May 7 “Let Loose” Event - new iPads

[Cmaier]

iPad16,6 - Geekbench

Benchmark results for an iPad16,6 with an ARM processor.

browser.geekbench.com

Amazing how much performance has increased since M1. People whine when apple doesn’t hit 20% improvement in a year, and then they come along and update twice in a year and make up for it. And you get base M4 multicore scores that look like M2 Pro multicore scores.

And M4 base is like 15% faster in multicore than M1 Max.

 

[Jimmyjames]

Amazing how much performance has increased since M1. People whine when apple doesn’t hit 20% improvement in a year, and then they come along and update twice in a year and make up for it. And you get base M4 multicore scores that look like M2 Pro multicore scores.

It’s incredible really. The 2019 Mac Pro doesn’t feel that long ago and now an iPad is 2.5 times faster in single core and nearly 50% faster multi-core. I know the Mac Pro can sustain performance and offers expansion etc, but in terms of cpu, it’s crazy.

 

[Jimmyjames]

Well it’s started.

 

[Jordanmarkstone]

Well it’s started.

View attachment 29322

Well it’s started.

View attachment 29322

Amazing.

 

[Cmaier]

Well it’s started.

LOL.

 

[Jimmyjames]

Amazing.

Incredible.

 

[dada_dave]

Well it’s started.

View attachment 29322

I know I’m over thinking here but why is object detection removed, just cause it’s an outlier?, and did they explain how they got IPC w/o SME?

GB Compute score = 53792

iPad16,6 - Geekbench

Benchmark results for an iPad16,6 with an ARM processor.

browser.geekbench.com

Very nice indeed. And iOS 17.5 which makes more sense.

 

[exoticspice1]

gb6.3 added support for sme

 

[Yoused]

Well it’s started

Wait, does that say what I think it says? What does "M3M" mean?

I looked at the base M3->M4 comparison, and just IPC has improved a lot. Like, nearly 15%, based on SC ÷ GHz.

 

[Jimmyjames]

Wait, does that say what I think it says? What does "M3M" mean?

I assumed M3 Max. None of that post makes sense to me. We have next to no knowledge about the M4, and they have dissected one score to determine it’s shit.

I looked at the base M3->M4 comparison, and just IPC has improved a lot. Like, nearly 15%, based on SC ÷ GHz.

That’s my impression.

 

[Cmaier]

Wait, does that say what I think it says? What does "M3M" mean?

I looked at the base M3->M4 comparison, and just IPC has improved a lot. Like, nearly 15%, based on SC ÷ GHz.

you aren’t applying the “apple sucks” filter.

 

[Cmaier]

I assumed M3 Max.

So he’s also assuming that a MBP or Ultra is going to run at the same clock as an ipad?

 

[dada_dave]

Well it’s started.

View attachment 29322

I know I’m over thinking here but why is object detection removed, just cause it’s an outlier?, and did they explain how they got IPC w/o SME?



Very nice indeed. And iOS 17.5 which makes more sense.

Wait, does that say what I think it says? What does "M3M" mean?

I looked at the base M3->M4 comparison, and just IPC has improved a lot. Like, nearly 15%, based on SC ÷ GHz.

Oh I get it now, he removed object detection because he's saying that increase is due to SME alone (and because that doesn't count for some reason) and that the rest of the scores only go up by the same as the clock speed increase.

So he’s also assuming that a MBP or Ultra is going to run at the same clock as an ipad?

I suppose if his claim is that IPC hasn't increased then any increase in frequency doesn't help with that. Of course IPC is a function of clockspeed so my guess is that if the M4 was run at the M3's clocks, IPC will indeed have increased, it's just that Apple seems to also increase clocks to the point where IPC remains the same. The real question is PPW, if Apple can run the new chip at 4.4 GHz, get that performance, and keep power low then that's a win, just as M3 was over M2. The IPC thing seems to be a metric for how much the micro-architecture has advanced, but I'm not convinced it's always telling the whole story in that regards. Maybe it is, but I have my doubts.

