dave haynie

jdiamond said:

The biggest apparent technical issue seems to be the failure of the quad-die Apple Silicon chip.  Perhaps they could make up for this with a dual socket system?  

While there were rumors, I don't think Apple ever had a four die system. The two die produces are two die, edge to edge on a multi-chip module (MCM), with a huge and ridiculously fast "UltraFusion" interconnect that make them essentially one chip. So far, there is only one such interconnect per chip. AMD does their "chiplet" thing with multiple on chip links on an interconnect called Infinity Fabric, much more flexible but not as fast as Apple's. To make a four chip solution using what Apple has today, you'd need either two Ultrafusion interconnects per chip or a central four!]-way switch. 

Some more conventional "dual socket" architecture would have two SoCs connected via... PCIe? Some other interconnect? That leaves very few PCie lanes for anything else today, and it's no longer a unified architecture, but some kind of cluster. Very different. 

jdiamond said:

It'd also be great if they found a way to do a two tiered memory system, so you could add terabytes of DRAM to the embedded system.  It sounds difficult, but Apple already pulled something similar off with the old two tiered SSD/HDD systems.  You'd let the 192GB of on-die RAM page out to the 8 TB of DIMMs.

The DRAM isn't on-die, but it's on the MCM. It's conventional DDR5, but tighter timing on the MCM lets it run faster. They could support another two or four DDR5 channels for off-module DRAM, but tjat is a ton of extra signals to deal with. 

Also, you're not using off-module DRAM as paged memory, it's not all much slower, but the overhead would be killer. Rather, the OS would have to manage tiers of memory, like most NUMA architrctures impose. All critical OS memory gets allocated in on-module RAM. 

techconc said:

No matter how you spin it, the M2 Mac Pro is a real disappointment.  Full stop. 

The only excuse I’d give Apple for this disappointment is if they felt it were more important to formally complete the Apple Silicon transition than it was to provide a proper pro machine.  So, if this is a stop gap measure to hold us until this time next year, then fine.

The Mac Pro is meant to be the flagship device… the pinnacle of Mac performance.  Instead, it’s a Mac Studio with PCI slots.  At the very least, an M2 Extreme (2 M2 Ultra chips) is what users are expecting.  Apple seems content on comparing to a 4 year old Intel Mac Pro while ignoring the current Intel / nVidia 4090 based solutions.   That would address the CPU / GPU scalability concerns or at least help mute them.
The other concern is memory.  192 GB is fine for most solutions, but there are very high end needs which go well beyond that.

Yeah, I agree. Apple has kind of set their own trap at the high end. One must conclude that reallynhihh performance just does not scale when you're required to make a new gigantic chip or two, rather than reusing one, when the market is that small already. 

The problem with the idea of an M2 Extreme is that there is no M2 Ultra chip. That's two M2 Max chips connected end to end. Unlike most of the other chip to chip interconnects like AMD's infinity fabric or Intel's EMIB that's any reasonable kind of compact or routeable, just the eonergy to end connection. So Apple would, at best, need an extra chip for this with a 4-way port, and of course, support in their UltraFusion connection. Or potentially, more than one UltraFusion interconnect per chip. 

As well, their tight coupling of DRAM to CPU necessitates DRAM on an MCM and maybe lowers main board cost in most of the M-series computers. But it also precludes memory expansion. 

In the context of a powerful small desktop machine, the Studio makes perfect sense. But the power of a large workstation class tower PC is all about the expansion. It's not just that the Mac Pro can't compete with a nVidia RTX4090... it's that it can't compete with two, three, or four of them. 

What is the basis for the A16 cost estimate. It's barely an upgrade, really should have been called the A15+, on,y a small increase in transistor count, pretty much the same CPU, GPU, and NPU cores, slightly faster clocks. The TSMC 4nm process is really a tweaked 5nm process, too. Double the price for 6% more transistors and a 5% clock speed boost?