MediaTek Tianji 9200 was officially released on November 8, while Qualcomm Snapdragon 8 Gen 2 will be released the following week. MediaTek and Qualcomm's first mass production machines will be launched at the end of November, and the next generation Android flagship SoC battle is imminent.
Specification: starter fan
Last year's Tianji 9000 broke Qualcomm's monopoly on high-end Android SoC. With better technology, larger cache, and stronger memory support, even Qualcomm Snapdragon 8+Gen 1 released in the following six months failed to catch up.
As a result, this year's Tianji 9200 is a bunch of first launches: the first TSMC second generation 4nm process, the first CPU X3/A715/refresh A510, the first GPU Immortalis-G715, the first mobile SoC that supports mobile hardware optical tracking and Vulkan 1.3, the first Wi Fi 7 Ready, the first LPDDR5X 8533Mbps memory, the first UFS4.0 flash memory that supports multi cycle queues, and the first ISP that supports RGBW natively.
First generation TSMC second generation 4nm process
CPU part
GPU launches Immortalis-G715 MC11@981MHz
Initial LPDDR5X 8533Mbps memory support
UFS 4.0 flash memory support for the first MCQ multi cycle queue
Overview of lifting part
TSMC's second generation 4nm process integrates 17 billion transistors. For reference, Apple A16 is 16 billion, A15 is 15 billion, and A14 is 11.8 billion
CPU single core increased by 12%, multi-core increased by 10% (GeekBench 5), the application of cold start was about 10% faster than Tianji 9000, and Antu Rabbit ran more than 1.26 million points.
The peak performance of GPU is increased by 32% (GFXBench Manhattan 3.0 scenario), and the power consumption in the same performance is reduced by 41%. 2x floating point, FP16, AI performance, 3x triangle computing capability, first hardware level ray tracing, variable rate rendering, and Vulkan 1.3
The first LPDDR5X 8533Mbps memory, bandwidth is 13% higher than the current 6400Mbps LPDDR5
MCQ multi cycle queue UFS 4.0, 8-core 8-channel, random read and write performance is 16.7% better, with a maximum increase of 50%, twice the continuous read and write of UFS 3.1. It takes 45 seconds to transmit 20GB movies, 28% faster than UFS 3.1.
The AI performance of the sixth generation APU 690 has been improved by 35% (ETHZ AI benchmark V5), and 30 TOPS computing power
Imagiq 890 ISP， It supports up to 320 million pixels of sensors, 8K30FPS and 4K60FPS video recording. The first launch natively supports RGBW sensor, supports AI dual track capture (preview stream and shooting stream are conducted simultaneously) and similar Apple movie mode focus switching function
MiraVision 890， Up to 144Hz 2K screen, 240Hz 1080P, dual 2.5K 60Hz screen, and adaptive refresh rate
5G new dual card dual pass, low-power 5G, 5G millimeter wave support, global 13 channel satellite signal
Wi Fi 7 Ready has a peak value of 6.5Gbps in the 6GHz band. Compared with Wi Fi 6, the throughput has increased by 170% and the coverage has increased by 50% (all new features are supported, but the Wi Fi 7 standard has not yet been finalized, so the standard can only be Ready. In addition, the 6GHz band is not open in China).
Bluetooth 5.3, supporting LE Audio and AURACAST Bluetooth broadcast audio, up to 24bit/192kHz decoding (using LHDC-V), peak 8Mbps throughput
Test platform
The engineering machine of MediaTek this year is still the "brick" mold of the previous year, but somehow the back plate is basically sealed. From the thermal imaging diagram and the disassembly drawing, it should have a simple heat dissipation configuration like soaking plate (it should be a flagship, except Apple). In addition to the self body type, the heat dissipation environment this year is not bad.
In addition to the Tianji 9200, the test machine is the top one today, and there is no 12GB LPDDR5x 8533Mbps memory of the mass production machine, as well as 256GB UFS 4.0 flash memory, true ceiling configuration.
The rest include 6.59 inch 120Hz 2400x1080 LCD screen, which is equipped with native Android 13 and F2FS file system. The system is so simple that even the performance mode is an independent app.
Performance measurement: CPU slightly increased, GPU soared
For the CPU and GPU, we found Tianji 9000/9000+, Snapdragon 8+/888/870 and A16 for performance comparison.
Streaming saving means that the CPU peak performance is not improved much, but the GPU performance is directly cracked like a flash flood. MediaTek unexpectedly gave GPU a "reverse false mark" of increase, with the actual average increase exceeding 35%.
