Apple's M5 Max Chip Achieves a New Record in First Benchmark Result

[rog]

I need an M5 Max iPod Shuffle

 

[chars1ub0w]

The M1-M2 Pros/Maxes had the same CPU core count.

No difference in multi score on Geekbench between M1 Pro and M1 Max.

 

[chucker23n1]

The improvement over M4 max is only 14%? That’s not a lot. In the worst intel days the improvements are around 5%

Given that the Max’s CPU configuration is now the same as the Pro’s, this was inevitable.

The Pro goes from 10p4e to 6s12p.

The Max from 12p4e to 6s12p.

So the Pro’s bump will be better, but only because, relatively speaking, the Max used to be better than the Pro.

 

[conmee]

Thanks! Then, looking at the CPU performance only, I wonder how much double the memory bandwidth affects the benchmarks.

Going from 20 GPU cores to 40 GPU cores doubles the memory bandwidth because the chipset has double the memory controllers?

I don't think the memory bandwidth is based on number of GPU cores. I think Apple can just dial up the bandwidth based on CPU clock and type/number of memory chips (someone can correct me if I'm wrong). Keep in mind the M5 Max gives you 614GB/s whether you get an M5 Max with 32 GPU cores or 40 GPU cores. Either way, should have a big impact on AI benchmarks/workloads. I used Geekbench AI on the M5 I returned as a baseline, so with the M5 Pro being double the bandwidth, I expect to see double(?) the results...

For reference, the M5 scored CPU Single Precision/Half Precision/Quantized of 5337/8497/6929 and GPU Single Precision/Half Precision/Quantized of 13659/25668/24464. How these numbers relate to real-world workloads, not sure. Someone else would have to weigh in.

 

[conmee]

Given that the Max’s CPU configuration is now the same as the Pro’s, this was inevitable.

The Pro goes from 10p4e to 6s12p.

The Max from 12p4e to 6s12p.

So the Pro’s bump will be better, but only because, relatively speaking, the Max used to be better than the Pro.

They may start tweaking clock, power, cache, other items like Intel did for i3/i5/i7/i9 differentiation.

 

[User 6502]

This doesn’t look like a huge upgrade, in fact it’s a bit disappointing. Oh well, let’s hope M6 family brings more innovation. And most importantly more ram for local models.

 

[conmee]

This doesn’t look like a huge upgrade, in fact it’s a bit disappointing. Oh well, let’s hope M6 family brings more innovation. And most importantly more ram for local models.

Looks like they're hanging their hat on the improvements to AI workloads via the neural accelerators in the GPU cores and the increased memory bandwidth. The GPU performance is impressive. Maybe they just wanted to get a successful launch of the Fusion Architecture out the door with M5 and as you say, we'll see more innovation and CPU combinations and performance with M6.

 

[svish]

Good to see these numbers. Apple Silicon chips are very powerful and the new Macs will be great machines.

 

[WilliamG]

The improvement over M4 max is only 14%? That’s not a lot. In the worst intel days the improvements are around 5%

Apple absolutely knocked it out of the park with the M1 Max, compared to any MBP before it. And since then it’s just been iterative updates.

 

[Frantisekj]

How are you managing to get by without any neural accelerators or ray tracing? 😉

idk how neuropsychguy but I do not do gaming or 3D and use my own neural accelerators 😀, otherwise 2012 MBP 😀

 

[chucker23n1]

Compared to M1 Max, 77.3% single and 131% multi.

Assuming M5 Pro scores are roughly the same, compared to M1 Pro, 79.3% single and 136.8% multi.

Not bad for ~4.5 years.

 

[conmee]

idk how neuropsychguy but I do not do gaming or 3D and use my own neural accelerators 😀, otherwise 2012 MBP 😀

Do you have the GeForce GT 650M in that 2012 MBP for some CUDA goodness? I kept my 2014 15" MBP thinking I'd upgrade when Apple went back to Nvidia GPUs... we know how that worked out.

 

[Frantisekj]

Do you have the GeForce GT 650M in that 2012 MBP for some CUDA goodness? I kept my 2014 15" MBP thinking I'd upgrade when Apple went back to Nvidia GPUs... we know how that worked out.

