[wintercharm]

Big takeaways:

- A12 has 8MB on chip “SoC cache”

- Big core L1$ = 128kB; Little Core L1$ = 32kB

- For the big core, L2$ is a whopping 128 instances 6MB per core/thread, 8MB at 64KB/inst.

- Little core L2$ is 32 instances, 1.5MB per core/thread, 2MB at 64KB/inst

- A12 GPU uses memory compression!

- A12 Big has 2.38 GHz base clock and 2.5 GHz 1 core boost

- A12 Little has 1.538 GHz all core, and 1.562 GHz 2 or 3 core boost, and 1.587Ghz 1 core boost

- A11 and A12 have a 7-wide decode (up from 6 on the A10) and 6Int ALU (up from 4 on the A10)

- Apple’s microarchitecture seems to far surpass anything else in terms of width, including desktop CPUs

- SPECint/fp Numbers show that it’s got 2x the speed of and any all other mobile SoC’s. 3x perf/watt if you normalize speed / power consumption.

- SPECint/fp numbers also show that the A12 is faster than a skylake server cpu (in core-for-core IPC). Not a perfect comparison, but far better than Geekbench

[ksec]

The A12 CPU Core is nearly if not exactly the same as A11. Only difference is clock speed, cache and node improvement. Think of A12 as a optimisation / Tock improvement. ( Anandtech didn't do deep dive on A11 last year due to shortage of staff )

So basically Apple's A11 has already reached Skylake like IPC since last year.

Which makes me wonder why Apple is still paying Intel substantial amount of premium for x86 on Mac? There is at least $100 BOM cost saving, translate to $200 in RSP. In an ideal world, iPad would have had 8 years of iteration and eating away PC market shares. It seems the vision of iPad died along with Steve Jobs 7 years ago today.

Steve, we miss you.

[schuke]

The iPad has so much potential yet the best effort they gave it was an average stylus and a clumsy keyboard cover. Now that the Surface has not only caught up but developed its unique forte the iPad has even less opportunity to innovate and lead.

[tgb]

What does "in terms of width" mean in this context?

[duhast]

width in this context means number of CPU execution ports or how many INT/FP operations CPU core can perform at the same time

[jakelarkin]

probably the processor-memory bus. GPUs have wider ones