On November 6, the new flagship of MediaTek, Tianji 9300, was officially released. Since the appearance of the Arm size core structure, The first mobile SoC with full core CPU and the first mobile SoC with 22.7 billion transistors This is also the first time that MediaTek has dropped its CPU, GPU and AI from competing products of the same generation in an all-round way. It means You, Snapdragon 8 Gen 3 and A17 Pro. The vivo X100 series, which launched the Tianji 9300 for the first time, will not be released until November 13, but a simple experience and test of the Tianji 9300 engineering machine will show that it is really the king of the next generation of mobile SoC performance. Specifications and official data Tianji 9300 ——New Fage Zhi ⭐ TSMC's third-generation 4nm process (Tianji 9000's N4 is the first generation, and Tianji 9200's N4P is the second generation)
The number of transistors soared from 17 billion to 22.7 billion from 920 in Tianji
CPU is one 3.25GHz X4+three 2.85GHz X4+four 2.0GHz A720
Initial LPDDR5T 9600Mbps memory support
GPU launches Immortalis-G720 MC12 (GeekBench display frequency is 1.3GHz)
The second generation ray tracing engine claims that the ray tracing performance has been improved by 46%, supports global illumination, and can run the ray tracing game at 60 frames
The seventh generation AI processor APU 790 consists of a performance core and a general-purpose core. Integer and floating-point performance have doubled, 45% less power
With hardware memory compression, the first mobile SoC capable of running 33 billion parameter AI large voice model
Tianji 9200 It is 3.05GHz X3+3x2.85GHz A715+4x1.8GHz A510 refreshed, Immortalis-G715 MC11 981MHz
Tianji 9200+ It is 3.35GHz X3+3x3GHz A715+4x2GHz A510 refreshed, Immortalis-G715 MC11 1.148GHz。
Snapdragon 8 Gen 3 One 3.3GHz X4+three 3.15GHz A720+two 2.96GHz A720+two 2.27GHz A520
Promotion of official propaganda:
CPU peak performance increased by 40%, and power consumption decreased by 33% under the same performance ， Application startup speed increased by 16%
The peak performance of GPU is increased by 46%, and the power consumption is reduced by 40% under the same performance
Daily power consumption decreased by 10% to 13% (web browsing decreased by 10%, short video by 11%, video recording by 12%, video playback by 13%), and Wi Fi hotspot power consumption decreased by 30%. High quality original+WeChat call, claiming that the frame rate increased by 15.5%, the power consumption decreased by 12.3%, and the power consumption of live broadcast+news decreased by 11.2%.
Up to 4K 120Hz or 2K 180Hz screens are supported, with AI depth of field image quality engine, and daily display power consumption is reduced by 10%
Up to 6.5Gbps Wi Fi 7, Xtra Range 2.0 technology can increase 4.5m signal coverage
Support Android's first 3-antenna two-way Bluetooth (connection speed increased by 42%, minimum 35ms delay)
The 5G baseband supports the four carrier aggregation and situational awareness functions of Sub-6GHz. It claims that the 5G Carton will be reduced by 20% and the wilderness standby will be increased by 4 hours
Dual security chips, respectively responsible for power on protection and physical isolation of sensitive data, with memory token expansion (MTE)
Tianji 9300 22.7 billion transistors , is really far ahead: Apple A16 is 16 billion, A17 Pro is 19 billion, and Apple M2 is 20 billion. Even just released Apple M3 also "only" has 25 billion transistors, while Qualcomm has not announced the number of transistors for several generations.
Historic Cancel small nucleus ， The CPU consists of four X4 super cores and four A720 super cores. The highest frequency X4 has a larger cache. From the perspective of Android SoC, the four super cores and four large cores of Tianji 9300 are actually closer to the concepts of Apple A series and Intel's P core (performance core) and E core (energy efficiency core).
Initial LPDDR5T 9600Mbps memory The speed is 12.5% higher than that of the previous LPDDR5x 8533Mbps, which is the frequency that people think can be reached only after LPDDR6 (Tianji 9000 was the first LPDDR 5x 7500Mbps memory two years ago, and Tianji 9200 was the first LPDDR5x 8533Mbps).
