M3 Chip Generation - Discussion Megathread

[name99]

I'd be surprised if Nuvia wasn't using off the shelf Arm core interconnect all along. It would easier to sell the company to a standard Arm interconnect implementor and it would lower the start up costs on their first die they had to get out the door ( if didn't get sold). Either way it doesn't make much sense to go very far off the 'reservation' with a smaller team just getting started on a new design and market target optimizations. Chunks of the die subcomponents they'd be trying to just buy and use.

Possibly so but a constant theme in Apple's patents, from day one, is evolving the interconnect in multiple ways.
Some are attaching QoS tags to EVERY transaction, to ensure that QoS is always maintained, some is power tweaks (like a few side bits that say which bytes of a wide bus are valid, along with twiddles to only ship bytes that matter – stuff like this is done on a few interconnects but Apple has extended it so that the common case of transporting zeros in various places/ways is much cheaper), some is control, testing, and HW debugging tweaks, etc.
You can't just add this stuff if the FULL set of functionality you want requires, say, 151 wide bits and ARM only gives you 140bit wide connection.

All this stuff is valuable, and I can't believe the Nuvia folks would want to give it up unless they really have no choice. (Of course maybe there WAS no choice, and it's scheduled for the next design?)

 

[name99]

You misread. It's not two clusters. It's two *cores* that can go to 4.3GHz. And it's not designed with two special cores, they'll just pick the two most efficient cores. (This is what Intel and AMD do.)

There's more weirdness here.
The current usual situation is that all cores in a cluster run at the same speed (or, some could be powered off, or, in principle, some could run at half or third speed). But you need everyone at the same speed so that you can communicate with the L2 core without losing a few cycles crossing a timing domain.

Now in principle you could put each core on a separate timing domain, also separate from the L2, and just accept that the cycles lost communicating with L2 don't matter if you are OoO enough. Possibly they have done that. Its an obvious idea, you just need to run the numbers and see if it makes sense, and the fact that Apple don't do it (even for just the E-cores, as far as I know, where there is more of a tradeoff to low energy vs performance) suggests overall it's not worth doing.

So maybe there's an additional asterisk here that, when you boost one core to these higher speeds, you also lose the rest of the cores in the cluster which have to be slept. Which may be a good idea for some use cases, especially if you don't care about total energy usage (ie wallpower not battery).
But it is an important asterisk - it only works for running one lightly threaded app. (aka getting a good GB6 ST number?)

 

[name99]

If a 12P Oryon really only does 50% better than a 4P+4E M2, that's... pathetic. And if you take them at their word that single-core is substantially better than M2... that's even worse.

I think they've put out a lot of really confusing numbers and we won't really know what's what until we see some actual running systems (by which point M3 will likely have ben out for 6-9 months).

But if the single- and multi-core scores are even close to accurate, it suggests that they have a really serious scaling problem, worse than Apple did with the M1 Ultra.

I hope it does turn out to be a good chip - the more quality competition, the better. So far, I'm not feeling it.

There are many claims flying around.
The ONLY one that matters is the claim of substantially faster than M2 in ST,
M2 ST GB6: 2841
Oryon ST GB6: 3227 (presumably in overdrive mode, though it's unclear; they claim this 14% better performance at simultaneously 30% lower power)

(a) Ignoring the power issue (because it's not clear EXACTLY what they are claiming) they appear to be running their core about 15% faster than M2 and getting 15% better ST. So impressive that they are at M2 IPC, but they don't appear to be BEYOND M2 IPC.
Of course M2 IPC, but at higher frequency AND lower power would be very impressive -- if that's achieved...

(b) If you see someone talking about ANY number other than this ST number, you now know that they're an idiot who doesn't understand what's hard vs what's easy, and they can be tuned out. Adding lotsa cores is easy; Apple could trivially do it (and probably have for the M3 Max Pro, maybe even for the M3 base). Making one core fast is what's hardest.

