Become a MacRumors Supporter for $50/year with no ads, ability to filter front page stories, and private forums.

ediks

macrumors 6502
Original poster
Nov 17, 2015
452
808
I have a question. Isn’t the neural engine in A12 more powerful than the one in the A13 since it can perform 5 trillion 8bit operations per second compared to the 1 trillion of the A13?? It’s also bigger in size(4,64 vs 5,79) .It looks like a big downgrade to me. Could it be the reason the XS series had mediocre battery life?

source : iPhone XS and 11 keynotes
 

Attachments

  • AC62E869-F88D-4B8D-983C-2BE6632F251A.jpeg
    AC62E869-F88D-4B8D-983C-2BE6632F251A.jpeg
    156.8 KB · Views: 1,570
It is only physical size of block in chip. Check GPU Core - it is same size as A12 and it is 20% faster.
You can put more transistors to same or smaller size of chip or make that unit or core faster in the same or smaller size.
 
  • Like
Reactions: nordique
It is only physical size of block in chip. Check GPU Core - it is same size as A12 and it is 20% faster.
You can put more transistors to same or smaller size of chip or make that unit or core faster in the same or smaller size.
Yeah but still , the number of operations per second is lower. 600 billion/sec on A11, 5 trillion/sec on A12, 1 trillion/sec on A13. I just think it’s strange that the number went down yet the NPU is 20% faster (not talking about GPU)
 
Yeah but still , the number of operations per second is lower. 600 billion/sec on A11, 5 trillion/sec on A12, 1 trillion/sec on A13. I just think it’s strange that the number went down yet the NPU is 20% faster (not talking about GPU)
Its possible that Apple decided that it was overkill, and that they could save a bit of power and die space by shrinking it.
 
  • Like
Reactions: ediks
Yeah but still , the number of operations per second is lower. 600 billion/sec on A11, 5 trillion/sec on A12, 1 trillion/sec on A13. I just think it’s strange that the number went down yet the NPU is 20% faster (not talking about GPU)

Where did you get 1 trillion for A13 ?
From Apple event "The A13 Bionic uses the same eight-core neural engine design as the A12 but is up to 20 percent faster and uses 15 percent less power."

You can't be faster and do 5 times less operations...

EDIT : Yes you can, as I posted below

iphone_cpu_breakdown_2019_a13_bionic.jpg
 
Last edited:
Where did you get 1 trillion for A13 ?
From Apple event "The A13 Bionic uses the same eight-core neural engine design as the A12 but is up to 20 percent faster and uses 15 percent less power."

You can't be faster and do 5 times less operations...

View attachment 911727
It was shown just minutes before this slide

I included the old ones too
 

Attachments

  • 7B6B0A75-3FC3-4AB5-901A-ED53BCBCAEA7.jpeg
    7B6B0A75-3FC3-4AB5-901A-ED53BCBCAEA7.jpeg
    144.4 KB · Views: 306
  • A3605D57-515D-4E44-8927-769092336FB4.jpeg
    A3605D57-515D-4E44-8927-769092336FB4.jpeg
    111.2 KB · Views: 308
  • 1223BE57-4AF9-4573-871D-FB1D2289C6CA.jpeg
    1223BE57-4AF9-4573-871D-FB1D2289C6CA.jpeg
    98.8 KB · Views: 302
Yeah I found it before you posted it but "That efficiency boost is despite fitting a record 8.5 billion transistors inside and upping performance by roughly 20 percent across the board — and Apple says the chip now has the most machine learning (ML) performance, too, with an eight-core neural engine that adds “6x faster matrix multiplication.”

Probably make 6x more or faster matrix multiplication in one operation - so it is equivalent to 6 trillion operations with 1 matrix multiplication in A12 and 20% faster makes sense...
 
  • Like
Reactions: Tekguy0
Yeah I found it before you posted it but "That efficiency boost is despite fitting a record 8.5 billion transistors inside and upping performance by roughly 20 percent across the board — and Apple says the chip now has the most machine learning (ML) performance, too, with an eight-core neural engine that adds “6x faster matrix multiplication.”

Probably make 6x more or faster matrix multiplication in one operation - so it is equivalent to 6 trillion operations with 1 matrix multiplication in A12 and 20% faster makes sense...
That makes sense. But i’m wondering why they didn’t use the same metrics and write 6 trillion. But the wikipedia states that the speed is for 8bit operations for all of the chips.
 
Because they are quite truthful when comes to numbers of performance or battery and because it is 1 trillion.
If you have 10 operations but one operation is adding
and have 10 operations but one operation is multiplication... For 6x6 result you must do 6 adds or one multiplication.
So yes number of operations is lower but you should trust that it is 20% faster overall (or 20% faster for machine learning or other camera or image processing (deep fusion, potrait mode, night mode etc..)
 
  • Like
Reactions: nordique and ediks
Register on MacRumors! This sidebar will go away, and you'll see fewer ads.