M4 Rumors (requests for).

Jordanmarkstone said:
Just clicked on his profile to see what he’s been posting at MR….

“

Oops! We ran into some problems.​

This member limits who may view their full profile.”

Hahaha.
Click to expand...
I haven’t been over there in a long time, but the last thing I remember him posting was insulting me by comparing me to Jim Keller, and me responding with some sort of snarky remark about Jim switching companies a lot.
 
Not sure about this guy’s track record, but while discussing ML on ASi, a self-described Apple leaker chimed in, and now I’m irrationally excited.

x.com

x.com x.com

IMG_3321.jpeg
 
I should add that, in that thread, there’s also a nice comparison between Apple’s offerings and NVidia’s H100, including power, performance, memory, memory throughput, and cost.
 
It will be interesting to see if M4 Ultra is back to the ultra-fusion design. M3 Max was a different layout from M3 Pro – the previous versions looked exactly like Por was a cut-off of the Max at the GPU end. But they still could go with the variant Max, applying ultra-fusion, on M4.

November is not a typical release window for Studio/Mac Pro platforms. Apple is feeling pressure around intelligence modeling though, and my shift their window slightly because of that.
 
It seems like Apple's GPU development slowed down?

M4 is using the same GPU cores as M3 just clocked higher. In fact the CPU at least the e-cores are also the same as M3. Wonder if Apple has bigger plans for M5?
 
exoticspice1 said:
It seems like Apple's GPU development slowed down?

M4 is using the same GPU cores as M3 just clocked higher. In fact the CPU at least the e-cores are also the same as M3. Wonder if Apple has bigger plans for M5?
Click to expand...
I’ve heard whispers that they’ve been focusing on throughput for peripheral connectivity. Should be interesting. Especially computer to computer connections.
 
exoticspice1 said:
It seems like Apple's GPU development slowed down?

M4 is using the same GPU cores as M3 just clocked higher. In fact the CPU at least the e-cores are also the same as M3. Wonder if Apple has bigger plans for M5?
Click to expand...
The M3 GPU was such a big change from the M2 that the M4 GPU isn’t doesn’t really qualify as a slowdown - especially since the first M4 chip launched so soon after the M3. Plus M2 to M3 didn’t have a GPU clock speed increase, probably due to the complexity of the new cache structure, so, taken together, in my mind, we were due for a simple GPU clock speed boost. That said, I was mildly disappointed that it didn’t have increased matmul especially given the (possibly transitory) market conditions. But overall, not surprising. Most new GPU generations are every 2 years from Nvidia/AMD.

Altaic said:
I’ve heard whispers that they’ve been focusing on throughput for peripheral connectivity. Should be interesting. Especially computer to computer connections.
Click to expand...
Interesting.
 
dada_dave said:
The M3 GPU was such a big change from the M2 that the M4 GPU isn’t doesn’t really qualify as a slowdown - especially since the first M4 chip launched so soon after the M3. Plus M2 to M3 didn’t have a GPU clock speed increase, probably due to the complexity of the new cache structure, so, taken together, in my mind, we were due for a simple GPU clock speed boost. That said, I was mildly disappointed that it didn’t have increased matmul especially given the (possibly transitory) market conditions. But overall, not surprising. Most new GPU generations are every 2 years from Nvidia/AMD.


Interesting.
Click to expand...
To clarify, I’m biased toward those whispers; we’ll see
Edit: I don’t know shit😄🥰🤗
 
Last edited:
Altaic said:
Not sure about this guy’s track record, but while discussing ML on ASi, a self-described Apple leaker chimed in, and now I’m irrationally excited.

x.com

x.com x.com

View attachment 31115
Click to expand...

Can’t comment on the LLM performance post. The claims about FP16 performance are however nonsensical. M2 Ultra peak performance for FP16 is the same as for FP32 at around 13.5TOPS (or 27TFLOPS if you count in FMAs)
 
Last edited:
leman said:
Can’t comment on the LLM performance post. The claims about FP16 performance are however nonsensical. M2 Ultra peak performance for FP16 is the same as for FP32 at around 13.5TOPS (or 27TFLOPS if you count in FMAs)
Click to expand...
I wondered if he was referring to the performance of the entire soc, but no, he says gpu. Weird.
 
Jimmyjames said:
I wondered if he was referring to the performance of the entire soc, but no, he says gpu. Weird.
Click to expand...

Even then it won’t make any sense. The matrix coprocessor offers less than 16TFLOPs of FP16, the CPU cores add a tiny bit on top of that. Still way too short of the claimed figure.
 
Jimmyjames said:
This person gets it!
Click to expand...
This new one whose cropped up in the last couple of months in the Apple Silicon forums, at least new to me, Minghold, is way worse. I’d take mi7chy any day over them. Seriously. Like mi7chy is just a troll, but Minghold is actually kind of disturbing - posting heavily conspiracy theory stuff.
 
dada_dave said:
This new one whose cropped up in the last couple of months in the Apple Silicon forums, at least new to me, Minghold, is way worse. I’d take mi7chy any day over them. Seriously. Like mi7chy is just a troll, but Minghold is actually kind of disturbing - posting heavily conspiracy theory stuff.
Click to expand...
I haven’t encountered this person. Now I’m half way between checking his posts out and avoiding entirely!

Edit: first post saw something about KDE Plasma…I”m out.

Edit2: second post I read is some drivel about “your computer isn’t yours”. Are we sure this isn’t another mi7chy alt lol.
 
Last edited:

Similar threads

T
Request for Adobe Acrobat Pro benchmarking
Replies
9
Views
701
dada_dave
D
fooferdoggie
Florida Senators Request More Federal Aid Despite Not Voting For Hurricane Relief
Replies
0
Views
189
fooferdoggie
fooferdoggie
Back
Top