Discussion Apple Silicon SoC thread

[Eug]

M1
5 nm
Unified memory architecture - LP-DDR4
16 billion transistors

8-core CPU

4 high-performance cores
192 KB instruction cache
128 KB data cache
Shared 12 MB L2 cache

4 high-efficiency cores
128 KB instruction cache
64 KB data cache
Shared 4 MB L2 cache
(Apple claims the 4 high-effiency cores alone perform like a dual-core Intel MacBook Air)

8-core iGPU (but there is a 7-core variant, likely with one inactive core)
128 execution units
Up to 24576 concurrent threads
2.6 Teraflops
82 Gigatexels/s
41 gigapixels/s

16-core neural engine
Secure Enclave
USB 4

Products:
$999 ($899 edu) 13" MacBook Air (fanless) - 18 hour video playback battery life
$699 Mac mini (with fan)
$1299 ($1199 edu) 13" MacBook Pro (with fan) - 20 hour video playback battery life

Memory options 8 GB and 16 GB. No 32 GB option (unless you go Intel).

It should be noted that the M1 chip in these three Macs is the same (aside from GPU core number). Basically, Apple is taking the same approach which these chips as they do the iPhones and iPads. Just one SKU (excluding the X variants), which is the same across all iDevices (aside from maybe slight clock speed differences occasionally).

EDIT:



M1 Pro 8-core CPU (6+2), 14-core GPU
M1 Pro 10-core CPU (8+2), 14-core GPU
M1 Pro 10-core CPU (8+2), 16-core GPU
M1 Max 10-core CPU (8+2), 24-core GPU
M1 Max 10-core CPU (8+2), 32-core GPU

M1 Pro and M1 Max discussion here:

Page 78 - Discussion - Apple Silicon SoC thread

Page 78 - Seeking answers? Join the AnandTech community: where nearly half-a-million members share solutions and discuss the latest tech.

forums.anandtech.com

M1 Ultra discussion here:

Page 109 - Discussion - Apple Silicon SoC thread

Page 109 - Seeking answers? Join the AnandTech community: where nearly half-a-million members share solutions and discuss the latest tech.

forums.anandtech.com

M2 discussion here:

Page 127 - Discussion - Apple Silicon SoC thread

Page 127 - Seeking answers? Join the AnandTech community: where nearly half-a-million members share solutions and discuss the latest tech.

forums.anandtech.com

Second Generation 5 nm
Unified memory architecture - LPDDR5, up to 24 GB and 100 GB/s
20 billion transistors

8-core CPU

4 high-performance cores
192 KB instruction cache
128 KB data cache
Shared 16 MB L2 cache

4 high-efficiency cores
128 KB instruction cache
64 KB data cache
Shared 4 MB L2 cache

10-core iGPU (but there is an 8-core variant)
3.6 Teraflops

16-core neural engine
Secure Enclave
USB 4

Hardware acceleration for 8K h.264, h.264, ProRes

M3 Family discussion here:

Page 215 - Discussion - Apple Silicon SoC thread

Page 215 - Seeking answers? Join the AnandTech community: where nearly half-a-million members share solutions and discuss the latest tech.

forums.anandtech.com

M4 Family discussion here:

Page 263 - Discussion - Apple Silicon SoC thread

Page 263 - Seeking answers? Join the AnandTech community: where nearly half-a-million members share solutions and discuss the latest tech.

forums.anandtech.com

 

[Eug]

The product also is drastically worse in some aspects like not supporting boot camp and since that was an official thing, it seemed to matter to some users. the people that actual look at benchmarks and understand the m1 implications can go either way: choose it because of drastic advantage or move away because of the drastic disadvantage.



Maybe but let's be honest. Final cut pro is their main piece to show how cool the m1 is. We have no iead how much dedicated hardware it uses and that will always be faster than using a CPU for the same task. All it shows is that having a fixed set of hardware and custom made software leads to superior results. which is not surprising really.

Boot Camp hasn’t been much of a draw for years now. Almost nobody cares these days. Much more important would be VM support so I’m surprised you didn’t even mention that.

BTW, it’s not just Final Cut. Video editing with LumaFusion has been silky smooth on iPad Pros for quite some time. I was doing stuff on my 2017 A10X model that made 2018 MacBook Pros struggle. Apple has been optimizing its software on Intel 1.5 decades but they could only do so much if the hardware support isn’t there. I suspect Apple felt hamstrung by Intel’s and AMD’s conflicting priorities and decided just to fix it themselves.

The iPad Pro A series chip proof of concept was very successful and I’m glad they brought that to the Mac, finally.