(c) It's worth noting that for all their cores, they neither mention GPU numbers nor use a sophisticated (ie M2 Pro level) memory system. I suspect this machine will be disappointing when stressed hard by gamers, or even when stressed hard across all twelve cores. It looks like it's been designed (not from the beginning, but in terms of the final sets of choices) as a Windows machine driven by concerns to match Windows marketing, rather than as a fully balanced machine.

(d) Of course who can know anything given what we saw? Maybe in six months we'll see full systems and get a full picture.

 

[camotwen]

1. Higher overclocked performance cores, 2 more efficient cores in pro/max. Smarter utilisation of efficient cores and switching (already Sonoma does that in m1/m2). This will provide better performance beyond whatever benchmarks.
2. Ray tracing in GPU (hopefully)
3. 12gb ram modules, which means m3 pro starts from 24gb, tops at 48gb.

 

[leman]

M2 ST GB6: 2841
Oryon ST GB6: 3227 (presumably in overdrive mode, though it's unclear; they claim this 14% better performance at simultaneously 30% lower power)

It’s more interesting than that. They claim that they can get M2 performance (the 2840 points) using 30% less power than Apple. They made no claims about how much power they need for that 3200 score.

Furthermore, they claim to get around 15-16K in GB6 multi core while consuming 50 watts. M2 Pro gets 15K at 40 watts. This is a source of confuSion for multiple people, myself included, as I don’t understand how they can simultaneously be more efficient than M2 in single core at the same performance, and be less efficient than M2 in multi core.

 

[leman]

Btw, since we are on the topic of predictions, my prediction is ECC RAM.

 

[APCX]

Any guesses at GB Single Core scores? 3500 seems reasonable.

 

[leman]

Any guesses at GB Single Core scores? 3500 seems reasonable.

That would require the CPU frequency of 4.4ghz, which is a bit on a high side. But who knows?

 

[Confused-User]

So maybe there's an additional asterisk here that, when you boost one core to these higher speeds, you also lose the rest of the cores in the cluster which have to be slept. Which may be a good idea for some use cases, especially if you don't care about total energy usage (ie wallpower not battery).
But it is an important asterisk - it only works for running one lightly threaded app. (aka getting a good GB6 ST number?)

I thought that that was directly stated. The 4.3GHz speed was only achieved when 10 cores were sleeping.

 

[TigeRick]

I thought that that was directly stated. The 4.3GHz speed was only achieved when 10 cores were sleeping.

Source? Cause I am expecting two clusters to be active..

 

[APCX]

Source? Cause I am expecting two clusters to be active..

I believe their event video yesterday stated that they could only use two cores at that speed. I could, of course be misremembering!

 

[Confused-User]

I believe their event video yesterday stated that they could only use two cores at that speed. I could, of course be misremembering!

You are not. That's correct.

 

[TigeRick]

You are not. That's correct.

But two cores running at 4.3GHz does not mean the rest of the cores not running at 3.8Ghz. Especially the two cores are part of two clusters which has same power rails...

 

[Confused-User]

But two cores running aat 4.3GHz does not mean the rest of the cores not running at 3.8Ghz. Especially the two cores are part of two clusters which has same power trails...

It is precisely because cores generally share the same power rails and clocks within a cluster that it's a safe assumption that the cores not running at 4.3GHz are shut down (or, perhaps, almost/entirely idle, in some sort of low-power state?).

It possible they're not, but it's very unlikely. We won't know for sure until someone competent outside Qualcomm gets their hands on one.

 

[TigeRick]

It is precisely because cores generally share the same power rails and clocks within a cluster that it's a safe assumption that the cores not running at 4.3GHz are shut down (or, perhaps, almost/entirely idle, in some sort of low-power state?).

It possible they're not, but it's very unlikely. We won't know for sure until someone competent outside Qualcomm gets their hands on one.

I tend to believe Qualcomm will shut down or reduce the power of the last cluster cause they want to maintain power to the two clusters with high-speed cores. Remember Apple only designed the M2 Pro/Max with 8P+4E, 12 HP cores seems stretching power limit of SoC. That might explain why MT scores of X Elite consumes more power than M2 Pro as stated by @leman.

