Android Common Score and Game Load Analysis | Ainuoji - 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 often see some scoring software used in mobile phone evaluation and testing. Many users just look at the superficial scores in general, and do not know why some mobile phones have high scores or why the results of different scoring software are so different. In fact, different running software has different emphasis on hardware, which also results in different results of running software.

The author is , originally from Baidu Qualcomm Bar.

Test procedure:
Basemark X
3dmark GT1 scene
3dmark GT2 scene
GLBenchmark 2.7 T-Rex
GLBenchamrk 2.5 Egypt HD
BaseMark ES Taiji
RR3=EA Real Racing 3
NFS17=EA Top Speed Car 17
Antutu， You all know
Asphalt 7=Gameloft Wild Speed 7, version 1.05, medium quality
Modern Combat

--
Take the average value after running for a period of time, The following are the loads of each frame, with a resolution of 1280 × 720
The explanation is not exact, it is only for reference
---

The number of vertices per frame, in M

Basemark X, 3dmark Scene 1, and GLBenchmark have a lot of vertices
Compared with ordinary games, it is much more. Of course, in the game, RR3 has many vertices. At least it is much more than modern war 4, NFS17
As for Antutu, the number of vertices is very small... Less than in current games

The number of pixel renderings and texture maps per frame, in M

BaseMark X shows the pressure of adversity - if you want to reach full frame 60fps, your GPU needs at least 2400M pixel processing power/fill rate
The pressure of 3dmark and GLB 2.7 is not small - 3dmark Scene 2 focuses on pixel squeezing
Antutu also made a lot of efforts on pixel squeezing
Of course, the current games are still pixel based

Number of vertex commands per frame, in M (million)
Considering the complexity of processing, there may be a small number of vertices, but a single vertex needs to do some complex shader calculations. But the result doesn't seem much worse

Number of pixel rendering instructions per frame, in M (million)
The blue one is the calculation of high precision FP32, and the red one is the calculation of medium precision FP16

BaseMark X is still under pressure
3dmark comes from the desktop, which is basically high-precision calculation of FP32
GLB is mainly based on the low precision calculation of FP16
Others, and in the game, most of them are half and half

Instruction data for single vertex/pixel rendering (average) The profile reflects the complexity of the command
The pixel part (red) is not too poor
However, the vertex (blue) command is much more complicated than the game

Bandwidth occupation of each frame, in M
Blue is the texture reading bandwidth
Red is the total reading
Green is the total write
Basically, to render a frame, the GPU needs to read 20-30M data, and then 50fps basically needs a bandwidth of 1.5GB/s.
We found that the NFS17 in the game is relatively bandwidth intensive, and 50fps at 720p is about 2G/s
Of course, the violent BaseMark X and GLB2.7 require 100M read per frame. If 50 frames are used, the bandwidth is 5GB/s I went there.

What's worse is antutu... It uses uncompressed texture. Although the picture is simple, the texture to be read per frame reaches 200MB (of course, there must be optimization problems)
So the 3D of antutu has become a bandwidth benchamrk to some extent:
Even if I can reach 30fps, the bandwidth of 6GB/s has been removed: the limit of 8.5GB/s of dual channel LPDDR2-1066 is almost unbearable

Sometimes you wonder why Adreno 320 has 720p and 1080p, but Anthare has the same score in 3d?
In fact, the GPU still has some spare power, but the bandwidth has been consumed by the rabbits

At this time, Octa GS4 equipped with dual channel LPDDR3 has advantages. With 13GB/s bandwidth Breaking the bottleneck of 30fps
That's why the score of 544MP3 in 3d mark and baseline x is the same as that of S600 A Dreno 320 similar
However, in Antu rabbit, it can break 9000 and close to 10000 points, while Adreno 320 only gets 7000 points

Some improvements in memory performance of S600 compared with 8064 have also led to an increase in Angu's score
Of course, I'm not saying that Anthare 3D itself is a bandwidth oriented benchmark. In fact, it squeezes the Fragment shader, but when the frame rate reaches or approaches 30 fps, the LPDDR2 equipped by most SoCs at present becomes a bottleneck. Its unreasonable point is that the bandwidth demand is an order of magnitude higher than the actual game.

-------------------
To sum up, I think you have a certain understanding of the focus of several benchmarks

As for whether running points is entertainment in the end -- running points is not practical, nor is running points entertainment——

Learn more about new cool devices, stay tuned