 

[mikegg]

So how is gaming performance on these? A CPU is no good without a good GPU...

Vastly superior to AMD's best iGPU and Intel's Tigerlake.

Final Cut Pro on M1 continues to impress.
M1 MBP with 8 GB RAM vs 2019 12-core MP with 192 GB RAM and upgraded GPU.

View attachment 34195

View attachment 34196

Plus, the Mac Pro's timeline playback was all stuttery at that clip I linked, but the M1 played it perfectly.

But @DrMrLordX said the M1 must be slower than the 4900HS because it loses in Cinebench, which is just a niche product that Ryzen owners use to stretch their e-pe....

People don't use Cinebench on their laptop. They do video editing, web browsing, opening apps, transferring files, doing designs, etc. All of these things, the M1 wrecks Renoir in performance.

The M1 chip is the fastest laptop CPU you can buy, period. It's also the fastest laptop chip for the vast majority of the time, period.

 

[mikegg]

The conversation was about whether $1000 is a great value for what you get from an MBA. If you don't want to participate in that discussion, and instead want to pigeon-hole people into camps, put words in their mouths, and then attack them for being in that camp and taking stances they're not taking, and be combative and oppositional -- whatever, enjoy your adolescence. If you want to discuss why you think $1000 is a great value for an MBA in the context of the current laptop market, then please redirect your response in that direction.

MBA at $1000 is a great value.

Yes, you can get 16GB of RAM and 512GB SSD, and maybe a low-end discrete GPU on a $1000 Windows PC.

But with the MBA, you're getting:

• A higher resolution, higher quality screen
• Better touchpad (Mac touchpads are the best, period)
• Significantly faster CPU performance
• Competent GPU performance (assuming $1000 PC gets you entry-level discrete GPU)
• Significantly better battery life
• Significantly better portability
• Significantly better build quality (MBA is all all metal)
• Significantly cooler and quieter laptop
• Significantly better overall system responsiveness

Believe me, I've been trying to switch from Macbooks to Windows laptops because I absolutely despise the touchbar. I found that Macbooks were competitive with Windows laptops in terms of hardware and quality when they used Intel chips. Now Macbooks seem like a no brainer with Apple Silicon.

With Windows PCs, anytime you want a high quality, high brightness, high resolution screen that can match Macbook Retina screens, the price instantly increases to Macbook prices. This is why I haven't switched to Windows.

I've been trying to tell PC master race nerds for years that people who buy Apple products aren't idiots. Apple phones, tablets, and laptops hardware are genuinely worth the price.

The dumbest thing Apple has done to Macbooks is the touchbar, which increased the price and decreased user experience.

 

[teejee]

Also, Intel is heavily leveraging it's AVX-512 instruction set with targeted instructions that accelerate AI, inferencing, machine learning etc and all the other stuff, so there are other ways of getting there without having to design and integrate custom hardware for these tasks.

.

AVX-512 has been out on the market for 5 years with very low adoption rate in SW compared to AVX2. So regardless of how good AVX-512 is, it doesn't give advantages to most users.

And rapidly decreasing market share among power users (to AMD) doesn't help for the future of AVX-512 either.

 

[jeanlain]

Final Cut Pro on M1 continues to impress.

I'm always cautious when benchmarks involve exporting to H.264 and H.625, as the results heavily depend on the dedicated encoding hardware, which the Mac Pro may not have (at least, it doesn't have quick sync). Sure, in the end the M1 can be faster than the Mac Pro, but this doesn't necessarily reflect the performance of the M1 CPU cores.

 

[mikegg]

I'm always cautious when benchmarks involve exporting to H.264 and H.625, as the results heavily depend on the dedicated encoding hardware, which the Mac Pro may not have (at least, it doesn't have quick sync). Sure, in the end the M1 can be faster than the Mac Pro, but this doesn't necessarily reflect the performance of the M1 CPU cores.

In the end, does it really matter? It looks like only ~20% of the M1 die is actually dedicated to the CPU. The rest are the GPU and accelerators.

If you want to compare CPU to CPU, you can though we already know what the CPU performance is like compared to Intel and AMD.

I think what's more interesting now is how the chip performs overall in actual common applications and how those applications take advantage of the neural engine and accelerators inside the M1.

 

[jeanlain]

That's disingenuous. The M1 is designed for low power usage and long battery life. The 5950X (with the 20.6W single core power usage) is a 105W TDP chip. Expecting its single core power usage to be lower than the M1 is absurd, they're designed for totally different things.