Test data comparison ↑
Percentage of gap between each SoC and Tianji 9200 (Tianji 9200 leading range) ↑
CPU：
12% and 5.8% stronger than Tianji 9000 single core multi-core
6.5% and 2.8% stronger than Tianji 9000+single core multi core
Single core is 7% stronger than Snapdragon 8+, and multi-core is almost equal
GPU：
48% to 79% better than Tianji 9000
35% to 46% better than Tianji 9000+
15% to 37% better than Snapdragon 8+
Milestone surpasses Apple A16 by 1% to 22%
Ratio of Tianji 9200 to each SoC ↑
After several generations of crazy improvement by the manufacturer, the GPU performance of Tianji 9200 is now 2.5 times that of Snapdragon 870 and twice that of Snapdragon 888/Tianji 8100. The increase is somewhat magical.
Screenshot of Teacher Fei Wei ↑
Due to the limitation of field test, the top LPDDR5x 8533Mbps memory and UFS 4.0 flash memory can only be reflected in the MEM score of Anthare. MEM test includes memory read/write speed, application storage speed, random read/write speed and sequential read/write speed.
The MEM score of Tianji 9200 rabbit is about 245000, while that of another Tianji 9000+flagship terminal is 210000, and that of Snapdragon 8+flagship using the overclocking version of UFS 3.1 is less than 200000.
The improvement of LPDDR5X memory+UFS 4.0 flash memory is second only to that of GPU, which leads to an increase of 16.7% to 22.5%.
According to the existing information, the sequential write of UFS 4.0 exceeds 3500MB/s, and the write exceeds 3000MB/s, which is 74% higher than the UFS 3.1 record (1950MB/s) maintained by MediaTek itself. Large game loading and large file transmission should speed up significantly.
Performance analysis: X3+A715+new A510
To analyze the CPU of Tianji 9200, we can't get around the Cortex-X3, Cortex-A715 and Cortex-A510 Refreshed architectures released at the end of June.
In terms of super core, the last generation of Cortex-X2 super core is to delete 32-bit circuits, while Cortex-X3 is specially designed for 64 bit, optimizing branch prediction, reading, decoding, etc., improving performance and reducing power consumption.
X3 focuses on upgrading the front-end scale, changing from 5 transmissions to 6 transmissions, increasing the disordered execution window from 288 to 320, and increasing the number of integer ALU units from 4 to 6. It introduced L2 BTB for large-scale application optimization of large instructions. Now there are L0, L1 and L2, and L0 BTB has increased by 9 times. Although its scale is still smaller than Apple's A14 core two years ago, it is certainly stronger than the previous public version of the arm architecture.
On PPT, the preset frequency for X3 by arm is 3.3GHz. At this time, X3 is 22% higher than X2 in SPEC06, 27% higher in SPEC17, and 25% higher in Geekbench 5. If the frequency of 16MB L3 cache+3.6GHz is used, it will be 34% stronger than the "latest mainstream notebook".
The big core A710, which had the weakest energy consumption ratio in the previous year, is the most obvious power consumption improvement of this generation: The performance is improved by 5%, and the power consumption is reduced by 20% under the same performance.
The A715 has also undergone significant changes. It completely abandons the burden of 32-bit support and focuses on improving the throughput. The pipeline width and depth have also been improved (if not the throughput). There are branch predictors and prefetching enhancements learned from the X series core. Arm indicates that the A710 can already provide the same level of performance as X1.
In addition, many online articles wrote that "A715 only occupies 1/4 of the previous chip area", which is actually a translation error. In fact, the encoder is 4 times smaller than A710, not the overall chip area.
The small core A510 refreshed, here referred to as the new A510, is the first small core architecture that supports SVE and SVE 2. In turn, it adds 32-bit support, but this is an optional part, and the manufacturer may not add 32-bit support. In the case of more complex chips, arm claims that it also reduces power consumption by 5% without affecting the chip area.
For the Tianji 9200 with a pure 64 bit X3 and A715 architecture, the four 1.8GHz new A510s are responsible for all 32-bit applications. The Qualcomm Snapdragon 8 Gen 2 uses X3+2xA715+2xA710+3xA510, and the two A710s are reserved for compatibility with 32-bit applications.
GPU side, Immortalis-G715 MC11@981MHz The number and frequency of the cores of Tianji 9000 are in Mali-G710 MP10@850MHz above. The above test also shows that even if When the performance mode is turned off, the performance loss of Tianji 9200 is only between 2% and 8%, and the output is far more stable than expected.
Fage features: energy consumption ratio, energy consumption ratio, or TM energy consumption ratio!
Fage did not write about the performance of 35% GPU in the press conference, but kept emphasizing the improvement of energy consumption ratio.