Hynix announced in August that it would complete LPDDR5T verification on Tianji 9300, and Micron announced on October 26 that it would start shipping samples of 9600Mbps LPDDR5x, which is also the only 1 β (1-beta) memory on mobile terminals at present. Micron 1 β process node claims to save 30% of power, and claims to be born for Snapdragon 8 Gen 3, but it is expected to be mass produced next year.
The bandwidth of LPDDR5T is the same as that of the 9600Mbps LPDDR5x. It is unknown whether there is a power consumption difference between them.
Immortalis-G720 GPU The number of cores is one more than that of the previous generation. GeekBench's display frequency is as high as 1.3GHz, which is more exaggerated than that of Tianji 9200+'s 1.148GHz.
GPU supports global lighting technology. It has twice the performance of multiple anti aliasing, twice the texture throughput, and twice the performance of pixel blending. The variable speed rendering performance can be improved by 86%. introduce Delayed vertex shading (DVS) reduces memory access and bandwidth usage to improve performance and reduce power consumption.
Imagiq 990 ISP The built-in AI semantic separation video engine supports up to 16 layers of image semantic segmentation (Snapdragon 8 Gen 3 is 12 layers).
Introduced OIS optical anti shake special core, upgraded depth of field and spot engine, and supported full Pixel focusing plus 2x lossless zoom, up to 3 microphones for noise reduction (claimed to reduce wind noise by 99% at 25km/h).
The 14bit HDR video recording cooperated with Sony claimed to increase the dynamic range by 4.8 times; Cooperate with Samsung to improve the resolution of 200 million pixel sensors by 30%; Intelligent perception in cooperation with Howell reduces power consumption by 53%.
Introduction to test platform Comparison between engineering machine and vivo S16 Pro with 6.78 inch curved screen
The MediaTek engineering machine is still familiar, with an extremely advanced 3.5mm headphone port, dual SIM card+TF card slot. The biggest appearance change comes from the screen. Last year, LCD was the main one, with only a few curved screen OLEDs. This year, OLED curved screens are all. Its cooling conditions are slightly different from those of mass production machines. It only has a basic soaking plate, and the battery capacity is still very small. But in order to "facilitate" the replacement of memory and flash memory, it is said that SoC and memory are not stacked like mass production machines. Its specific configuration is 6.73 inch 1080P+OLED screen, the top 16GB LPDDR5T 9600Mbps memory, 512GB UFS 4.0 flash memory, and native Android 14 system.- The room temperature remained between 22.8 and 23.8 degrees throughout the whole journey, the screen brightness was 280 nits, and the body temperature was fixed to 22 degrees before starting.
- Forced normal mode, without using performance mode, fan, heat dissipation clip and other tools throughout the process.
This year, there are finally strict and controllable test conditions, but the setting of 280 nit brightness is not good for high scores. The high brightness screen will grab the heat dissipation margin, which will have a significant impact on the GeekBench 6 and rabbit tests that have been lit for a long time. Tianji 9300 real pole Limited performance. It is estimated that it will take the same production machine to see. Now let's take a look at the test results of the Tianji 9300 "reserved version". CPU test: Reverse virtual label, increased by 47.5% Before the beginning of the wave, it was the old saying: Performance test/running score is like an exam. A high running score may not be a good experience, but a poor running score must be a bad experience.
With excellent testing tools, the project will be closer to the actual use of users, and the results will be stable. The most close to this ideal goal are the old GeekBench and GFXBench. Antu Rabbit has too many projects, and its proportion is difficult to control. Moreover, the official standards and scoring coefficients change frequently, and it often runs "inflation", so it is called an entertainment rabbit. So we mainly take the data of GeekBench 6 and GFXBench 5.0 as the standard, and take Anthare as the example Picture I Music Complementary. Tianji 9300 engineering machine GeekBench 6 runs ↑ GeekBench 6 and GeekBench 5 are two sets of test systems. GeekBench 6 is a multi-core single load, which is closer to daily application than GeekBench 5. It is dominated by Clang6 and does not mainly call local libraries like GeekBench 5, so it cannot reflect the impact of system optimization as before, but it is just used to test the theoretical performance of SoC.- CPU single core performance It is still the world of Apple A series. This year, the single core performance of Android camp is just close to Apple A15, and there is still a gap of about 30% between Android and A17 Pro.
