M1 Pro/Max SoC

[altaic]

Just having a look at the die shots from the event. MR captured them both:

I guess we now know why the internal name was Jade and Jade Chop. The M1 Pro looks like the M1 Max with the die snapped off right in the middle of the massive GPU.

Also, that GPU...! It's like 1/3 of the die! The M1 Max is more like a GPU with some other stuff added in than a CPU with a GPU tacked on.

 

[Gnattu]

The M1 Max is more like a GPU with some other stuff added in than a CPU with a GPU tacked on.

Just like the SoC made by AMD for PS5 and Xbox Series X.

 

[Gnattu]

Even better die shots side by side. It is so clear that the M1 Pro is a cut down M1 Max.

 

[Macintosh IIcx]

Just like the SoC made by AMD for PS5 and Xbox Series X.

Yep, and unified memory architecture. Interesting how Apple has gone the way of the game console approach to a certain extent when you also think of the tight OS integration. (Yes, yes I'm aware of the many dissimilarities also, but you get it I'm sure)

 

[Boil]

Just having a look at the die shots from the event. MR captured them both:

I guess we now know why the internal name was Jade and Jade Chop. The M1 Pro looks like the M1 Max with the die snapped off right in the middle of the massive GPU.

Also, that GPU...! It's like 1/3 of the die! The M1 Max is more like a GPU with some other stuff added in than a CPU with a GPU tacked on.

View attachment 1870174

Even better die shots side by side. It is so clear that the M1 Pro is a cut down M1 Max.

Disappointed there wasn't a new Mac mini today, wondering what the new design is really going to be, hoping it can handle a maxed out M1 Max SoC...

Also thinking about the rumored smaller Mac Pro (come on new Cube), and hypothesizing that a Jade 2C SoC (SiP?) might have better cooling, because the Cube is larger than a Mac mini...?

Regardless, my thoughts towards a solid Apple silicon Mac desktop lineup (no AIOs, just boxen)...

Mac mini

• 10-core CPU (8P/2E)
• 32-core GPU
• 16-core Neural Engine
• 64GB RAM
• 1TB SSD
• 10Gb Ethernet port
• (4) USB-C ports
• (2) USB-A ports
• HDMI port
• 3.5mm audio port

$2999


Mac Cube

• 20-core CPU (16P/4E)
• 64-core GPU
• 32-core Neural Engine
• 128GB RAM
• 1TB SSD
• (2) 10Gb Ethernet ports
• (6) USB-C ports
• (4) USB-A ports
• HDMI port
• 3.5mm audio port

$4999


Mac Pro

• 40-core CPU (32P/8E)
• 128-core GPU
• 64-core Neural Engine
• 256GB RAM
• 2TB SSD
• (2) 10Gb Ethernet ports
• (8) USB-C ports
• (4) USB-A ports
• HDMI port
• 3.5mm audio port
• Some PCIe 4 (PCIe 5?) slots for the audio folk...?

$7999

 

[altaic]

View attachment 1870174

Even better die shots side by side. It is so clear that the M1 Pro is a cut down M1 Max.

That one is really clear, too. Much easier to make out the structure. If you take AnandTech’s analysis of the M1, it shouldn’t be hard to annotate the new dies:

Edit: Oddly, it looks like the M1 Max has a 32 (2x16) core neural engine. The spec says 16 cores…

 

[altaic]

Tried my hand at amature annotation:

 

[Boil]

Tried my hand at amature annotation:

View attachment 1870417

Dual 16-core Neural Engines...?

 

[altaic]

Dual 16-core Neural Engines...?

Yeah, that’s what I mentioned in the prior post. It’s based on AnandTech’s analysis of the M1, so unless they messed up, it looks like there’s a 32 (2x16) core neural engine, contrary to the specs.

Also, I didn’t separately label the CPU L2 caches. Pretty obvious where they are, though.

Edit: It’s interesting to note that the M1 Max has 4x the SLC of the M1. That’s a lot of cache! CPU L2 caches look to be the same per core.

 

[Boil]

Yeah, that’s what I mentioned in the prior post. It’s based on AnandTech’s analysis of the M1, so unless they messed up, it looks like there’s a 2x16 core neural engine.

Thought I missed that, but I see you edited the post... ;^p

32-core Neural Engine can only be a good thing, yeah...?!?

If so, then that means the Jade 2C & Jade 4C could have 64-core & 128-core Neural Engines...!!!

 

[tlab]

Tried my hand at amature annotation:

View attachment 1870417

You forgot to highlight the frolicking whales in the top left and the face of Cthulhu above the neural engine

 

[altaic]

You forgot to highlight the frolicking whales in the top left and the face of Cthulhu above the neural engine

I thought the same thing! Couldn’t stop looking at it…

 

[MrUNIMOG]

Edit: Oddly, it looks like the M1 Max has a 32 (2x16) core neural engine. The spec says 16 cores…

Noticed the same about that Neural Engine structure appearing twice in the M1 Max.
Very odd that Apple doesn't mention it. Maybe only one is active, for whatever reason..

It’s based on AnandTech’s analysis of the M1, so unless they messed up

I don't think there is a mistake, since the exact same structure appears in the A14 too and it doesn't seem like it could be anything but the Neural Engine there.

 

[thingstoponder]

View attachment 1870174

Even better die shots side by side. It is so clear that the M1 Pro is a cut down M1 Max.

Cut down as in literally cut down or binned? That sounds like a big waste of good silicon.

 

[dugbug]

Cut down as in literally cut down or binned? That sounds like a big waste of good silicon.

the design was .. 'chopped', not the physical chips.

 

[Gnattu]

Cut down as in literally cut down or binned? That sounds like a big waste of good silicon.

If the bottom half is mostly defective they can just remove that part to increase yield. So yes it is binning, but there is no good reason to keep that part silicon so they just remove that. They are not wasting lot of good silicon because they will keep the whole die if they can make a 24-core GPU version out of it.

 

[thingstoponder]

If the bottom half is mostly defective they can just remove that part to increase yield. So yes it is binning, but there is no good reason to keep that part silicon so they just remove that. They are not wasting lot of good silicon because they will keep the whole die if they can make a 24-core GPU version out of it.

I don’t buy that. What are the odds the only part of the chip that isn’t working is a significant portion of the GPU but everything else including the CPU cores is working? This explanation doesn’t make sense. They showed two separate dies with two separate transistor counts. I have no reason to believe they aren’t two separate dies.

I’ve never heard of cutting a chip down after testing showed defective parts. They fuse the dead parts but they don’t cut them. There’s no reason for that and they would have to design the chip to be cut down perfectly. What’s the point? When Intel sells an 8 core die as a 4 core SKU they don’t cut half the chip off, they fuse the cores off.

 

[Gnattu]

I don’t buy that. What are the odds the only part of the chip that isn’t working is a significant portion of the GPU but everything else including the CPU cores is working? This explanation doesn’t make sense. They showed two separate dies with two separate transistor counts. I have no reason to believe they aren’t two separate dies.

I’ve never heard of cutting a chip down after testing showed defective parts. They fuse the dead parts but they don’t cut them. There’s no reason for that and they would have to design the chip to be cut down perfectly. What’s the point? When Intel sells an 8 core die as a 4 core SKU they don’t cut half the chip off, they fuse the cores off.

They are two different dies, but that gigantic 400mm2 M1 Max on 5nm could have a surprisingly low yield. This could be a way to recycle some defective dies, and they have to resize the die to accommodate the smaller packaging of M1 Pro if they want to do so. It's just my speculation so sorry for the confusion.