I'm not sure to follow. There may a be a tradeoff between pure performance and efficiency, and you would have a point of the M1 was less powerful than the Zen 3 core. But it's not.
And do Ryzen laptop CPUs use a different design compared to desktop? I'm not aware of that. They just use lower frequency, perhaps less cache, binning. But the core design isn't radically different, is it?
Zen 3 laptop CPUs should be 20-30% more efficient than their predecessors, just like their desktop brethren. This won't be enough to match the M1.

Also, the M1 core is many times more power efficient than intel's TGL, which is a laptop part (the A14 uses 5W vs 20W for TGL to reach similar SPEC scores.)

 

[Gideon]

Or M3...
Zen to Zen2 was ~2.5 years.
Zen2 to Zen3 was ~1.5 years

There MIGHT be a Zen3+ competing against the A15, but I'd expect for most of its lifetime, except perhaps for a few months, Zen4 will be competing against A16.

It is realy impressive how well Apple executes.

But just to note, Zen to Zen 2 was an outlier when AMD was essentially broke. They have commited to a 12-15 month cadence (15 months in practice) and they've kept it so far. Zen 4 will be out in Q1 - early Q2 2022, not late 2022.

 

[Carfax83]

A Zen 3 core at peak spends 19W-20W. A Firestorm core at peak spends 5W. Ignoring the I/O die, apple cores are significantly more efficient at peak performance.
By the time Zen 4 comes out, it will be competing against M2.

Yes, and a great deal of the M1's power efficiency comes from being on 5nm. When Zen 4 is on 5nm, AMD can get 30% less power for the same performance.

AVX-512 doesnt come close to dedicated accelerators for ML. Its like asking a CPU to run Graphics... why? Sure you can run it but GPUs do it better. You can run ML code on CPU but dedicated accelerators run it better. In the end of the day, no one cares which part it runs. If its faster, its faster.

My point is that CPUs will get faster and more efficient at these tasks as well. And the cost of R&D for developing and integrating custom hardware might eventually be seen as problematic as the average consumer just doesn't use those applications which leverages AI and machine learning on a regular basis.

Someone earlier brought up the analogy of PhysX. Many years ago, Ageia tried to sell PC gamers on a dedicated physics processing card in the P4 era. Back then, it wasn't seen as a strange idea because CPUs were so much weaker back then. A P4 could not run one instance of cloth physics simulation at an acceptable framerate for instance. Eventually Nvidia purchased Ageia and ported the API so that it could run on their GPUs so developers could implement mostly aesthetic effects like extra particles, cloth and smoke simulation. This typically called for getting a dedicated GPU PhysX card as attempting to run GPU PhysX on your main rendering card could tank performance to unacceptable levels.

Fast forward to the Core i7 era when CPUs got higher core counts, SMT, 256 bit SIMD and much higher IPC, suddenly CPUs were now powerful enough (in combination with physics software overhauls) to run all of the effects that hardware accelerated PhysX was capable of. In fact, cloth physics algorithms can run faster on modern CPU than they do on GPUs, or so I've read.

So now today, developers no longer care about implementing hardware accelerated physics in their games, and Nvidia has stopped trying to get consumers to buy dedicated GPU PhysX cards and PhysX runs primarily in software, which is actually much better for gameplay.

Renoir runs much higher than 15W. 15W is just the advertised TDP, which means little to nothing as to how much the chip actually spends. Unfortunately, renoir laptops are also quite rare and finding information on how much power they use is a nightmare, but I assure you when its running multicore benchmarks it doesnt run at 15W to achieve that performance.

The M1 sucks up power in multithreaded workloads as well, despite lots of technical advantages over Renoir as amrnuke alluded to.

 

[Gideon]

Some Adobe Creative Cloud benchmarks from Pudget:

Apple M1 MacBook vs PC Desktop Workstation for Adobe Creative Cloud

Apple has recently launched MacBook Air and Pro models using the new Apple M1 chip based on the Arm instruction set. While we do not usually examine performance for laptops, we wanted to see how these new chips compare to a desktop PC.

www.pugetsystems.com

 

[Carfax83]

I think the file size is a function of bitrate. Puget Systems did a series of tests linked below. The summary of export performance (speed) is also below.

Here's the thing. Hardware accelerated encoding is very nice for real time or performance sensitive encoding. If I were going to stream a game on YouTube, I would certainly choose NVENC over anything else. But for movies, I would choose a software solution because if you want maximum quality and flexibility, offline encoding is the best.

It's similar to why offline rendering is still usually done on CPUs, even to this day.