 

[theorist9]

Well it’s started.

View attachment 29322

If the test had been done apples-to-apples, then a reasonable non-clickbait headline would be: "Most of the IPC uplift in the GB6 CPU test suite is from Object Detection."

But you really can't say that from what this poster has presented. There's no indication these were tested under identical conditions--test variation itself can be well over 10%. And the fact that Horizon Detection shows (in this comparison) an IPC reduction in the M4 suggests they were not (based on the relative clocks, zero IPC uplift would give you a 6% increase).

 

[Yoused]

So he’s also assuming that a MBP or Ultra is going to run at the same clock as an ipad?

Do you remember a couple decades ago when the first GHz processors started showing up? And then in a couple years, we had a few 2GHz processors, and some people were saying that in just a few years we would have 10GHz machines?

Where are we now.

 

[Cmaier]

Do you remember a couple decades ago when the first GHz processors started showing up? And then in a couple years, we had a few 2GHz processors, and some people were saying that in just a few years we would have 10GHz machines?

Where are we now.

I was in the race for the first GHz processor! And i won the race twice! (Once at RPI with F-RISC/G, and once at AMD with Athlon). (At least I think so? Maybe somebody beat those in their respective categories but I can’t recall anyone).

But physics stopped the race from continuing full speed ahead. Power consumption, of course. And transistor toggle speed began to get overwhelmed by interconnect parasitics, which haven’t scaled the same as transistor channel size.

 

[dada_dave]

If the test had been done apples-to-apples, then a reasonable non-clickbait headline would be: "Most of the IPC uplift in the GB6 CPU test suite is from Object Detection."

But you really can't say that from what this poster has presented. There's no indication these were tested under identical conditions--test variation itself can be well over 10%. And the fact that Horizon Detection shows (in this comparison) an IPC reduction in the M4 suggests they were not (based on the relative clocks, zero IPC uplift would give you a 6% increase).

Indeed.

It would appear that non-SME IPC (because again we're not being allowed to count that, very important ) though could be within a few %, like the M3 with M2. Also within each subsection it's supposed to a geometric mean and then the final score is supposed to be a weighted average of the two subsections. Without knowing why GB 6 is supposed to be 65-35% Integer-FP, it becomes hard to know how those weights are supposed to change if you simply remove a test. Maybe they stay the same ... maybe not.

https://www.geekbench.com/doc/geekbench6-benchmark-internals.pdf

 

[mr_roboto]

It would appear that non-SME IPC (because again we're not being allowed to count that, very important )

I mean attributing this to SME at all is extraordinarily dubious. We don't even know whether M4 supports it, and as I just pointed out to someone over at the other place, there's no reason to believe that Primate Labs shipped any SME code in GB6. To the contrary, there's lots of reasons to expect that they haven't, starting with the essentially total lack of any hardware on the market which supports SME. It's a very recent extension, even Arm's own cores don't generally support it yet.

 

[dada_dave]

I mean attributing this to SME at all is extraordinarily dubious. We don't even know whether M4 supports it, and as I just pointed out to someone over at the other place, there's no reason to believe that Primate Labs shipped any SME code in GB6. To the contrary, there's lots of reasons to expect that they haven't, starting with the essentially total lack of any hardware on the market which supports SME. It's a very recent extension, even Arm's own cores don't generally support it yet.

According to GB release notes they did in the latest update:

Geekbench 6 Release Notes

But you're right we don't know what capacity or what's affected or how or even if M4 truly has it. Even so, a huge percentage of the increase in performance is in Object detection. Something has changed about it far more than the others. Attributing that to SME is definitely the most plausible explanation though not definitely the definite explanation.

 

[exoticspice1]

Here is GB5. It seems mostly clock improvements and a bit of IPC.

iPad16,6 - Geekbench

Benchmark results for an iPad16,6 with an ARM processor.

browser.geekbench.com