Why Apple A17 Pro's 10 % CPU Performance Increase Is Fine
Some people seem disappointed in the performance increase of the newly announced Apple A17 Pro. To offer some perspective this post assesses the CPU performance from the perspective of chip design as well as end user.
Chip Design Perspective
A17 uses TSMC N3 node, while the A16 uses N4, an optimized N5 node.
The greatest benefits of the new node are performance and density increase and power reduction. TSMC claims around 15 % performance gain from N5 to N3. Given that last years A16 already benefited from some efficiency gains, the jump from N4 to N3 should be smaller than TSMC numbers.
Apple put 19 % more transistors in the A17 compared to the A16. As per their keynote they reworked several parts of the SoC.
The GPU is now 20 % bigger, with a updated architecture and ray tracing support.
The Neural Processing stays at 16 cores, but almost doubles its throughput to 35 trillion operations per second. In the past the smallest data type supported by Apples NPUs was FP16. One way to achieve the doubling of NPU TOPS is data types that only need half as many bits like FP8 or INT8. This number format is also supported by the latest NVIDIA GPUs and is more frequently used to replaces FP16 for inferencing. So I think that the additional transistors weren't spend here for the most part.
Than there are several smaller functions added that take something out of the transistor budget as well like a USB3 controller, and HW decoder for ProRes and AV1.
After taking all this into account I can imagine that there weren't many transistors left over to spend on improving the CPU. A tradeoff had to be made. So the 10 % performance increase that can be observed e.g. in Geekbench is largely achieved by increasing the frequency. Of course one can now ask, whether Apple should have made a different tradeoff and prioritize CPU performance over the other things. Which brings us to the end user perspective.
End User perspective
For the end user, what would have changed if CPU got 20 % faster, instead of only 10 %. Nothing really. Apples CPU already delivers high end desktop class performance. And phones generally are used for less demanding tasks, like browsing the web, social media or streaming videos.
One frequent use case thats needs a lot of performance is gaming. But gaming performance generally benefits more from increased GPU performance compared to increased CPU performance. For the first time with the new A17 there are gaming console games playable on a Smartphone. Its not clear, whether that would have been possible without the big GPU improvements.
Another indicator that CPU performance of the A17 is fine can be found by comparing it to the Geekbench 6 scores of current Android flagship SoCs.
KPI | A17 | A16 | Tensor G3 | SD 8 Gen 2 |
---|---|---|---|---|
Process Node | TSMC N3 | TSMC N4 | Samsung 4 LPP | TSMC N4 |
Transistor Count | 19 B | 16 B | ||
GB 6 SC Score | 2894 | 2554 | 1772 | 1990 |
GB 6 MC Score | 7190 | 6459 | 4072 | 5270 |
NPU TOPS | 35 TOPS | 17 TOPS |
The performance of the Qualcomm Snapdragon 8 Gen 2 in the Samsung Galaxy S23 Ultra is comparable to that of an an Apple A14 from three years ago. Despite this the S23 Ultra is generally regarded as a great phone.
The Google Tensor G3 in the newly launched Pixel 8 does even worse. Its comparable to Apple A13, launched with the iPhone 11 in 2019. And at least Marques Brownlee in his Pixel 8 review seems totally fine with that performance level. He doesn't once mention the CPU performance.
Conclusion
Apple had only a limited amount transistors to design new features for the A17 Pro. They decided that having 50 % more CPU performance compared to the best Android SoC is enough. Instead they invested in many other parts of the SoC, mainly the GPU, thereby enabling to play gaming consoles games on a Smartphone for the first time. Thats an ok tradeoff to me.