They also address the quality issue, stating that you couldn't really see a difference except in low bitrate modes, unless you zoom in. If we're talking say iPhone 4k video, you can forget it. The 150Mbps source videos they used as example is impossible to create from an iPhone Pro Max, the max they record at is around 63Mbps. These are professional quality videos to start with.

They said this in that article:

One thing to keep in mind is that our testing was done with "VBR, 1 pass" since hardware encoding currently doesn't support 2 pass encoding. If you were to use 2 pass encoding with software encoding, the quality difference would be a bit more pronounced (although it would also take significantly longer).

 

[jeanlain]

Some Adobe Creative Cloud benchmarks from Pudget:

...which are silly. They don't even mention Rosetta. Do they know that the M1 does not have an X86 CPU?

 

[nxre]

Yes, and a great deal of the M1's power efficiency comes from being on 5nm. When Zen 4 is on 5nm, AMD can get 30% less power for the same performance.

No. 7nm to 5nm only gives a 10% efficiency improvement for the same performance (Looking at the Kirin 9000 vs the Snapdragon 865+, we’re seeing a 10% reduction in power at relatively similar performance. Both chips use the same CPU IP, only differing in their process node and implementations - https://www.anandtech.com/show/16226/apple-silicon-m1-a14-deep-dive/3)

The M1 sucks up power in multithreaded workloads as well, despite lots of technical advantages over Renoir as amrnuke alluded to.

On cinebench, which seems to be the benchmark that most favours Zen 3, M1 uses 15W in MC and 3,8W in SC to achieve its peak results. Would be curious to see how much Renoir spends to achieve its MC results.

https://twitter.com/x/status/1328777333512278020

 

[nxre]

We have no good information on how Zen 3 cores scale down beyond the 5600X core which is still very much performance oriented rather than efficiency oriented, and it very well could be the case that AMD beats, is roughly equal to, or loses to Apple at that power threshold. We just don't know yet.

M1 achieves its peak SC performance at 5W. 5800U Cezanne, using Zen 3 cores, peak performance will come at 4,4Ghz, which looking at Zen 3 5600X power draw, seems to be around 10-11W. How will exactly AMD overcome this deficit, realistically? And, most importantly, why would they? Laptops are not close to being a big market to them.

 

[coercitiv]

Laptops are not close to being a big market to them.

Laptops are the second biggest market for them, after servers.

 

[Gideon]

Using Zen 3 cores, peak performance will come at 4,4Ghz which looking at Zen 3 5600X power draw, seems to be around 10-11W.

They will probably clock a bit higher (ST), looking a the diff between Renoir and Matisse desktop vs Vermeer. But you are right about power draw, they certainly won't make up the difference.

While I doubt AMD will reach anywhere near M1 efficiency, let's not forget that going to 7nm AMD doubled perf/watt (with Less than half coming from the process, according to them) And with Zen 3 they managed to improve power-efficiency by another 20% on the same node. Considering their trajectory and what patents reveal about Zen 4 I'd be surprised if Zen 4 doesn't offer at least 50% improvement in that category. Leaks saying Genoa is 96 Cores, seems to hint at it as well ( they can't really increase the TDP, as it would require water cooling).

Laptops are the second biggest market for them, after servers.

And both really need perf/watt.

 

[Eug]

I'm always cautious when benchmarks involve exporting to H.264 and H.625, as the results heavily depend on the dedicated encoding hardware, which the Mac Pro may not have (at least, it doesn't have quick sync). Sure, in the end the M1 can be faster than the Mac Pro, but this doesn't necessarily reflect the performance of the M1 CPU cores.

Of course it’s hardware accelerated. That was my whole point. Apple has built in excellent hardware acceleration, and leveraged it in such a way that it is a vastly superior user experience. Unfortunately, QuickSync just doesn’t cut it.

AMD based hardware support is better than QuickSync and Apple does support it on the Mac Pro (but not MacBook Pro), at least in some workflows.

 

[nxre]

Laptops are the second biggest market for them, after servers.

Really? I didn't know that, I had an idea intel had a monopoly on laptops, my bad.

 

[nxre]

While I doubt AMD will reach anywhere near M1 efficiency, let's not forget that going to 7nm AMD doubled perf/watt (with Less than half coming from the process, according to them) And with Zen 3 they managed to improve power-efficiency by another 20% on the same node. Considering their trajectory and what patents reveal about Zen 4 I'd be surprised if Zen 4 doesn't offer at least 50% improvement in that category. Leaks saying Genoa is 96 Cores, seems to hint at it as well ( they can't really increase the TDP, as it would require water cooling).

I wish we had good benchmarks on performance per watt, its getting extremely hard to compare when we don't even have matching data sets between chipsets and all we can do is guess some numbers.

