View Single Post
chucker
 
Join Date: May 2004
Location: near Bremen, Germany
Send a message via ICQ to chucker Send a message via AIM to chucker Send a message via MSN to chucker Send a message via Yahoo to chucker Send a message via Skype™ to chucker 
2021-11-05, 13:20

Quote:
Originally Posted by psmith2.0 View Post
Interesting story out on the AS front.

The future sounds bright.



I suppose "third generation" is M3? Or is the new M1 Pro/Max considered the second generation and what's coming next (M2 in 2022?) considered third-generation?
Given it says 40 cores, I think the answer is neither: this is an M1 Extreme / M1 Ultimate (or whatever stupid branding suffixes they come up with).

So far, we have:
  • the M1, which contains 4 Firestorm and 4 Icestorm cores, making it a bit of an A14X
  • the M1 Max, with 8 Firestorm and 2 Icestorm cores, up to 32 GPU cores, and four LPDDR5 RAM chips
  • the M1 Pro, with 6 or 8 Firestorm and 2 Icestorm cores, up to 16 GPU cores, and two LPDDR5 RAM chips

I bring this up (and in this particular order!) because there was a rumor around April that a chip design called "Jade C-Die" was going to arrive in late 2021, and that, derived from that, we'd see a "Jade C-Chop", and a "Jade 2C-Die" and "Jade 4C-Die". Based on that, John Siracusa made this mockup.

What actually ended up happening is surprisingly close to that rumor. If we consider the M1 Max the "Jade C-Die", then the M1 Pro is the "Chop" version, with portions chopped off. In Siracusa's mockup, basically the right third. And in the actual M1 Pro and Max? The bottom third.

https://images.anandtech.com/doci/17019/M1PRO.jpg vs. https://images.anandtech.com/doci/17019/M1MAX.jpg

The M1 Pro is basically an M1 Max with half the LPDDR5 chips, half the SLC blocks, and half the GPU cores chopped off. So that part of the rumor seems to have been right on the money. Extrapolating from that, suppose the "Jade 2C/4C-Die" part is, too. If you look at Siracusa's diagram again, that basically means you take the M1 Max, and add another whole M1 Max underneath it (2C), and then you add a whole block of those two on the side (4C, a 2x2 grid).

And that, the supposition goes, is what they'll do for the high-end desktop Macs (space grey Mac mini? 30-inch iMac Pro? Mac Pro? All three of those?).

The downside would be: even on this very high end, you'd be constrained in your RAM choices by the very specific amount you got when you ordered it. Plus, it won't be a lot. The current Mac Pro goes up to 1.5 TB RAM (yes, that's 1536 Gigabytes). This setup would "only" go to 256 GB.

On the other hand, it would lead to astounding performance levels, assuming your code (just as before with the iMac Pro and Mac Pro) can take advantage of heavy parallelization. The M1 Extreme would have a memory bandwidth of 800 GB/s, and the Ultimate of 1,600 GB/s, simply because what they'd do is double the memory chips again.

And for the same reason, such a Mac with an M1 Extreme would have up to 20 CPU cores and up to 64 GPU cores, and a Mac with an M1 Ultimate with the cited up to 40 CPU cores and up to 128 GPU cores. It's simply four M1s Max in a 2x2 grid.

Quote:
Originally Posted by psmith2.0 View Post
I assumed the "generation" count would be designated by the M1, M2, M3, etc. labeling, but I'm the first to admit that I don't understand anything about this stuff beyond the lightest surface.
No, I think you understand just fine, it's just that the choice of word "generation" is unfortunate here. "Third wave of releases" would be more apt.

There was the A14 and soon after the M1, then a year later, there were the M1 Pro and M1 Max, and now, some time next year, we might see the M1 Extreme and M1 Ultimate. (I don't think Apple would actually use "Ultimate", but I also can't think of a less stupid adjective. "M1 Blowaway"? "M1 Offthecharts"?)

Quote:
Originally Posted by psmith2.0 View Post
I just assumed any new Mac Pro coming in 2022 would be on an M2 Pro/Max, but maybe that's not how it works? The M1, and all its variations, could be in the picture for some time and it won't be a yearly M1, M2, M3 type of thing at all.
Yes and no!

It seems what Apple is doing here is a bit similar to Intel. They introduce a bunch of CPU cores like the A14's Firestorm (on the "performance cores") and Icestorm (on the "efficiency cores"). Then they move that up to the M1. Then further up to the M1 Pro/Max a year later. And then all the way up to the Extreme/Ultimate another year down the road. But, in the meantime, they've also introduced another generation (here, it fits) of CPU cores, namely the A15's Avalanche and Blizzard.

So far, we have not seen a Mac with those cores. My guess is we'll see an M2 MacBook Air (and perhaps Mac mini, and perhaps even 24-inch iMac) in spring/summer of 2022, and that the M2 will use Avalanche/Blizzard. That is, each of the cores inside will be a little faster and better™ than the M1's. In either late 2022 or some time 2023, then, we will see the M2 Pro/Max, replacing the M1 Pro/Max that just came out.

Apple's calculation being that pros need a lot of cores (and special features like ProRes acceleration) more than they need the most current cores.

This is similar to how Intel rolls out a microarchitecture like Skylake — on the desktop in 2015, on the laptop in 2016, and on workstations/servers in 2017. But meanwhile, by 2017, desktops and laptops already get newer microarchitectures.
  quote