1. The CPU of Tianji 9200 claims that the power consumption of web surfing and video recording will be reduced by 20%, WeChat will be reduced by 15%, and the power consumption of personal hot spots will be reduced by 70%
2. The video super division energy efficiency of APU 690 is increased by 45%, the power consumption of 4K30FPS AI night shot is reduced by 25%, and the noise reduction ability is increased by 20
3. After upgrading to a 6nm wireless connection chip, Bluetooth power consumption is reduced by 33% and Wi Fi power consumption is reduced by 70%
4. For the 5G part, it is claimed that the power consumption of heavy app download scenarios will be reduced by 19%, the power consumption of web browsing/video/games and other network light load scenarios will be reduced by 10%, and the standby power consumption will be reduced by 13%
5. The power consumption of heavy loaded games is reduced by 25%, the average power consumption of 120FPS MOBA games is reduced by 21%, the glory of the king is reduced by 15%, and the god is reduced by 20%
6. The 30 minute test of the official Yuanshen 60FPS extremely high image quality+5G network shows that the average frame rate is 15% higher than that of the previous generation, and the jitter rate is 63% lower
7. There is also a package cooling capacity related to power consumption, which is increased by 10%. The temperature rise speed from 20 ℃ to 95 ℃ is 4 times slower than that of the previous generation.
Physically, GPU may have the biggest increase in energy consumption ratio. Immortalis-G715 MC11@981MH The peak performance of is improved by 32% (GFXBench Manhattan 3.0 scenario), and the power consumption in the same performance is reduced by 41%.
As we all know, including the original god, the bottleneck of mobile games now lies in the CPU, and the Snapdragon 8/Tianji 9000 generation GPUs are not full. but Strictly speaking, the CPU is not lack of performance, but overheating and frequency reduction.
The increase in the GPU energy consumption ratio of Tianji 9200 is far more beneficial than the increase in peak performance. It can now run at a lower frequency with lower power consumption, leaving the limited thermal load of the body (generally 6-7W for mainstream mobile phone bodies) to the CPU, so that the CPU can run at a higher frequency.
In the field test, the room temperature was 21 degrees, the naked eye "fixed brightness" was 200 nits (the brightness could not be measured on the spot), and the Xumi City of the original God had a very high quality of 60FPS for 15 minutes. The highest temperature of the Tianji 9200 engineering machine was only 35 degrees, which was significantly smoother than the current flagship of Xiaolong 8+and Tianji 9000+, and the irregular frame drop was significantly less (however, the field did not allow the frame rate to be measured). Unexpectedly, the hope of conquering the original god was rekindled so soon.
MediaTek launched hardware ray tracing this year. The main effect is to use light tracing to create soft shadows and more realistic reflection and refraction effects. MediaTek did not disclose the information about the light chase unit, which is actually far beyond the level of "one frame is smooth, two frames of e-sports", because the light chase game Tencent "Breakthrough in the Dark Zone" on the spot can run 30 frames, and the light chase on the mobile end is at least a little hopeful.
X3/A715/Energy efficiency curve of new A510 ↑
MediaTek's "obsession" with power consumption control is most evident in CPU frequency setting.
In the case of more advanced manufacturing process and architecture, and more suitable for high frequency, except for the doubling of L2 cache of A510, the CPU frequency and other cache configurations of Tianji 9200 are consistent with those of Tianji 9000, making it an excellent model for the improvement of X3/A715/new A510 architecture.
Although the PPT of arm did not mention the specific energy consumption ratio increase of X3, from the official intergenerational energy consumption ratio curve, X3 has the steepest curve (it can save more power under the same energy). The A715 with 20% improvement in performance and energy efficiency is also set at the same frequency as before.
Arm said that when maintaining the core configuration of 1+3+4, the X3+A715+new A510 will increase by 12% compared with the X2+A710+A510 last year. From the final score of 6% increase in multi-core, its power consumption decline should be very objective (A715 itself has 20% increase in energy consumption ratio), which in turn leaves a very sufficient space for the future Tianji 9200+to raise frequency.
MediaTek is resisting the pressure of public opinion, focusing on the CPU of Tianji 9200 to suppress power consumption The current line seems to be "one generation is responsible for raising the specification and one generation is responsible for reducing power consumption".
Qualcomm next door, on the other hand, takes the opposite route. According to the current information, Qualcomm Snapdragon 8 Gen 2 is a new combination of X3+2xA715+2xA710+3xA510. This is the new DSU-110 large and small core scheduling architecture of arm, which provides 1+3+4, 1+4+4, 2+2+4, and 8+4+0.