 

[DrMrLordX]

Apple could probably made this slower in all aspects and on 7nm and sell it for $599 and it would be a much, much bigger win, financially and user-base wise. Les so tech-wise.

Apple's got a closed ecosystem user-wise. They can afford to tinker with profit/loss and maybe bump prices up a bit just to keep the ball rolling. Having strong technology locked up in their ecosystem just makes it more likely that they can drag people into their way of doing things. Once you go Mac, it's hard to get out for a variety of reasons. There really isn't any need for them to lower prices and use an older node, when they can just raise prices and use the bleeding edge.

But @DrMrLordX said the M1 must be slower than the 4900HS because it loses in Cinebench, which is just a niche product that Ryzen owners use to stretch their e-pe....

I see someone's still upset that he got caught mis-using hyperbole. Be more careful about what you say next time.

They do video editing, web browsing, opening apps, transferring files, doing designs, etc. All of these things, the M1 wrecks Renoir in performance.

Actually, I think if people did enough of those things all at once, that the 4900H (and maybe HS) would stand a chance of stacking up pretty well against the M1, albeit at much worse battery life/higher power draw.

The M1 chip is the fastest laptop CPU you can buy, period. It's also the fastest laptop chip for the vast majority of the time, period.

That's . . . contradictory. Is English your first language? Honest question.

 

[Eug]

Really? I didn't know that, I had an idea intel had a monopoly on laptops, my bad.

AMD CPUs are in a fair number of laptops, but IMO most of them are junk. They seem to lean towards the market with bad battery life and annoying international keyboards, at least around here. All the nicer WIndows laptops with good battery life seem to be Intel.

Maybe that will change with the new cores, but in the past AMD CPU in a laptop = red flag. I wonder how long it will take for AMD to shake that reputation.

 

[Eug]

Re: Pricing of the new M1 Macs:

While I don’t think Apple will hit the $599-$699 price point any time soon with Mac laptops, there is now a education specific SKU with decreased storage at 128 GB, available at $799. It’s not for the broad edu market, but is aimed at edu institutional purchases.

https://www.reddit.com/r/apple/comments/jxl3hs

It’s also available in 5-pack configurations for $779 each.

 

[mikegg]

The M1 chip is the fastest laptop CPU you can buy, period. It's also the fastest laptop chip for the vast majority of the time, period.

That's . . . contradictory. Is English your first language? Honest question.

How is it contradictory? The first sentence speaks to the CPU of the M1 only. The second sentence refers to the entire chip, aka SoC. Is English your first language? Honest question.

 

[shady28]

Here's the thing. Hardware accelerated encoding is very nice for real time or performance sensitive encoding. If I were going to stream a game on YouTube, I would certainly choose NVENC over anything else. But for movies, I would choose a software solution because if you want maximum quality and flexibility, offline encoding is the best.

It's similar to why offline rendering is still usually done on CPUs, even to this day.



They said this in that article:

Placebo effect. Below is the video review. What he said here was :

"...hardware encoding does result in a slight loss in video quality but in our testing it was fairly minimal it is only on surfaces with very fine detail like pavement where you might be able to see a noticeable difference in quality if you pause and zoom in the video"

There are some videos out there that show where hardware has better quality on some things.

Like I said before, if you're a pro use a server farm and do multi-pass encode. None of this is relevant to 99.9% of people.

 

[Eug]

Placebo effect. Below is the video review. What he said here was :

"...hardware encoding does result in a slight loss in video quality but in our testing it was fairly minimal it is only on surfaces with very fine detail like pavement where you might be able to see a noticeable difference in quality if you pause and zoom in the video"

There are some videos out there that show where hardware has better quality on some things.

Like I said before, if you're a pro use a server farm and do multi-pass encode. None of this is relevant to 99.9% of people.

Don’t forget, not all hardware encoders are created equally. For example, commercial encodes for stuff like UHD Blu-ray and video streaming are hardware based. Could you imagine Netflix trying to do software encoding?

I’m no expert but I do recall many years ago that Intel QuickSync hardware encodes were of worse quality encodes but perhaps that’s changed. And as you say, the quality difference was most obvious with low bitrate encodes.

BTW, speaking about bitrates. Netflix’s 4K can look excellent but needs about 15 Mbps. However, crazy as it may sound Apple sometimes has bitrates that nearly triple that on Apple TV+. I think they are often in the 20-25ish Mbps range IIRC, but can peak at well over 40 Mbps! Clearly Apple isn’t religious about saving space with their HEVC encodes.