 

[Confused-User]

I tend to believe Qualcomm will shut down or reduce the power of the last cluster cause they want to maintain power to the two clusters with high-speed cores. Remember Apple only designed the M2 Pro/Max with 8P+4E, 12 HP cores seems stretching power limit of SoC. That might explain why MT scores of X Elite consumes more power than M2 Pro as stated by @leman.

"stretching power" - Not even close.

It is indeed a power issue, in the sense that thermals and battery are a power issue. But it's not power directly - Intel and AMD push WAY more power through their chips. (Edit to add: I can imagine they might have limits on power delivery per cluster. But if so that's a design choice, probably informed by the inability to usefully run more than one core/cluster at the max boost speed due to thermals.)

And of course, 12 P cores is a big part of why it uses more power than the M2. Obviously. But the efficiency of each core is a factor, and all the other parts of the chip (GPU, NPU, ISP, NoC, I/O, etc.) play a big part as well.

But as for what actually is going on when the two P cores are running at 4.3GHz - that's very likely pretty much the same as what Intel does when it clocks up its "favored core(s)" to max clock. Google that term for specifics.

 

[MayaUser]

Again all of that data how it will transfer to an 13" windows on arm laptop...because this is happening since 2016, 7 years...7 years of garbage use case of windows on arm....and in all events they show very nice charts for what one chip will do and when the final package arrives to me as an customer...its garbage, some of them had a relative good performance but very poor battery, or some had almost 16h battery life but garbage performance. Maybe im the only fool who always believed windows on arm Qualcomm chip and brought 2 different generations devices and they failed. Just because it was build by new people...new cores etc doesnt mean nothing to me anymore on paper...show me these charts to be true on the consumer end product otherwise... Apple is doing for the last 3 years what Qualcomm and Microsoft windows on arm couldnt after 7(not to include surface rt). And let me tell you, my windows on arm on the M2 is working night and days better than any current gen Qualcomm chip. Yes, its nice to talk and find out what you think...but i personally on my own money i felt on my fingers and on my personal time not to trust this anymore (even the translation windows layer is garbage compared to Rosetta2) until contrary proof

 

[senttoschool]

It’s more interesting than that. They claim that they can get M2 performance (the 2840 points) using 30% less power than Apple. They made no claims about how much power they need for that 3200 score.

Furthermore, they claim to get around 15-16K in GB6 multi core while consuming 50 watts. M2 Pro gets 15K at 40 watts. This is a source of confuSion for multiple people, myself included, as I don’t understand how they can simultaneously be more efficient than M2 in single core at the same performance, and be less efficient than M2 in multi core.

They could have significantly overclocked the SoC for the ST that beat M2 Max but declined to show power figures.

In addition, perhaps M2 Pro's efficiency cores boosted its MT without using much more power.

 

[leman]

They could have significantly overclocked the SoC for the ST that beat M2 Max but declined to show power figures.

I wouldn't be surprised if that 4.3Ghz boost incurs a substantial power cost, but I am more confused by the fact that they claim performance parity at 30% lower power. So how comes that there 12-core solution both uses more power and fails to deliver higher performance than M2 Pro?

An earlier report claimed that Qualcomm has had issues with their power management unit. So it's possible that the efficiency drops substantially at slightly higher power levels because the controller can't handle it.

In addition, perhaps M2 Pro's efficiency cores boosted its MT without using much more power.

An E-core is about as performant as 1/3 P-core, so I doubt this is it.

 

[MayaUser]

So Mark G says that M3 Pro will come with 8Perf cores and 6Ef and M3 Max will come with 12P cores and 4E cores and 40gpu cores with ray tracing
So the gains for both M3 Pro and M3 Max compared to M2 Pro/Max should be bigger than what we gain from M1 pro/max oto M2 Pro/max
So for me, another year another mac year will be, that QC BS it will be again dead on arrival for semi and proffesional use when it will come 6 months later in actual devices

 

[MRMSFC]

But it is an important asterisk - it only works for running one lightly threaded app. (aka getting a good GB6 ST number?)

