Well it’s happening. Unfortunately we are still mostly in the dark regarding actual products and configurations, and judging by Apples modus operandi, we are likely to remain so until they are ready to take orders.
However.
Apple has made quite a few statements, and I’ll refer to them and use them as a springboard for some speculation.
In the
keynote, Johny Srouji made these statements:
1.27.48-1.30.48 (A whole new level of performance)1.31.23 (no blue field at ”desktop” power levels) (much higher level of performance while consuming less power)(A whole new level of graphics performance to the mac)(family of SoCs specifically for the mac product line)-1.33.16
In the developer State of the union video
https://developer.apple.com/videos/play/wwdc2020/102/
Sri Santhanam
08.30-09.30(Building a family of SoCs tailored specifically to the Mac. Practically speaking a whole new level of performance.)11.00 (Our goal is to provide maximum performance within each of (Mac) enclosures.)-12.00 (Bringing our high performance GPU architecture to the Mac)-13.00 13.15 (unified memory architecture) 14.25 (every mac will have powerful graphics)
And Craig Federighi in a
video interview 48.30 about the dev kit.
“It gives a sense of what our silicon team can do when they are not even trying. And they’re gonna be trying.”
So - family of SoCs tailored to the power/thermal envelope of their respective product lines. Confident about the performance being much improved. Referring to unified memory and graphics power being strong across the board.
My speculation is that this implies that they won’t use a standard PC memory solution. Todays multicore CPUs pull data from narrow memory subsystems. And the bandwidth problem gets worse if you try to run graphics code at the same time.
If they are going to offer substantially better performance other than in very niche benchmarks, they need to do something about the memory subsystem. The iPad solution has been going to a 128-bit wide memory path, to LPDDR4x in the 2018 model, and likely LPDDR5 in the next.
But while that is comfortably better than what will be the going standard in x86 space, it doesn’t scale up nicely for systems that go beyond 10W power draw. I see two industry standard possibilities. One is GDDR6 a la consoles, the other is HBM.
Both offer bandwidth that allow graphics on a SoC to scale to high levels. (Next iPad Pro is likely to be roughly at PS4 level or thereabouts, so lets take that as a base line for fanless Macs.)
GDDR6 is relatively cheaper than HBM, but Apples volumes may reduce the difference, so HBM may be a more scaleable approach.
Of course, this means you can’t upgrade your memory capacity after purchase, but that is largely the case already for their product line, and the benefit is huge for performant unified memory systems.
Benchmarks, as always, will downplay the benefits of a strong main memory systems, so tech articles will likely describe it as mainly benefiting graphics (where it
would make a huge difference) but all code that shuffle data to feed the SoC would gain.
(Also of note is that typical cross platform benchmarking like SPEC or GB will miss a large point of the change, which is the integration of dedicated coprocessors for various tasks.)
If the new Apple silicon Macs ship with a strong unified memory subsystem, that will in itself be a huge competitive advantage over what is currently projected in Win-x86 space.
I’m leaning towards HBM in stationary Macs, and either that or a 256-bit interface to LPDDR5 in MacBook Pros. We’ll see in a year or so.