- CPU multi-core Tianji 9300 outperformed A17 Pro in multi-core performance, 8.1% stronger than Xiaomi 14, 29.5% stronger than Xiaolong 8 Gen 3, 47.5% stronger than Tianji 9200
You know, our Tianji 9300 tester is tested against the screen brightness of 280 nits, and our previous tests were all run with the lowest screen brightness (The screen power consumption will compress the available heat dissipation margin, resulting in a low score) 。 GPU： Overriding Snapdragon 8 Gen 3, leading 8.9% to 24% In Vitro ray tracing test of GFXBench 5.0 and BaseMark ↑
GFXBench is an old GPU testing tool. Compared with 3DMark, the rendering pipeline and API it uses are closer to traditional mobile games and suitable for testing traditional GPU performance. GPU part The leading range of Tianji 9300 is even more exaggerated, directly killing the list. The official logo said that the GFX Manhattan 3.0 scene increased by 47%, which is also a reverse virtual logo. Several scenes are at least 50% stronger than Tianji 9200, and 62.7% stronger than the most leading scenes. The Snapdragon 8 Gen 3 is not weak in fact. This generation of GPU has also increased by 25%, but Tianji 9300 is too fierce, falling behind 8.9% to 24.4%. Compared with Snapdragon 8 Gen 2, it is 42% to 51.8% stronger. The A17 Pro, which GPU has been squeezing toothpaste for several years, need not mention. It has not yet caught up with the Snapdragon 8 Gen 2 and Tianji 9200+, and has been abandoned by Tianji 9300 by 55% to 79%. It can only sit at the children's table. AI and optical pursuit performance: True · far ahead, increased by 50% to 100% For ray tracing, the industry's testing tools have not yet been unified (after all, last year's Tianji 9200 just introduced hardware ray tracing into the mobile field). The mainstream is BaseMark's In Vitro test and 3DMark's Solar Bay test.
But two tests The percentage of light chasing is not high. Solar Bay's percentage of light chasing is only 10% to 15%. The percentage of light chasing tested by In Vitro is 25-30%, so the latter may be more able to test the performance of light chasing (if the percentage of light chasing is too low, GPUs with strong basic performance can also force up the total score). Tianji 9300 claimed that the light chasing performance was improved by 46%. In the In Vitro test, the previous record holder is Tianji 9200/9200+, which is about 3000 to 3100 points, Tianji 9300 is more than 50% stronger. In the Solar Bay test of 3DMark, the A16 scores more than 4000 points, and the A17 Pro, which boasts four times of light chasing performance, scores about 6700 points (at this time, the power consumption exceeds 10W). The leading version of Snapdragon 8 Gen 2 scores about 5600 points, and Tianji 9200+scores about 6500 points. Although the field staff did not allow 3DMark to run, they predicted with their toes that Tianji 9300 was probably the first. For AI testing, the industry divergence of testing tools is even greater. Unlike CPU/GPU, which can borrow experience from desktop platform, mobile AI is the fastest growing end AI, and no experience can be borrowed. In addition, there is no large-scale AI application, and it is still uncertain which aspect of mobile AI should focus on, let alone a unified testing tool. Now there are mainly four test tools: ETHZ's AIBenchmark 5.1 (i.e. Zurich test), AI Mark V3, AItutu V2, and MLPref 3.0. The most pressing one is AIBenchmark, which has been used by MediaTek. It supports 22 models, more than four times the number of 3-5 models tested. Its highest test chart resolution is 1536x1024, while the others are 224p, 300p and 512p. The score of Tianji 9300 tester is 3145 points, while that of Snapdragon 8 Gen 2 leading version is about 2100 points, Snapdragon 8+is more than 1800 points, Tianji 9200 is about 1600 points, and Tianji 9000 is more than 1000 points. Fage's lead is still around 50% This is partly due to the 9600Mbps LPDDR5T memory. After all, AI testing is very sensitive to memory performance. MediaTek also has an official AI demonstration tool, which is not as sophisticated as those at the vivo developer conference, but the AI performance of mobile SoC can really play with local GPT text creation (big language model) and stable diffusion, and the text and image generation speed is very fast. Finally, let's look at entertainment projects. Anthare test covers too many things and has too many variables. For example, with the same configuration, the 2K screen will be about 20000 times higher than the 1080P screen (because scanning and sliding scenes will be more dominant), and even the memory size will gain. The SoC of the same generation often violates the perception of strength, for example, the iOS platform is always low. In addition, its hardware detection of new machines often deviates from the sky, so it can only be used as a qualitative testing tool. The biggest advantage of Anthare test is that it will display the temperature rise and power consumption during the test process (the screenshot of running points dare not cut the temperature curve, probably the bellows/refrigerator running points). We have The total score of Tianji 9300 engineering machine and Antu rabbit is close to 2.14 million, and the high brightness bright screen results in a low score, but it is still more than 100000 points higher than the normal temperature Snapdragon 8 Gen 3 mass production machine It is not too much behind the dragon 8 Gen 3, which is 10 degrees below zero. The score of Tianji 9300 mass production machine has already been close to 2.25 million:- Tianji 9300 engineering machine, 2.13 million in total, 507000 CPU, GPU 928000, memory 389000, UX 312000.
