Doug S (foo.delete@this.bar.bar) on November 21, 2020 8:53 am wrote:
>
> Why can't "adapt it as best they can" also mean "beat Intel and AMD performance at the high end"? They've
> got this 8 big / 4 little chip coming up. If they were building the Mac Pro in 2021 instead of 2022, they
> could use four of those working together as chiplets and get a 32 core system. AMD seems to have had no problem
> making chiplets work well, despite having embarked on that path when they had few engineering resources compared
> to what Apple (or Intel) can muster. The next generation of that 8+4 chip built on TSMC N3 in 2022 would
> make a killer Mac Pro (two chiplets for the "entry level" Mac Pro and four for the beast)

In the past the difference was that workstation and server applications could afford lower yields and higher power making greater integration on larger die viable.

> If they're going to scale their own GPU all the way to Mac Pro as evidence suggests, having all
> those GPU cores on the chip isn't a waste as it would be if they were planning on having the Mac
> Pro ship with a third party GPU in a PCIe slot. After all AMD (and Intel) are developing chiplet
> based GPUs, so there's no reason Apple can't do that too - and how knows how long they've been working
> on this already in the bowels of the spaceship. You need more memory channels in a bigger system,
> well you get more for every chiplet. Same with I/O and number of displays supported.
>
> So it isn't clear to me at all that "adapt it as best they can for
> their desktops" isn't also "provide a market leading solution".

With chiplets maybe it will work out that way this time but AMD went the chiplet route for lack of an alternative and chiplets are not new to Intel who used them when they were either desperate or for packing density in portable applications. Intel is not failing now for lack of using chiplets.

What are AMD and nVidia uses for their largest GPUs?