hattig

LanceMeAlot said:

I have a mackbook air M1 for working purposes. For office, and online work it’s fantastic but as soon as you use MS teams for a meeting, sharing a screen and video conferencing, the cpu get’s too hot. And because there is no fan inside MacOS makes the cpu’s work slower to avoid heat problems. And then the machine will get very slow. Therefore I would always buy the 13” mackbook pro to avoid this problem.

MS Teams does that on 16 core behemoth desktop systems - it's not the system, it's the appalling software from MS.

blastdoor said:

If Geekbench is running through Rosetta then these numbers are actually pretty impressive. 800/1100 = 72% of full performance when running emulated code? That's really impressive. 

That's right.

Also consider that Geekbench is a benchmark - it's a pure x86-64 kernel running in isolation, so it's 100% worst case scenario.

Most applications will call out to the system to do even the simplest of things - which is pure native.

jdb8167 said:

I’m pretty sure Rosetta only uses the “big” cores which is why these results show up as a 4 core CPU and not 8. You also can’t rely on the reported clock speed. These results are from a kit that isn’t supposed to be benchmarked running an x86-64 translation to Aarch64. They shouldn’t be used to infer anything about native ARM performance. 

It’s also interesting that the date of the tests were from before WWDC which likely means they came from a developer with very early access. 

Rosetta is a translator, rather than a runtime emulation. Even the JIT feature will be implemented by translating finalised Intel Code into ARM (when you mark a memory page as executable post-JIT, Rosetta kicks in and does its work - I guess). **HOWEVER** there will be some things that need to be trapped as they are not emulateable, and therefore I suspect Rosetta provides a very thin VM/system layer around the executable. Also binary translators really cannot infer intent, so there is only so much they can do - they will generate code that is less optimal, especially when coming from x86 with all its quirks.

Yes, these are likely in-Apple or in-close-third-party leaks.

If you make your own silicon designs, then you have two costs - upfront research, design, testing, and then manufacturing.

The more you sell, the more you amortise the first cost over all of the sold devices. Apple is already investing the money into the first part for the iPhone/iPad SoCs - it is an incremental additional cost to do a new design for Macs. But let's say it's $500m for a 5nm design at this stage. Apple will ship 16m Macs a year. That's $31 per Mac. It'll be lower because a lot of that cost is shared with iPhone SoC. 

A 5nm wafer processing cost might be $12k (dropping as it matures) right now. It's 80% denser than 7nm. Let's assume they use that for the additional CPU cores and GPU cores the desktop chip would have and the resulting chip is 10.1mm x 12.6mm just like the A12Z. Even with poor yields you get over 300 fully working die per wafer out of ~450. That's a fabrication cost of $40 for each fully working die (and you'll get partially working dies as a bonus as well, for your lower-end products).

I imagine that Apple want to grow the Mac business and will use lower priced Macs with those recovered dies to achieve that - although there's a certain level of quality that Apple won't go below.

How much does Apple pay Intel for the chips in their laptops currently? Sure, it's discounted over list price, but it will be far, far higher than $71.

These numbers are extremely rough estimates.