First Geekbench for base M3

[MayaUser]

So if these are accurate
M3 faster than M1 Pro
M3 Max as fast as M2 Ultra and faster than M1 Ultra
At this pace...M3 Ultra will be..

 

[jeanlain]

It looks like the new GPUc core can bring improvements on specific tasks using mesh shaders or RT, and perhaps very complex shaders benefiting from of dynamic caching, otherwise, it isn't much faster than its predecessor.

 

[senttoschool]

3nm was advertised with 15% higher performance at the same power, so this is what we are getting. Did you expect something else?

Absolutely I was. Hell, if I dig into some of your old posts, I'm sure you were expecting more ST than this.

3nm isn't just 15% faster at the same ISO power. The logic density is 1.7x more as well with slightly improved SRAM cache. We're also going up from A15 to A17 Pro. All in all, I expected much better performance than this.

 

[headlessmike]

So..Geekbench 6 multi-core scores for CPU performance:

• M2 Ultra: 21,182 (Mac Pro, averaged)
• M3 Max: 21,084 (16-inch MacBook Pro, highest score listed so far)

The M3 Max has 4 more P cores than the M2 Max (same number of E cores). So, part of the improvement is because the number of cores is somewhere between the M2 Max and M2 Ultra. The other part is down to improved designs and higher clock speeds thanks to improved power efficiency (presumably).

 

[hagjohn]

Absolutely I was. Hell, if I dig into some of your old posts, I'm sure you were expecting more ST than this.

3nm isn't just 15% faster at the same ISO power. The logic density is 1.7x more as well with slightly improved SRAM cache. We're also going up from A15 to A17 Pro. All in all, I expected much better performance than this.

What is the chip, you produce, performance level?

 

[leman]

Absolutely I was. Hell, if I dig into some of your old posts, I'm sure you were expecting more ST than this.

I was indeed, but at a higher power consumption. It seems Apple decided to stay with the same power per core. I also hoped to see a bigger IPC increase. So yes, I do find some of the aspects of this upgrade underwhelming, but I don't none of that has to do with a 3nm process.

 

[Pressure]

I was indeed, but at a higher power consumption. It seems Apple decided to stay with the same power per core. I also hoped to see a bigger IPC increase. So yes, I do find some of the aspects of this upgrade underwhelming, but I don't none of that has to do with a 3nm process.

It could still be reserved for the high-end desktop SoCs.

Although I think Apple should just make bigger and more power hungry desktop SoCs for that segment anyway ... or give us an Ultrafusion GPU add-on instead of just doubling up SoCs with it 😅

 

[leman]

It could still be reserved for the high-end desktop SoCs.

I was hoping to see that, but now I am more sceptical.

Although I think Apple should just make bigger and more power hungry desktop SoCs for that segment anyway ... or give us an Ultrafusion GPU add-on instead of just doubling up SoCs with it 😅

Not in this generation. There is a distinct possibility that we will get an M3 Extreme Mac Pro though.

 

[Confused-User]

Absolutely I was. Hell, if I dig into some of your old posts, I'm sure you were expecting more ST than this.

3nm isn't just 15% faster at the same ISO power. The logic density is 1.7x more as well with slightly improved SRAM cache. We're also going up from A15 to A17 Pro. All in all, I expected much better performance than this.

I was indeed, but at a higher power consumption. It seems Apple decided to stay with the same power per core. I also hoped to see a bigger IPC increase. So yes, I do find some of the aspects of this upgrade underwhelming, but I don't none of that has to do with a 3nm process.

Once the A17 shipped it was crystal clear that we weren't going to see a major IPC boost in the M3. Clearly they had to trade their winnings in IPC back to get faster clocks. What could possibly make you think we'd see big IPC gains, when most of the gains in the A16 and A17 were from clocks?

But we don't know what the clock in the M3 will top out at, in a desktop form factor. Leman talked about 4.3GHz; there's no reason to think we won't see that in the Studio and Pro (and maaaaaybe the Mini). Even faster isn't out of the question, though it would be a little surprising.

And yes, I too am disappointed, to the extent that I expected big IPC wins earlier this year. But the other thing I expected was for them to work on scaling for multicore tasks and the initial indications are that they exceeded all expectations there- though we need a bit more data before we call the game.

