M3 core counts and performance

Some things I wanted to comment on...

- if that 3227 GB6 run holds up as a normal result for an unloaded M3 machine, both the overall geomean score and nearly all the individual benchmarks show better scaling than the clock speed ratio between M3 and M2.

- Did anyone else notice what might have been Johny Srouji's biggest flex? It wasn't anything technical, it was when he made a point of mentioning that this time, his department was able to deliver three M3 family members to the market at the same time. I think that's pretty significant. The staggered release schedules for the M1 and M2 families left some questions open about how much Apple was going to invest in growing the AS team to support the Mac. Launching all of A17 Pro, M3, M3 Pro, and M3 Max in Q4 2023 says they've grown their organization a lot.

- along those lines, unlike M1 and M2, the M3 Pro mask art isn't just a slightly edited crop of the M3 Max artwork. That was a significant labor / compute time saver for the M1 and M2 families, but this time they eschewed it in favor of tailoring the M3 Pro for lower costs and the M3 Max for being, well, "max".
 
I think M3 will also be short lived, N3B is expensive and low. So M4 with A18 N3E tech might come Q4 2024/ Q1 2025.

Remember that leaked up to A19 and M5 IDs. Apple is already testing M4 chips.
 
Still rising
Single core: 3227
I think the ~3200 score results are going to be the “real” score. There’s a few results in that range already, it’s what we expected, and not many things can make the CPU run faster (other than extreme cooling, as we saw with the A17 Pro results).

In the context of Apple SoCs this upgrade may not look like a mind blowing update but in the context of competitors CPUs I think the gap hasn’t been wider since the M1 launch (Intel pushed some important new CPUs around the time the M2 was in the market). The Intel 14900K, a just-released 24 core (8+16) desktop CPU with a TDP of 125W (325W of maximum turbo power, per The Verge) has a Geekbench score of 3125 (Single core) and 22055 (Multi core). The M3 Max is roughly in that ballpark, being a laptop CPU, using a tiny fraction of that power. Intel hasn’t been able to get ahead in performance even with their desktop CPUs this time.
Snapdragon X Elite is admittedly much closer to Apple’s SoC in performance (although the reported TDP is 80W, so I don’t quite believe their efficiency claims), but it hasn’t been released yet.
 
I think the ~3200 score results are going to be the “real” score. There’s a few results in that range already, it’s what we expected, and not many things can make the CPU run faster (other than extreme cooling, as we saw with the A17 Pro results).

I suspect that this is the Max running at higher frequency than detected by Geekbench. Otherwise we would have already seen scores in 32xx for the base M3.

In the context of Apple SoCs this upgrade may not look like a mind blowing update but in the context of competitors CPUs I think the gap hasn’t been wider since the M1 launch (Intel pushed some important new CPUs around the time the M2 was in the market). The Intel 14900K, a just-released 24 core (8+16) desktop CPU with a TDP of 125W (325W of maximum turbo power, per The Verge) has a Geekbench score of 3125 (Single core) and 22055 (Multi core).

Intel will be struggling to get to higher ST, as they just tweak the clocks further and increase the power consumption. But Meteor Lake should be out soon, and it might allow the to increase the clocks some more. AMD will likely have a larger boost with Zen5.
 
I suspect that this is the Max running at higher frequency than detected by Geekbench. Otherwise we would have already seen scores in 32xx for the base M3.
Has this happened before? I don’t recall Geekbench misreporting frequency. I though this might have been because of the higher bandwidth of the Max chip.

Intel will be struggling to get to higher ST, as they just tweak the clocks further and increase the power consumption. But Meteor Lake should be out soon, and it might allow the to increase the clocks some more. AMD will likely have a larger boost with Zen5.
Oh, I thought Meteor Lake wouldn’t be out until late 2024. After a quick Google search, looks like the plan is to launch Meteor Lake (desktop) late 2024, and Meteor Lake (laptop) late 2023. So unless their laptop chip somehow beats their (previous gen) desktop chip in single core, the situation shouldn’t change much, their highest-performance chip is probably still going to be the 14900K.
 
Has this happened before? I don’t recall Geekbench misreporting frequency.

I don't even know how GB gets the frequency to begin with, they probably rely on the same private APIs to estimate it as I did in my A17 investigations.

I though this might have been because of the higher bandwidth of the Max chip.

I would be shocked if bandwidth has anything to do with it. None of the GB6 subsets are bandwidth-heavy.
 
Has this happened before? I don’t recall Geekbench misreporting frequency. I though this might have been because of the higher bandwidth of the Max chip.
GB seems to use an API which reports the time averaged frequency over some interval. If the CPU dips below its maximum freq at any time during that interval, the value it reports is lower than expected.