It’s the end result of my generation’s education system, teach to the test!

So Mark G says that M3 Pro will come with 8Perf cores and 6Ef and M3 Max will come with 12P cores and 4E cores and 40gpu cores with ray tracing
So the gains for both M3 Pro and M3 Max compared to M2 Pro/Max should be bigger than what we gain from M1 pro/max oto M2 Pro/max
So for me, another year another mac year will be, that QC BS it will be again dead on arrival for semi and proffesional use when it will come 6 months later in actual devices

Well it’s too early to say.

I’m also skeptical that QC just happened to show up with a cpu this great despite being behind for the past few years, but I’m willing to give it a chance. Maybe they found something brilliant?

 

[leman]

I’m also skeptical that QC just happened to show up with a cpu this great despite being behind for the past few years, but I’m willing to give it a chance. Maybe they found something brilliant?

Well, they did buy the startup of the engineer who has designed all Apple CPUs up to date, so it’s hardly a surprise.

 

[Homy]

People say Apple is in BIG trouble or even doomed but the more important question is what would happen next if SnapDragon X Elite is in fact ”better” than Apple Silicon? Would that be a world changer? Would Mac users ditch all their Macs and go buy ARM PCs? Would Apple go bankrupt? The answer is no. Because ARM PCs don’t run macOS and using Macs, iPhones or iPads is more than just having the fastest CPU or GPU. Neither is Apple always competing with everyone. They compete with themselves by doing their best and what’s best for their users.

I also agree with Ars Technica saying ”No matter how good the hardware is, it's only part of the puzzle. The Apple Silicon transition succeeded in part because Apple's Rosetta 2 compatibility layer made x86-to-Arm code translation mostly invisible to the user, and because the company had already broken legacy technologies like 32-bit app support and certain kinds of drivers in previous macOS releases. Qualcomm will be relying on another company's software to succeed, namely the Arm version of Microsoft's Windows 11. Sometimes, even apps with Arm-native versions will download and install their x86 versions by default on the Arm version of Windows, leaving performance on the table. Drivers written for the x86 versions of Windows, and the hardware that relies on those drivers, won't work. Many games, particularly those that depend on non-DirectX 12 APIs and anti-cheat software, won't run. Backward-compatibility is one of the defining features of the x86 versions of Windows; it's much more limited in the Arm version".

 

[Chuckeee]

Does it seem like this snapdragon processor is following the procedure that Intel typically does?

Lots of PR about the chip itself and a continuous stream of PR as that ship gets further and further along but it will be six months to a year before there’s an actual consumer product that hosts the chip.

Also, do you think they will be following the current x86 procedure that when (and if) a commercial board is available, will this chip be mounted via a socket?

 

[leman]

Lots of PR about the chip itself and a continuous stream of PR as that ship gets further and further along but it will be six months to a year before there’s an actual consumer product that hosts the chip.

I think this makes a lot of sense from the marketing perspective. They need to generate a lot of hype and get developers on board to ensure that there will be enough high quality software compatible with their platform on release. Right now Windows on ARM is in somewhat sad state, so releasing a high-end product outright can be very risky. Qualcomm seems to be playing it by targeting the creatives first, which is a smart move, as there is a limited number of players in the field and their hardware will work well for this kind of work.

Also, do you think they will be following the current x86 procedure that when (and if) a commercial board is available, will this chip be mounted via a socket?

It most certainly won’t be. Qualcomm’s target market are premium laptops. I also doubt that they will be cheap. They will be competing directly with Apples MBP. I think Qualcomms target customer is a creative or a tech enthusiast/developer who wants a fancy fast energy efficient laptop but prefers Windows.

I am curious to see if their strategy will work out. They’d certainly do fine against M2 Pro, but I wonder how good their value proposition will be against M3?