 

[leman]

Once the A17 shipped it was crystal clear that we weren't going to see a major IPC boost in the M3. Clearly they had to trade their winnings in IPC back to get faster clocks.

What would be the mechanism by which IPC is traded for higher clocks? Dougall Johnson reported latency on some operations actually reducing rather than increasing.

I would rather speculate that there is some other bottleneck in the core. Maybe cache bandwidth or the amount of load/store units. Should this be the case we might see some IPC improvements in the future if Apple redesigns these parts. Another explanation is that the IPC is already approaching the limits imposed by code and that higher core utilization will only be noticeable in some corner cases.

But we don't know what the clock in the M3 will top out at, in a desktop form factor. Leman talked about 4.3GHz; there's no reason to think we won't see that in the Studio and Pro (and maaaaaybe the Mini).

To be honest I’m getting less and less optimistic about this. But let’s see. Maybe the larger red tops will feature overclocked “Pro/Max Plus” chips 😅

 

[Confused-User]

What would be the mechanism by which IPC is traded for higher clocks? Dougall Johnson reported latency on some operations actually reducing rather than increasing.

Traditionally, as I'm sure you know, the trade is made in the form of a longer pipeline. I have no idea what they actually did though, and I wouldn't care to speculate - I don't know enough and have zero data. But it does seem clear that they've made major changes which, all other things being equal, would be expected to give them some IPC boost. They're not getting it. (Or rather, they got it, and then had to give some back in other areas). What else would they give that up for?

I would rather speculate that there is some other bottleneck in the core. Maybe cache bandwidth or the amount of load/store units. Should this be the case we might see some IPC improvements in the future if Apple redesigns these parts. Another explanation is that the IPC is already approaching the limits imposed by code and that higher core utilization will only be noticeable in some corner cases.

Funny you should mention cache, I had just brought that up in the other thread a few minutes before. It's certainly possible. In any case, they have a bunch more tricks up their sleeves, so I'm sure we'll see some more gains in the future (unless they need to trade them for even higher clocks, I guess).

 

[Macintosh IIcx]

I would rather speculate that there is some other bottleneck in the core. Maybe cache bandwidth or the amount of load/store units. Should this be the case we might see some IPC improvements in the future if Apple redesigns these parts. Another explanation is that the IPC is already approaching the limits imposed by code and that higher core utilization will only be noticeable in some corner cases.

Regarding cache, I think that is where we will see a jump in the future. One important thing to note about N3B is that SRAM / cache density didn’t get much better from N5 whereas logic density did. So that looks like an obvious design compromise - take what you are given. The additions to the arch in A17P / M3 should give higher uplift in theory, so the P CPU might very well be a bit restrained by cache as things stand now. But not all workloads will suffer, so given the design limitations of N3B, I think it is fine.

Note that N3E should be even worse regarding SRAM density, maybe it is the same as N5, can’t remember.

Edit: I based my memory on this: https://www.anandtech.com/show/1883...n-schedule-n3p-n3x-deliver-five-percent-gains

 

[Kristain]

Regarding cache, I think that is where we will see a jump in the future. One important thing to note about N3B is that SRAM / cache density didn’t get much better from N5 whereas logic density did. So that looks like an obvious design compromise - take what you are given. The additions to the arch in A17P / M3 should give higher uplift in theory, so the P CPU might very well be a bit restrained by cache as things stand now. But not all workloads will suffer, so given the design limitations of N3B, I think it is fine.

Note that N3E should be even worse regarding SRAM density, maybe it is the same as N5, can’t remember.

Has it been established that M3 definitely is A17 as a few sources were saying it was possibly based on A16?

 

[Weaselboy]

Moderator Note:

Also see the discussion in the news thread:

https://forums.macrumors.com/threads/first-benchmark-results-surface-for-m3-chip-in-new-macs.2409507/

 

[leman]

Has it been established that M3 definitely is A17 as a few sources were saying it was possibly based on A16?

All evidence (of which there is very little) points to A17 and M3 sharing the same CPU and GPU cores. Three are no sources saying that M3 was based on A16, just some folks making random unsubstantiated claims.

