The impact of L2 cache frequency on mobile phone performance | Ainongji - 上海419贵族宝贝-上海后花园1314-上海品茶网-阿拉爱上海

Editor's note: We have always focused on the performance of mobile phones with respect to CPU frequency and GPU frequency. What impact will the frequency change of L2 cache have on mobile processors? Some third-party kernels and tools have allowed us to modify the frequency of L2 cache by ourselves, which also makes the test more intuitive.

Original text transferred from Baidu , Author

Test environment:

Mobile: HTC One

OS: 4.2.2
CPU: 1890MHz x4
Hotpulg：MPDecision
RAM: 533MHz 64bit
L2 Cache：1134MHz/1296MHz
GPU：450MHz
IO Scheduler：ROW

Antu rabbit evaluation

L2:1134MHz

L2:1296MHz

simulator

L2 1134MHz CPU 1890MHz

L2 1134MHz CPU 1566MHz

L2 1296MHz CPU 1890MHz

What we can see is that under L2 1.13G, the CPU 1.5G and 1.9G are only 9FPS lower than each other, while under L2 1.3G, the performance of 1.9G simulator reaches 131FPS

Under L2 1.13G, the scores of 1.9G and 1.56G differ by 7.8%, and the frequency increases by 20.7%

Under L2 1.3G, the score increased by 5.6% under 1.9G, and L2 frequency increased by 14.3%

If L2 can exceed 1.5G in the future (it was exceeded in the 8260 era), the score is estimated to be 135+

Sunspider

First of all, to make a digression, the Japanese spider test browser is OK, but the test performance is too bad The native browser Xiaomi 2S 1.7G can easily break 900, while Niuwan 1.9G+L2 can break 900ms. However, Xiaomi 2S native browser does need to load web pages faster, unlike T3 Sun Running Spider, the actual loading speed is not as fast as 8260. It also has something to do with the kernel. My overclocking kernel does not need 1.9G to break 1000ms, but the official 1.7G kernel can break about 950ms

L2 1134MHz

L2 1296MHz

Performance improvement: 16%

I think this is a great pressure on the second buffer, and the test results also think so

L2 1134MHz

L2 1296MHz

Score improved: 28.7%

Increase in L2 cache frequency:
(1296-1134)/1134=14.3%

Average improvement of next grade:
(8.5%+5.6%+16%+28.7%)/4=14.7% free friends can measure the soft solution, and there is also an improvement in visual inspection. When the L2 frequency is much lower than the CPU frequency, applications and scores with high requirements for cache and multi-core will naturally make L2 a bottleneck. What will be the result when the L2 cache exceeds 1.5G or even higher in the future?
Is it too far fetched to judge the performance of the CPU architecture by performance when the L2 frequency is far lower than the CPU frequency?

Soft solution
Yu Jian AB OP 1080p 60fps ref=16
BSPlayer 1.9.151 (Note that the 1.4 version will be higher, and even some machines will be 10FPS higher)

L2 1134MHz
avg 32.18

L2 1242MHz
avg 33.84

Lift 1.66FPS
5.16%

Comparison chart:

Average increase: 12.8%
L2 frequency increase: 9.5%