No “Extreme” chip coming to Mac Pro?

Cmaier

Site Master
Staff Member
Site Donor
Posts
5,666
Reaction score
9,312
Maybe, but they probably still want to keep single-chip solutions for laptops. I can imagine that the die connector comes with a power usage overhead.

They already distinguish the ”base” M1/2 chip from the “max” so I would expect them to continue that. But I could see this other solution for anything Max (or maybe Pro) and higher.

The die connectors shouldn’t add much power at all as long as they are passive. Looking at the M2 Max die photos, there are already some long communication paths, and only a small portion of each of those wires would have somewhat higher capacitance than now.
 

Yoused

up
Posts
5,987
Reaction score
9,721
Location
knee deep in the road apples of the 4 horsemen
it feels to me that the apple-y solution is to partition GPUs on one chiplet and CPUs on another. That would give them flexibility to scale core counts more appropriately for the specific product, instead of being limited to multiplying everything by powers of 2.

I have this insane belief that Apple will move to a "UPM" (unified processing module), skipping past everyone who said "they will eventually have to hyperthread". There will be a sectored rename file with a sort of spoke-like EU structure radiating out through it, where rename blocks and EU lines can be gated out during low-load operation. The rim will have fetch/decode/dispatch/capture/retire units with index blocks for register identification for each code stream, and all ops will just flow through as they come in. There might be one separate small core for handling certain privileged processes, just to be safe.

There are two reasons for this. The big one is that they can tape out one design that works across all products and fuse off capacity for the low end models. The Air will only have 8 or 10 thread capability while the high end will reach maybe three to four times that, all on the same UPM. The second will be about implementing 4x SVE/SVE2, simply by using long rename lines (which, frankly, they could do now, but the UPM would allow easy SVE access to any code stream without having to work it into low-power cores where it might be a tight fit).

Having a single processor design for all their SoC models would reduce costs a lot. They might be able to go with a simple fusion connect for the GPU, making it much easier to produce a range of products just by gluing two chips together.
 
Top Bottom
1 2