Such dips are probably common. Each P cluster in a M1 Max has a true top speed of 3228 MHz, and this speed is permitted only while 3 of the 4 cores in the cluster are asleep. When the macOS scheduler has two threads it needs to schedule on P cores, it prefers to wake more than one core in the first cluster rather than lighting up both clusters (presumably because this is less power).

So, when you look for M1 Max results in the GB browser, you'll find lots of clock speed numbers that are close to the correct maximum clock of 3228, but never quite get there.

You'll also find a significant number that are in the region of 100-200 MHz lower than 3228, which I'm not quite sure how to interpret.
 
Lol, Geekerwan label their geekbench runs now!?
1699015215573.jpeg
 
Still no M3 Pro results 😔
I’m sure it’ll be fine, but I’m dying to know how multi-thread performance stacks up going from M2 Pro 8P+4E to M3 Pro 6P+6E.
 
Last edited:
- Did anyone else notice what might have been Johny Srouji's biggest flex? It wasn't anything technical, it was when he made a point of mentioning that this time, his department was able to deliver three M3 family members to the market at the same time. I think that's pretty significant. The staggered release schedules for the M1 and M2 families left some questions open about how much Apple was going to invest in growing the AS team to support the Mac. Launching all of A17 Pro, M3, M3 Pro, and M3 Max in Q4 2023 says they've grown their organization a lot.

I came into the stream too late to catch that line, but I did notice the results. They’ve gone from a 1 year gap (M1 -> M1 Max), to 6 months (M2 -> M2 Max) to simultaneous (M3).

I think back before the MBPs fully transitioned there was all the speculation as to what release cadence we’d get at the other place. A number of folks assumed it would always be staggered because the iPad releases were done that way. But now that we are here, it’s pretty clear that Apple wants to close this gap.
 
Oryon vs. M3 Max is really something https://browser.geekbench.com/v6/cpu/compare/3372046?baseline=3326437

Oryon matching M3 single thread performance seemed impossible just a few weeks back 🤯 We don't know how efficiency stacks up of course, but it's hard to imagine Oryon being a total disaster on that front.

All things considered, the M3 lineup is solid. I think it's fair to say everyone here expected more single thread performance gain? I'm still curious what happened on that front. Like:
  • Was M3 supposed to be the actual successor to M1? Was M2 a product of M3 delays?
  • Why are performance gains from A15 > 16 > 17 mostly coming from clock speed? Have they hit a wall on instruction level parallelism as some have speculated? Or is the "next big core" stuck in development? Where is armv9 and SVE(2)?
 
One thing to be aware of with these high Oryon scores, is that all of them were achieved with massive cooling from the fans due to lack of fan control under Linux currently. We’ll see what happens when they are released, but so far, the Windows scores are considerably lower.
 
One thing to be aware of with these high Oryon scores, is that all of them were achieved with massive cooling from the fans due to lack of fan control under Linux currently. We’ll see what happens when they are released, but so far, the Windows scores are considerably lower.
Yeah I know. Geekbench on Fisher Price OS typically doesn't score as high as Linux/macOS anyway. So while the Linux benchmark benefits from more thermal headroom, I'm inclined to blame the difference more on Windows power management and scheduling.

Edit: plus, unless Qualcomm has gone full Intel and the cores are using >20W at peak boost, thermal headroom should be fine 😜
 
Yeah I know. Geekbench on Fisher Price OS typically doesn't score as high as Linux/macOS anyway. So while the Linux benchmark benefits from more thermal headroom, I'm inclined to blame the difference more on Windows power management and scheduling.

Edit: plus, unless Qualcomm has gone full Intel and the cores are using >20W at peak boost, thermal headroom should be fine 😜
Always happy to bash windows, but in this case, it’s not clear how much of a factor that is. What Anandtech specifically mentioned, is the fact that Linux is getting huge cooling due to unfinished fan control under Linux. While I’m prepared to believe Linux will score higher, I’m not convinced scores will be as high as currently boasted.
 
Always happy to bash windows, but in this case, it’s not clear how much of a factor that is. What Anandtech specifically mentioned, is the fact that Linux is getting huge cooling due to unfinished fan control under Linux. While I’m prepared to believe Linux will score higher, I’m not convinced scores will be as high as currently boasted.
We'll see. I'm happy to lean towards Windows pulling down the ST score at this point (I agree that increased thermal headroom definitely helped the multi-core score though)

Another data point (granted, these tests are using virtual machines and GB5 not GB6):
macOS M1 Max native = 1773 points https://browser.geekbench.com/v5/cpu/10662889
Parallels Linux VM on M1 Max = 1690 points (~94%) https://browser.geekbench.com/v5/cpu/10660486
Parallels Windows VM on M1 Max = 1546 points (~87%) https://browser.geekbench.com/v5/cpu/10664539
 
Back
Top