- Snapdragon 8 Gen 3, refrigerator Xiaomi 14 Pro (- 10.3 degrees, teacher Feiwei), total score 2.18 million, CPU 493000, GPU 904000, memory 455000, UX 330000
- Snapdragon 8 Gen 3 QRD has a total score of 2.15 million, 494000 CPUs, 927000 GPUs, 395000 RAM and 346000 UX
- Snapdragon 8 Gen 3, Xiaomi 14, total score 2 million, CPU 430000, GPU 834000, memory 420000, UX 318000 (WHYLAB)
- Snapdragon 8 Gen 3, one plus 12, official total score of 2110000, CPU 496000, GPU 914000, memory 362000, UX 338000
- Snapdragon 8 Gen 2, Xiaomi 13, total score 1.57 million, CPU 397000, GPU 600000, memory 296000, UX 277000 (WHYLAB)
- A17 Pro， IPhone 15 Pro Max has a total score of 1.497 million, a CPU of 368000, a GPU of 526000, a memory of 264000, and a UX of 339000 (WHYLAB)
- The total score of Snapdragon 8 Gen 2 is about 1.6 million, with 380000 CPU and 600000 GPU
- The total score of Tianji 9200+is about 1.65 million, and that of 3GHz Snapdragon 8+is about 1.3 million
Energy efficiency and game measurement: The most concerned part In fact, as soon as the Tianji 9300 "4 X4+4 A720" was revealed half a year ago, everyone knew that the performance of Fage would be a winner this year, and everyone was always concerned about energy efficiency. But for the time being, it can only be tested laterally through games, Still have to wait for the vivo X100 series of these mass production machines to finally hammer. In terms of manufacturing process, only Apple A17 Pro is TSMC 3nm this year (although the actual energy efficiency seems to be no improvement over 4nm), Snapdragon 8 Gen 3 is N4P (the second generation 4nm of TSMC like Tianji 9200), and Tianji 9300 is the third generation 4nm. At the communication meeting, MediaTek said that MediaTek and TSMC in Taiwan have a very close relationship. Even for processes with the same name, MediaTek is very selective, generally leading other manufacturers by half to one generation (Apple has always been the largest customer of TSMC, while MediaTek and AMD rotate in the second and third places). In terms of architecture, the biggest improvement this year is the X4 super core, which can reduce the power consumption by 40% under the same performance (although the PPT preset is a 3nm process), while the A720 is a minor change, which can reduce the power consumption by up to 20% under the same performance. Neither Qualcomm nor MediaTek uses the function of DSU-120 to turn off some L3s, and neither can turn off individual cores (Tianji 9300 has three clusters, at most three frequencies, and Qualcomm has four). Because the manufacturing process is approaching the limit, more important innovations in the next few years should come from architecture design. And Android ecosystem has always been It is based on 8 cores, so Qualcomm and MediaTek are reducing the number of small cores to the number of large cores. The CPU of Snapdragon 8 Gen 2 is 5 big and 3 small. By Snapdragon 8 Gen 3, it has become 6 big and 2 small, and Tianji 9300 is extremely "big and bold" and cancels the small core directly.
