Apple Reveals M6 as First-Ever 2nm Chip

[Populus]

I guess we'll have to wait until someone (I miss AnandTech) analyzes the new M6 cores, and compare the M5 super-cores against the M6 performance-cores, and the super-cores. But I have a hunch there's little difference in the core architecture beyond the node shrink, and the true new core architecture is going to debut with the M7.

Also. They say the M6 super cores are aimed at single threaded tasks but... the M5 has four of them... I wonder if the 4 cores in the M5 are also aimed at single threaded tasks or multi thread tasks.

But from a raw power point of view, a chip with 4 super cores (M5) should be more powerful than a chip with 4 performance -middle- cores and 2 super cores (M6), right?

I guess we'll have to wait until the first real world reviews and benchmarks.

 

[chucker23n1]

But I have a hunch there's little difference in the core architecture beyond the node shrink, and the true new core architecture is going to debut with the M7.

Possibly. The node shrink could allow an even higher clock (M5 ran at 4.61 GHz), although they can't keep that up for many more generations.

With Geekbench 6, the single-core generation-over-generation improvements, adjusted for clock speed increases, looked as follows:

So in the ARM Mac era, the M2 and M3 didn't actually increase IPC much at all. But M4 certainly did, and M5 a little bit as well.

But interestingly, Geekbench 7 seems to penalize this even further:

I haven't yet read into what Geekbench 7 does differently.

Either way, I can imagine Apple wanting to do big IPC jumps every three generations. We see them with the M1 and M4, and maybe M7 will indeed be the next time.

But from a raw power point of view, a chip with 4 super cores (M5) should be more powerful than a chip with 4 performance -middle- cores and 2 super cores (M6), right?

We don't know enough about either the M5 or M6 cores to answer that, I would say.

 

[darkblu]

Possibly. The node shrink could allow an even higher clock (M5 ran at 4.61 GHz), although they can't keep that up for many more generations.

With Geekbench 6, the single-core generation-over-generation improvements, adjusted for clock speed increases, looked as follows:

View attachment 2656572

So in the ARM Mac era, the M2 and M3 didn't actually increase IPC much at all. But M4 certainly did, and M5 a little bit as well.

But interestingly, Geekbench 7 seems to penalize this even further:

View attachment 2656573

I haven't yet read into what Geekbench 7 does differently.

Either way, I can imagine Apple wanting to do big IPC jumps every three generations. We see them with the M1 and M4, and maybe M7 will indeed be the next time.



We don't know enough about either the M5 or M6 cores to answer that, I would say.

Are these charts taking into account the SME effect in m4 and later? Because it they do not that score hike by m4 could be a bit.. artificial (pun intended)

 

[BorisDG]

I guess we'll have to wait until someone (I miss AnandTech) analyzes the new M6 cores, and compare the M5 super-cores against the M6 performance-cores, and the super-cores. But I have a hunch there's little difference in the core architecture beyond the node shrink, and the true new core architecture is going to debut with the M7.

Also. They say the M6 super cores are aimed at single threaded tasks but... the M5 has four of them... I wonder if the 4 cores in the M5 are also aimed at single threaded tasks or multi thread tasks.

But from a raw power point of view, a chip with 4 super cores (M5) should be more powerful than a chip with 4 performance -middle- cores and 2 super cores (M6), right?

I guess we'll have to wait until the first real world reviews and benchmarks.

They are still 10-wide just like M5. It seems not big architectural change other than the node and changing some core configurations - removing S for P cores. (They are 7-wide vs 6-wide for the E). I think M6 is mostly focused on AI ie the 2x NPU (Dual core)

Also it seems, it's still running on AMX 6 and it has same cache configuration. Every previous generation had an upgrade.

 

[chucker23n1]

Are these charts taking into account the SME effect in m4 and later? Because it they do not that score hike by m4 could be a bit.. artificial (pun intended)

I've never understood the logic of this concern. Obviously a CPU that supports a certain extension is going to use that extension in benchmarks, as it would also use it in real-world use.

 

[BorisDG]

Are these charts taking into account the SME effect in m4 and later? Because it they do not that score hike by m4 could be a bit.. artificial (pun intended)

M4 was CPU year (A18 Pro) compared to the slow M3 (GPU year - 1st with Ray-tracing via A17 Pro) which was Frankenstein for cores arch. (A17 Pro + A16 Bionic). M5 was GPU year (A19 Pro), but also it introduced the Super cores with more cache. M6 should use the same Apple 10 GPU (2 years always), so it has to be CPU year, but than (it's always juggle between the two) it's still 10-wide, which was the big jump in M4. So no idea. It feels like weird chip, but we have to see. I'm super curious!

 

[darkblu]

I've never understood the logic of this concern. Obviously a CPU that supports a certain extension is going to use that extension in benchmarks, as it would also use it in real-world use.

I take this as a ‘no’ : ]

It skews the result toward a single type of gb workload which was nerfed before. And it’s not like the AMX unit was introduced with the isa extension — the unit was there already, just gb had no way of using it. A benchmark that was already using the unit would’ve shown completely different results in your chart.

 

[darkblu]

M4 was CPU year (A18 Pro) compared to the slow M3 (GPU year - 1st with Ray-tracing via A17 Pro) which was Frankenstein for cores arch. (A17 Pro + A16 Bionic). M5 was GPU year (A19 Pro), but also it introduced the Super cores with more cache. M6 should use the same Apple 10 GPU (2 years always), so it has to be CPU year, but than (it's always juggle between the two) it's still 10-wide, which was the big jump in M4. So no idea. It feels like weird chip, but we have to see. I'm super curious!

Oh I know the m4 was a notable improvement cpu-wise. Just not to the degree suggested by SME benchmarks.

 

[chucker23n1]

It skews the result toward a single type of gb workload which was nerfed before.

Yes, because the newer CPU is better at the workload than its predecessor, which is the entire point of a benchmark.

 

[DBR2025]

Time to sell the iPad Pro m5 and get a new shiny iPad Pro m6.

Damn son - show me the money.