Some key takeaways, the A17...
- has SUBSTANTIALLY MORE transistors at 19 billion. That's 19% more than A16.
- runs at slightly LOWER power of ~7.5 watts versus ~8.5 watts on A16 (to be confirmed)
- increases clock speeds on the P-cores by 10%. We don't know if the E-cores also got a boost.
- P-cores have wider decode & execution units, improved branch prediction.
- GPU cores have improved efficiency, mesh shaders, and dedicated hardware ray-tracing silicon.
- has an additional GPU core for a total of 6 cores vs 5 in the A16.
- doubles the performance of the neural engine to 35 TOPS
- adds AV1 decode support
- adds USBC controller
So, if one is to take an objective look at that:
✅ higher clock speed
✅ 19% more transistors
✅ same or lower power consumption
✅ extra GPU core
✅ dedicated ray-tracing hardware
✅ 2 x faster neural engine
✅ AV1 decoder
✅ USBC
This is a SUBSTANTIAL upgrade. It's easy to get lost on the ~10% speed bump, but there's a lot more going on here. That 10% is really only the P-core performance. However what we need to see are real-world results once you factor in workloads that utilise the neural engine, as well as instances where AV1 is being decoded. Tasks bound to the neural engine will be TWICE as fast. That's SIGNIFICANT.
I would say Apple has utilised ALL of the purported benefits of TSMCs 3nm process in a very balanced way.
✅ more efficient = SAME battery life with 19% more hardware, and 10% speed increase
It could have been balanced differently. If they kept the GPU at 5 cores, we'd have probably gained some battery life and consumed less power. Assuming that the base design of the GPU is largely the same, with the ADDITION of mesh shaders and ray-tracing, which only benefit performance when those features are engaged.
Or they could have pushed the clock slightly higher while dropping back to 5 GPU cores, which would have improved performance to say 12%, but we'd be missing a GPU core and battery life would remain the same.
It's all trade-offs, right?
I guess where most people are disappointed is that the apparent architectural changes to the P-cores have not resulted in substantial clock-for-clock / IPC improvements? Someone smarter than I will be able to explain that widening the decode/execution units affects IPC versus clock speed.
By the looks of things, the P-cores may be very similar to the A16 - which is not a bad thing. Moving to a new process node can introduce its own bugs and issues, so keeping the main logic of the CPU cores the same means less to have to fix. Instead, using the new node to add additional silicon to other areas while ramping clock speed is a safer bet. It's likely then, that the A18 may have some further architectural changes to the processor cores over A17.
How does this related to M3 generation?
So I guess the big question is: is this A17 the precursor to M3? In all likelihood, yes.
Given that we know the A16 was meant to ship with the ray-tracing GPU, AND considering the M1 was A14, the M2 was A15, we can ONLY ASSUME the M3 is going to be based on A16. HOWEVER, note that the A16 was actually largely meant to be the A17 with ray-tracing. So, one can conclude that the M3 will have feature parity with the A17.
The bigger question to ask is what DIFFERENCES will there be between the M3 Mac chips and the A17.
What differences were there between an M1 and an A14? What differences were there between an M2 and an A15?
Ignoring core-counts...
Clocks will be faster on the M3 over the A17, that's almost a certainty. We're looking like 3.9GHz - 4.0Ghz is a realistic estimate.
Will memory subsystem be different?
Lots to discuss!