 

[Pressure]

Moderator Note:

Also see the discussion in the news thread:

https://forums.macrumors.com/threads/first-benchmark-results-surface-for-m3-chip-in-new-macs.2409507/

It feels like it is an entirely different species that comment in the news section threads and these threads 😅

 

[APCX]

Not Geekbench, but as it’s the only new result I’ve seen, there is a New Gfxbench result. 446fps Aztec High Tier offscreen. Don’t take it too seriously as Gfxbench has some issues.

GFXBench - Unified cross-platform 3D graphics benchmark database

The first unified cross-platform 3D graphics benchmark database for comparing Android, iOS, Windows 8, Windows Phone 8 and Windows RT capable devices based on graphics processing power.

gfxbench.com

 

[Populus]

There are already benchmarks from the M3 and M3 Max, but I’m really impatient to see the M3 Pro benchmarks, for a couple of reasons. First and foremost, for me it’s the sweet spot of raw performance at a still relatively affordable price, so it is my target device, wether the M2 Pro or the M3 Pro Mac mini. And, secondly, because the M3 Pro is the only chip that has reduced the performance cores in favor of the efficiency ones, and it has actually 3 billion transistors less than the M2 Pro. And the memory bandwidth issue. So, I need to see the benchmarks to know wether it is worth waiting for the M3 Pro Mac mini, or just getting an M2 Pro mini.

 

[Pressure]

There are already benchmarks from the M3 and M3 Max, but I’m really impatient to see the M3 Pro benchmarks, for a couple of reasons. First and foremost, for me it’s the sweet spot of raw performance at a still relatively affordable price, so it is my target device, wether the M2 Pro or the M3 Pro Mac mini. And, secondly, because the M3 Pro is the only chip that has reduced the performance cores in favor of the efficiency ones, and it has actually 3 billion transistors less than the M2 Pro. And the memory bandwidth issue. So, I need to see the benchmarks to know wether it is worth waiting for the M3 Pro Mac mini, or just getting an M2 Pro mini.

The M3 Pro appears to be slower in something like GFXBench but again it's old and dated. Or it could be the binned 14-core GPU version and the M2 Pro could be the full 18-core GPU.

 

[APCX]

The M3 Pro appears to have a small GPU speed improvement in something like GFXBench but again it's old and dated.

View attachment 2306324

To emphasise how bad gfxbench is, the M3 Pro scores lower than the M2 Pro!

 

[Pressure]

To emphasise how bad gfxbench is, the M3 Pro scores lower than the M2 Pro!

The A14 scores higher than the M3 in some of the offscreen tests 😂

 

[Populus]

There are already benchmarks from the M3 and M3 Max, but I’m really impatient to see the M3 Pro benchmarks, for a couple of reasons. First and foremost, for me it’s the sweet spot of raw performance at a still relatively affordable price, so it is my target device, wether the M2 Pro or the M3 Pro Mac mini. And, secondly, because the M3 Pro is the only chip that has reduced the performance cores in favor of the efficiency ones, and it has actually 3 billion transistors less than the M2 Pro. And the memory bandwidth issue. So, I need to see the benchmarks to know wether it is worth waiting for the M3 Pro Mac mini, or just getting an M2 Pro mini.

EDIT: I just realized that the M2 Pro comes with 6 performance cores and 4 efficiency cores (10 in total) not 8 p-cores and 4 e-cores like I thought. So overall, M3 Pro has indeed 2 more cores, even if those are efficiency cores.

EDIT 2: I see on Apple configuration website that there’s an option for a 12 core M2 Pro, for 350€.

 

[APCX]

The A14 scores higher than the M3 in some of the offscreen tests 😂

Yeah, it’s not great these days, but it’s one redeeming feature is a searchable database of results.

 

[Populus]

The M3 Pro appears to be slower in something like GFXBench but again it's old and dated. Or it could be the binned 14-core GPU version and the M2 Pro could be the full 18-core GPU.

View attachment 2306324

Thank you, although I’m more interested in the CPU.

 

[Populus]

Okay, comparing the GB6 scores of the M2 Pro and the M3, I can see the M3 single core (~3000) is higher than the M2 Pro single core (~2600), and the multi core score of the M3 (~11.700) is almost the same as the multi core score of the M2 Pro (~12.000). So it will probably be a good idea to wait for the M3/M3 Pro Mac mini.