing it's important to embrace the change that does have a few drawbacks. Without that jump very few would be willing to accept the drawbacks. So Apple took their time till they had a big enough gap before they jumped the Intel ship.

The M1 is the entry level machine: there's no reason to not believe they don't have more CPU cores, more GPU cores, more ML cores, more RAM, more cache etc. all somewhere along the path between idea and product already. And as Apple's key staff knew about this for long enough: they've had time to get ahead already.

So I'd still expect the CPUs for the MP, iMac Pro etc. to be significant jumps in performance from whatever Intel has by then, or the Apple hating trolls will have a field day.
Apple seems to have plans according to Bloomberg to (eventually) make a Mseries SoC with 32 power cores (that's 8 times more than the M1!) Now adding more cores for most applications does not scale linearly, but still those now using a Xeon 28 core MP7,1: imagine an Apple Silicon with 32 CPU power cores ...
The same report also suggests Apple is working on 128 GPU cores ... that's 16 to over 18 times as much as the M1 has today (and with their tile approach Metal does scale pretty good over more cores). Given even that 7/8 core GPU in the M1 already beats gamer GPUs from a few years ago in performance, just imaging 18 times as many of them each more performant in itself...

Ref: https://www.bloomberg.com/news/arti...ac-chips-with-aim-to-outclass-highest-end-pcs

Add in the steady improvement they make on every component itself over the years, and the high-end might still flabbergast us all when it comes out in a product as those aren't expected in early 2021 yet.

From then on: I'd expect a linear growth. But that's more than fine and far better than what Intel gave us in the years past.
What Intel did / is doing was / is to boost the clock mostly: that gives a little more compute power at the cost of using a lot more electricity and thus generating a lot more heat - leading to thermal issues and/or throttling to avoid overheating.

The M series is also new: so there's more room to improve in the components inside it that it didn't inherit from the A series (which will be much more mature and harder to get significant improvements beyond the "chip making technology" itself.

Long term: they might run into trouble, nobody can predict that, after all when Apple moved from PowerPC to Intel CPU's: Intel looked good and promising - sadly they've dropped the ball it seems.

The real challenge and at the same time interesting is IMHO to see how they plan to get more I/O on the SOC, how to add (much) more RAM, how they'll make an expandable system, ... there's still plenty for Apple to reveal (or decide among options they still have in their product development pipelines). Obviously the M1 is far from what the MP8,1 will need - but OTOH the M1 is a promising start.

 

[MacBH928]

I'm not attempting to prove anyone wrong, simply laying out my opinion. I guess my point is, the MBA or MBP is my sweet spot computing device. I don't do any heavy video or photo editing, so an iMac and/or the 4 Thunderbolt variants of the 13" and 16" MBP's are overkill for me and extremely expensive to boot, so I'm narrowing my thoughts here specific to the M1 chip, not any of the larger more powerful chips that are coming next year.

That being the case, I don't think we see an M2 chip until mid 2022 at the earliest, and I think it will offer performance gains in the 15-25% range.

To wrap it all up, my point was for those in the market for a MBA or MBP now or soon, I think they would be doing themselves a disservice to wait an additional 18 months for an M2 that won't "change the game again" in my humble opinion.

I'm excited at the prospects of Apple Silicon's future and love the conversation!

I agree I don't think we will see any updates to the MBA-MBP(basic) with the current M1 until mid 2022. Its hard to believe a great leap will happen in less than 1.5 years, this was the great leap. I also don't think its worth it to release a spec. bump by Nov. next year then a big release by June 2022.

 

[1240766]

I will cross the bridge when it gets there...right now it is all speculation, I would rather enjoy the m1 now then think about the enjoyment of the m2 when it gets here....