M4 Mac Announcements

[Nycturne]

You'll have to explain that one to me. Cortex M-33's? Or would that be how many photos of Messier M33 you can process per hour? [You did mention getting into astrophotography....]

Messier 33, yes. It was this month's challenge for beginners over on Cloudy Nights, and I got a whopping 3 hours of time on it all month. But it takes about 23 minutes for my M1 Max to do a batch calibration + stack from the ~32GB data set that 3 hours produced. The work is CPU bound most of the time, and sometimes I/O bound, so I should get a rather noticeable jump.

If I had 10 hours of data, it easily would be longer than an hour to do the calibration + stack. I've got a guess on how much I can save here, but I'll wait until I have the numbers to share the guess and how off I am.

 

[tomO2013]

Question : have we established if Apple has caught up in the fish fingers and hungry hungry hippos benchmarks? - asking for “truth hurts” over at the other place!

it’s quite amazing to see folks arguing over how close the new M4 Max is the 4090 over at the other place.

Perhaps the big thing to consider…. What are the big incremental improvements that Apple has made, year on year…. At what power and performance…. In what form factors…

When I look at that, simply there is nothing close on the WinTel side.

I was talking with a server admin in our place today - he commented that the perf of the Mac mini Pro with 10G Ethernet makes it a very interesting for a compute dense rack option….

 

[Jimmyjames]

Question : have we established if Apple has caught up in the fish fingers and hungry hungry hippos benchmarks? - asking for “truth hurts” over at the other place!

it’s quite amazing to see folks arguing over how close the new M4 Max is the 4090 over at the other place.

Perhaps the big thing to consider…. What are the big incremental improvements that Apple has made, year on year…. At what power and performance…. In what form factors…

When I look at that, simply there is nothing close on the WinTel side.

I was talking with a server admin in our place today - he commented that the perf of the Mac mini Pro with 10G Ethernet makes it a very interesting for a compute dense rack option….

I don’t think there are any benchmarks out there other than the Russian ones from a couple of weeks ago (Geekbench ~3800 single core, 15000 multi-core). We should start getting them soon. If “truth hurts” is mic7hy, then it won’t matter in any case as they will just obfuscate and lie. As they did with @leman when they compared cycles render within Blender on macOS to the eevee one on Lunar Lake.

I’m almost reluctant to check it out in order to protect my mental health!

 

[Jimmyjames]

Speaking of which…first GB scores for M4 Pro.

Mac16,7 - Geekbench

Benchmark results for a Mac16,7 with an Apple M4 Pro processor.

browser.geekbench.com

 

[Cmaier]

Speaking of which…first GB scores for M4 Pro.

Mac16,7 - Geekbench

Benchmark results for a Mac16,7 with an Apple M4 Pro processor.

browser.geekbench.com

View attachment 32396

that’s around 10% higher than iPad Pro M4 on single core?

 

[Jimmyjames]

that’s around 10% higher than iPad Pro M4 on single core?

4.51ghz

 

[Jimmyjames]

M4 Pro OpenCL. Nice 30-35% increase!

Mac16,7 - Geekbench

Benchmark results for a Mac16,7 with an Apple M4 Pro processor.

browser.geekbench.com

 

[theorist9]

Speaking of which…first GB scores for M4 Pro.

Mac16,7 - Geekbench

Benchmark results for a Mac16,7 with an Apple M4 Pro processor.

browser.geekbench.com

View attachment 32396

There are eight leaked base M4 Mac scores on GB, and they all say this:

...which is the same clock as on the M4 iPad:

While the one you linked says this which, if real, means there's a 100 MHz clock bump on the Pro:

I'm imagining the marking guys begging the engineers to let the Max run even faster so they can break that 4000 point barrier on GB....

Also: Does the 16,7 model identifer mean Mini, 14" MBP, or 16" MBP?

 

[Jimmyjames]

There are eight leaked base M4 MBP scores on GB, and they all say this:

View attachment 32398
...which is the same clock as on the M4 iPad:
View attachment 32400


While the one you link says this which, if real, means there's a 100 MHz clock bump on the Pro:

View attachment 32399

Those are old ones from the Russian leakers. There are a few from the recents list that just showed up. The one I posted is the fastest. I suspect it may be the Mini?

 

[theorist9]

Those are old ones from the Russian leakers. There are a few from the recents list that just showed up. The one I posted is the fastest. I suspect it may be the Mini?

My eight included both the older Russian base M4 scores and the most recent base M4 on Oct 30, and they all showed the same 4.41 GHz clock.

 

[Jimmyjames]

My eight included both the older Russian ones and the most recent one on Oct 30, and they all showed the same 4.41 GHz clock for the base M4.

I think they may have been re-tested as iirc they are all Mac16,1?

 

[Jimmyjames]

Nice M4 Pro Metal score +35% roughly?

Mac16,11 - Geekbench

Benchmark results for a Mac16,11 with an Apple M4 Pro processor.

browser.geekbench.com

 

[theorist9]

I think they may have been re-tested as iirc they are all Mac16,1?

Sounds like you're saying all those may be been the same machines tested earlier. Or even just multiple tests of a single machine. So our sample size of base M4's for assessing clock has n=1. That's certainly possible!

 

[Jimmyjames]

Sounds like you're saying all those may be been the same machines tested earlier. Or even just multiple tests of a single machine. So our sample size of base M4's for assessing clock has n=1. That's certainly possible!

Yes that’s it. To be clear. it’s a total guess!

 

[Jimmyjames]

Geekbench AI GPU scores see a nice FP32/16 increase of around 30/35% which matches what were seeing generally in the gpu.

Mac16,7 - Geekbench

Benchmark results for a Mac16,7 with an Apple M4 Pro processor.

browser.geekbench.com

GB AI CPU also has nice increases. +30-35% for FP32/16 approx.

Mac16,7 - Geekbench

Benchmark results for a Mac16,7 with an Apple M4 Pro processor.

browser.geekbench.com

 

[exoticspice1]

i feel sorry for the M3 Pro folks.

 

[Jimmyjames]

i feel sorry for the M3 Pro folks.

I don’t. I have a 2018 mini!

 

[theorist9]

Nice M4 Pro Metal score +35% roughly?

Mac16,11 - Geekbench

Benchmark results for a Mac16,11 with an Apple M4 Pro processor.

browser.geekbench.com

View attachment 32401

I'm calculating a 50% increase over the Metal Score listed on Primate's website for the M3 Pro: 111,119/73,810 = 1.505

And the same for the Open CL score: 69,867/46,656 = 1.497

I don't know how well that will scale to the M4 Max, but if we use the 1.84 x scaling of Open CL scores between the M3 Max and M3 Pro, that would give ≈130k, which would put it about half-way between a 4070 laptop GPU (107k) and a 4080 laptop GPU (161k). That would be much better comparative GPU peformance than that of the M3 Max which, at 86k, had about the same Open CL score as a 3060 laptop GPU (84k).

If we scale by 1.84^2 to project an M4 Ultra's Open CL score, that's 237k, about the same as the 238k score of the 4080 desktop. Though by the M4 Ultra's release date, the 5000 series should be out.

 

[Jimmyjames]

I'm calculating a 50% increase over the Metal Score listed on Primate's website for the M3 Pro: 111,119/73,810 = 1.505
View attachment 32410

You might be correct. I grabbed the first one I saw!

 

[theorist9]

You might be correct. I grabbed the first one I saw!

I just used the values you posted. I can't seem to find the results myself—my searches for the M4 Pro on their website come up blank.