This seems right since the graphs Apple showed during the keynote and in their press release, they always show the M1 Pro and M1 Max as having the same performance/power for the CPU cores. That would imply that the CPU cores can't use more than the 200 GB/s memory bandwidth of the M1 Pro.
Got a tip for us?
Let us know
Become a MacRumors Supporter for $50/year with no ads, ability to filter front page stories, and private forums.
Geekbench score for M1 Max. It's great
- Thread starter Kung gu
- Start date
- Sort by reaction score
Wasn't this the case with the M1 Mini, MBP, and MBA? Each one was less capable of running at peak speed due to thermal constraints of each chassis? Seems reasonable the same would exist with the 14" and 16" MBPs.
The first Geekbench compute result is out:
OpenCL Score: 60,167
OpenCL Score: 60,167
MacBookPro18,2 - Geekbench
Benchmark results for a MacBookPro18,2 with an Apple M1 Max processor.
browser.geekbench.com
That 400GB/s is primarily there for the GPU not the CPU cores. Apple has limiters so the bandwidth can't all be skewed in on solidary narrow single direction.
The AMD and Intel CPUs also don't have an upper mid-range GPU hanging off the the primary RAM resource either.
If it is a large enough dataset that is stilling on the disk then having 2-3 PCI-e v4 SSDs to upload off of will fly faster. RAM speed is mitigated if the data is off on persistent storage.
…the chip is the same and the thermal envelope is what dictates the performance. You’re stating it like that wasn’t the intention of Apples design.
The M1 GPU runs at 1.28GHz clock, so for it to be able to pull in 400GB/s, the GPU's L1/L2/L3 caches (assuming running at full clock speed of 1.28GHz) would have to pull in over 2500 bits every clock to fully saturate the 400GB/s bandwidth. Quite unlikely for the data path to be so wide I would think. So I guess you're right in that there's natural limits for all the IPs in the M1 SoC. The question would be how wide is the internal bus width to the caches.
Impressive for a mobile chip, but not quite there to replace even the older generation of Pro GPUs in the Mac Pro line.
On Metal it’s up to 102,937 for the 1xVega II (over ~205K if doubling the score for Duo): https://browser.geekbench.com/v5/compute/414168
1xVega II on OpenCL, up to ~ 82K: https://browser.geekbench.com/v5/compute/658281
The Apple Silicone Mac Pro can’t come soon enough.
Last edited:
Yes, the teardowns will help eventually.
Except not on the battery thing. The MBP 13" four port (Intel ) 58Whr ... MBP 14" 70Whr. There is only a 12Whr difference there. (about 20%). There is no movement on the MBP 16" ( somewhat max'ed out on carry on limitations. ) [ looking at it from 16" -> 14" battery transition probably is deceptive. ]
The MBP 14" is wider (~2%) and deeper (~2%) and heavier than the MBP it replaces , so Apple added a bit of thermal heft. ( if the twin fans are identical in both system then would be even. But there too.... the MBP 13" only had one fan. So more power once under heavily load. )
if Apple's pictures aren't a kluge the side vents on the 16" were adjusted to be about the same as the 14"
14"
16"
The 16" is deeper so that side slot could run longer. But the Mag Safe ( and HDMI ) port is shifted closer to the front.
( those side slots are coupled to the audio system so suggestive that they are the same. Same specs for audio so not sure why they would use different components over the same length).
Apple's air flow pictures suggest there is also an air inlet on the back. output and ingest near each other isn't the best thermal design guideline but 16" may be slightly better off if they are gapped wider apart. ( of course Apple is super eager not so show any substantive thermal vents because it is a "magical" system. So no pictures from behind. ).
The "stretch" from the 14 to the 16 is more so by the hinge where there is no battery. the thicker chassis is probably playing a bigger contribution to get the 16 a bigger battery. (wider also) . That non-battery area expansion is a bit suggestive that the fans are slightly bigger (and better) in the 16" than the 14". ( maybe some more funky bigger hinge stuff going on, but the fan/heatsinks are primarily in that zone. )
That is low? Laptops with the RTX 3070 score 125k, so double the score.
And this is not even laptops with the RTX 3080.
It’s not a limitation of ARM architecture, it’s the limitation of Apple’s wide micro-arch.
Why is that a limit of the ARM architecture? Surely it's a limit of the M1 cores as designed by Apple.
Frequency limits are set by the desired TDP of the cores (generally trading lower TDP & hence power consumption against IPS), and the limits of the fabrication process (considering effects like quantum tunneling across very small silicon features).
I'm not aware of any frequency limitation in ARM that means you couldn't theoretically run it at the same speeds as other CPUs. They would just lose some of their advantages at these frequencies.
Update: here's a doc from ARM that mentiond 3.5GHz for their Cortex-X2 architecture - https://community.arm.com/arm-commu...d-processors-blog/posts/first-armv9-cpu-cores
This was even mentioned several years ago in 2017: https://wccftech.com/intel-arm-10nm-22nm-collaboration/
Last edited:
You joking? 😂 Anyhow, the burden of proof is on you, you are making some outlandish claims again and then deflecting. Prove us that no ARM CPU is ever capable of reaching higher clocks.
Couple of years ago you were probably claiming that „x86 CPUs can’t break 4Ghz“ or „electric cars are impossible“ 😁
Keep in mind this is was a openCL test. Not a Metal one. Should be interesting to see metal.Around a 1650ti was my prediction.
So its around a GTX 1660 ti in openCL.
I wonder what it is in Metal
It will probably be a few percent higher in Metal, but again, this is just a single sample, so no point in drawing premature conclusions. It is a disappointing number though.
The real test will be how well it compares in real-world workload tests instead of benchmarks.
I'd also like to see 16 and 24 core GPU numbers. My hope was that the 16-core would be about the same as the AMD 5600M, with the 32 core about twice as much (seems to be only 1.5x the 5600M).
MacBook Pro review embargo lifts on Monday at 9am so we should start seeing more benchmarks posted between now and then for various configurations. Geekbench may get updated as well to bette support the new chips.
Things are already looking very good though!
Things are already looking very good though!
I put some links (one from ARM) showing 3.5GHz design frequency for their more recent architectures (Cortex-X2, Neoverse), so 3.2GHz is clearly not a limit 
The question is why the "architecture" would be limited to any frequency? The fabrication may very well be....but the architecture is all about how the instructions are processed. There might be some race-conditions that create limits, but in theory you could just crank up the clock and it would just run faster - and hotter, of course. ARM no doubt gives design guidelines to implementers, and I expect there is a sweet spot beyond which ARM compares less favorably to Intel/AMD, which is while most chips are in the 2-3.2GHz range.
The question is why the "architecture" would be limited to any frequency? The fabrication may very well be....but the architecture is all about how the instructions are processed. There might be some race-conditions that create limits, but in theory you could just crank up the clock and it would just run faster - and hotter, of course. ARM no doubt gives design guidelines to implementers, and I expect there is a sweet spot beyond which ARM compares less favorably to Intel/AMD, which is while most chips are in the 2-3.2GHz range.
That's still lower than what x86 can do.I put some links (one from ARM) showing 3.5GHz design frequency for their more recent architectures (Cortex-X2, Neoverse), so 3.2GHz is clearly not a limit
again you deflect..you were proven wrong...now you tell us something else , so based on that, your opinion lose its trust
Frequency isn't usually a limit of architecture. There's a sweet spot for frequency/watt since it isn't linear and more exponential. It's better to target the frequency/watt sweet spot then scale out with more cores. I actually disable boost clocks on x64 because of this since the slight gain in frequency and performance isn't worth the more significant power consumption increase and heat output.