Qualcomm is equivalent to replacing a small core with a large one, and also has two A710 "mops". According to the data leaked by Samsung Galaxy S23/S23+, the X3 super size of Snapdragon 8 Gen 2 has been approved at 3.36GHz, which is more radical than the official arm. This has pushed its GeekBench single core score to 1500 and its multi-core score to more than 4800.
However, the power consumption control is conservative here. In addition, the new 1+2+2+3 structure adds core clusters, and there may be application adaptation and scheduling problems to solve at the initial stage.
For a few disappointing 32-bit applications in China, MediaTek added 32-bit support to the A510. Qualcomm reserved two A710s for this purpose. Fage is more radical here.
For mobile SoC, peak performance and power control are two ends of the seesaw. But as for the question of "which side is more important", the annual dark horse Tianji 8100 and Xiaolong 888 jointly told us—— No matter how strong the peak performance is, what's the use of overheating and failure to run? The improvement of energy consumption ratio is indeed more urgent than the improvement of peak performance. This time, I will make a big contribution.
Fage's turnkey plan, the possible improvements of the new machine next year
Now let's guess. In the peripheral configuration, the most practical next year may be Tianji 9200's support for adaptive refresh rate. The MiraVision 890 (which can be generally understood as a display chip) of Tianji 9200 has hardware support for LTPO, so it does not require the terminal manufacturer to change the system bottom layer for refresh rate adaptation. It can detect the frame rate required by the current picture, directly send the frame rate that should be displayed to the DDIC, and claim to reduce the power consumption by 35%.
Looking back on the past two years, OPPO、 Both Xiaomi and Vivo have made different plans for LTPO, and their power saving effects vary greatly, but there is a huge power consumption control gap between them and the LTPO screen of Apple's iPhone 13 Pro/14 Pro (Apple's LTPO can achieve high brush, brighter and more power saving). MediaTek, an upstream chip factory, provides hardware based native LTPO support. Maybe there is hope for Android's LTPO screen power consumption.
Also MiraVision 890, Tianji 9200 has its own HDR multi block hierarchical color management. At present, only the flagship of Huawei and Vivo have multi block HDR management similar to that of iPhone (for example, in the album, HDR management is done separately for the system and photos, and HDR is done by using the brightness margin of the screen to make the bright places brighter), and all of them are software solutions.
If Tianji 9200 has its own support, other manufacturers should be able to have similar support soon, so that the Android camp can also popularize HDR multi block layered color management. The typical upstream turnkey scheme is the most direct driving force for technology popularization. First, set up a flag. Next year, there should be many flagship ships that specifically mention similar functions in the press conference.
The Imagiq 890 ISP of Tianji 9200 still supports up to 320 million pixels, as well as 8K30FPS and 4K60FPS video recording. The first launch of RGBW sensor support (previously OPPO Mariana X purposely added corresponding support) is an important benefit for popularizing RGBW sensors. RGBW can make CMOS more sensitive, which in turn can make small CMOS more effective, helping to reduce the size of camera modules.
Others like AI dual track capture (preview stream and shooting stream are conducted at the same time) and focus point switching function similar to Apple movie mode will probably also appear in the Android flagship conference later (Lao Luo: what XX is installed is the supplier's technology).
Summary: the route dispute between MediaTek and Qualcomm
The mass production progress of Tianji 9200 is much faster than that of the previous generation, and the brand face is much more than that of the previous generation. At the end of the press conference, MediaTek said that the flagship mobile phone equipped with Tianji 9200 would be released at the end of November, and then all mobile phone manufacturers would take turns to take the stage.
Vivo said that Tianji 9200 would be launched for the first time, and OPPO was the first to be launched. Xiaomi and Median, as well as ASUS and Glory, who came to the stage later, did not mention the release time of the terminal. It can be inferred that the vivo X90 series will be released at the end of November, and the OPPO Find X6 series should also be released in December.
Finally, the battle situation of the next generation Android flagship is predicted. In addition to everyone's surpassing the epic GPU milestone of Apple A16 and making it possible to conquer "The Original God" again, the confrontation between MediaTek and Qualcomm may be a bit like this year's Tianji 8100 vs Snapdragon 888, which is a route battle between "energy consumption priority" and "peak performance priority".
If we say that last year's Tianji 9000 was assisted by TSMC, that day's Ji9200 was a positive confrontation between MediaTek and Qualcomm with its mobile concept of focusing on energy consumption ratio.
Follow our Weibo @ Love Computer
Follow our WeChat official account: playphone
Of course, we also pay attention to our Bilibili account: love computer