I do have the GeForce GT 650M but I probably never used any CUDA goodness... or idk about it.

 

[macbookj0e]

I have an M4 air which is perfect for me but the geek in me really wants an m5 max for absolutely no reason 😅

 

[wanha]

Really hope they come out with a M5 Ultra with 1TB+ memory Studio.. literally everyone wants to run local LLMs. Come on Apple!

I think you literally mean "figuratively everyone"

 

[wanha]

The improvement over M4 max is only 14%? That’s not a lot. In the worst intel days the improvements are around 5%

14% improvement year over year roughly equates to the performance doubling every 5 years

At 5%, it would take 14 years to double the performance

 

[chucker23n1]

looking at the CPU performance only, I wonder how much double the memory bandwidth affects the benchmarks.

We don't know for the M5 Pro/Max, but for the M1 Pro/Max, higher memory bandwidth was almost exclusively beneficial for the GPU cores, not the CPU cores.

The M1 Pro had 204.8 GiB/s (Anandtech says GB, but I'm assuming GiB), and the M1 Max had 409.6 GB/s. But as you can see here,

• for single-core operations, you barely get above 100
• even with all ten cores saturated, you get to about 242.95

So about 40% of that bandwidth aren't being used by the CPU.

Again, that's for the M1 Pro/Max. It's possible the design has improved enough that the story is different for the M5 Pro/Max, but I think you'll see a similar plateau where, whether you use two cores or all 18 of them, it plateaus well below max — unless you heavily use the GPU.

 

[ralph_sws]

This doesn’t look like a huge upgrade, in fact it’s a bit disappointing. Oh well, let’s hope M6 family brings more innovation. And most importantly more ram for local models.

Each subsequent generation will have smaller performance gains. You can't cheat physics; you can't infinitely increase clock speeds or add more cores. At some point, you'll start experiencing shorter battery life and throttling. The m5 already has poor heat dissipation. 2nm will help a little, but only temporarily.

 

[chucker23n1]

Going from 20 GPU cores to 40 GPU cores doubles the memory bandwidth because the chipset has double the memory controllers?

I don't think the memory bandwidth is based on number of GPU cores.

It's not; it's because of the memory controllers, as the person you were responding to said. Certainly that was the case for the M1 Pro/Max. Here's the M1 Pro:

Notice there's two memory controllers here.

And here's the M1 Max, which adds two more underneath.

Each memory chip was simply tied to a memory controller. As a result, they ran in parallel, which effectively increases the bandwidth.

 

[starlabjump]

The numbers are excellent, but the M5 Max chip would be more than I need.


A regular M5 would be sufficient for me, and if I were a bit greedy, the M5 Pro would be ideal.


That said, I will probably wait for the 2-nm M6 or a later generation, since my M3 Pro is still working great for me.

 

[datagov63]

This doesn’t look like a huge upgrade, in fact it’s a bit disappointing. Oh well, let’s hope M6 family brings more innovation. And most importantly more ram for local models.

This is for the 18/32 base version of the M5 Max laptop. In the Studio the numbers will be a little higher. Looks like a 30% bump in multi-process CPU performance and the GPU should be about the same. If you are coming from a M1 Max or M2 Max studio this will be a nice upgrade. If you have an M4 Max you probably won't notice much difference and this is as it should be.

 

[Nyx1]

It seems Apple has stagnated after a search, and the Core rebranding seems like a complete nonsense. Johnny Srouji is done, and we need new blood on the CPU team, like Jim Keller.

 

[Swm61230]

Not available for preorder in all countries. Wish you would do some better investigating before you report such nonsense.

 

[iBluetooth]

The improvement over M4 max is only 14%? That’s not a lot. In the worst intel days the improvements are around 5%

The improvements are much larger in AI or expecting up to 4x previous model. But, for sure we all like to have the first M1 jump of 80%!

 

[iBluetooth]

It seems Apple has stagnated after a search, and the Core rebranding seems like a complete nonsense. Johnny Srouji is done, and we need new blood on the CPU team, like Jim Keller.

It's hard to keep up the first advantage as now AMD and Intel have woken up and after 5 years will be ready with an answer to the first M1 revolution. But I would never think of saying Srouji is doing a bad job and Jim Keller is busy with his AI torrent chip company.