Energy efficiency curves of A520, A720 and X4 (the position of the small core is too small to see clearly) ↑ If the energy efficiency curve of Arm is true, the Ji 9300 has the confidence to cancel the small core: The A720 has the smoothest energy efficiency curve in three cores (a slight increase in power consumption can lead to a great performance improvement). The performance of one A720 tops that of three or four A520s, and the power consumption of three high-frequency A520s must exceed that of the low-frequency A720. According to the official energy efficiency curve, Effect after horizontal vertical conversion ↑ Tianji 9300 sets the A720 at a super conservative frequency of 2GHz, which means that it uses the large core as the small one. Reduce power consumption with larger cores running at low power consumption dessert frequency, and improve single core performance with super large cores. The only disadvantage may be that it is expensive (chip area is larger). The complete elimination of small cores may also bring about anti intuitive power consumption benefits:- The most basic one: The small core that "has low power consumption and performance, but needs to work for a long time" may not save power than the large core that "has higher power consumption and performance, but has shorter working hours".
- The Tianji 9300 has the first CPU in the Android camp to execute in full disorder. The small core of the public version of arm has always been executed sequentially, and the A520 is no exception. The small core of sequential execution, when working together with the large core and super core of out of order execution, will drag down other cores, which is more likely to cause congestion (especially when involving instruction and data access).
no Small kernels are a drag. On the one hand, they can improve the speed, and on the other hand, they can shorten the working time of large kernels, make them return to sleep faster, and even facilitate the migration of desktop games (developers do not need to schedule the backward order of small kernels).
After theoretical analysis, let's take a look at the actual measurement. Although it is lateral, only high-performance+energy efficiency can conquer the original spirit with low power ↓
The room temperature is 23.5 degrees, the screen has a fixed brightness of 300 nits, the original god has the highest picture quality of 60 frames, Xu Micheng runs for 30 minutes, the last average frame rate of Tianji 9300 engineering machine is 59.6 frames, and the average power consumption is 5.09W. Other media teachers who use the test route of MediaTek+260 to 280 nites concentrate on 59.7 frames, 4.85W to 4.96W. The iron burst is 58.8 frames, between 5.12W and 5.5W. You know, it's 300 nits brighter than us Love to play computer games The 200 nit test used at ordinary times is a bit higher. Under 200 nits, the iQOO Neo8 Pro of Tianji 9200+is 6.4W, and the iQOO 11S of Snapdragon 8 Gen 2 is 6.32W. Of course, if they are also adjusted to 300 nits, the power consumption will not be much higher, because 6.4W is close to the maximum power consumption that can be maintained for a long time by non game phones (the screen grabs the power consumption, then the CPU and GPU can only reduce the frequency, resulting in frame drop and jamming). (The cooling effect of the engineering machine's screen is still slow, but the back is not at all. You can even see the outline of the motherboard and battery ↑) The maximum temperature of the actual fuselage is 40.8 ℃. It's ridiculous to open the door. It's ridiculous to make the original god into a medium game's power consumption and heat. Like the running score, Tianji 9300 also broke a new historical record in the game. Snapdragon 8 Gen 3 may not be able to achieve similar effects even on game consoles. Although the highest frame of Xumi City is the same, the number of small carts in Tianji 9300 has plummeted from 30-40 times/10 minutes for the Snapdragon 8 Gen 2 model and 10 times/10 minutes for the new Snapdragon 8 Gen 3 model to 3.3 times/10 minutes. No kidding, in the original God, powerful CPU is more practical than anything. During the interval between receiving performance dogs, we made a simple test, and opened a full white picture in the picture library under different brightness ↑
The Tianji 9300 engineering machine is 3.5W at 680 nit full white maximum brightness, while the rest power consumption is 1.5W at 300 nit brightness, and the minimum brightness is 0.8W. The rest power consumption performance is quite normal.
CPU frequency and scheduling ↑ That is to say, when the original god Xumi City ran its map, 1.5W of the 5W was fed to the screen and other components, and the CPU and GPU added up to 3.5W (during the period, the A720 core averaged 1.34GHz, the three super core X4 averaged 1.47GHz, and the super super core X4 was 876MHz).
GPT broke out at the end of last year. This year, both Qualcomm and MediaTek focus on the "generative AI" at the end. Both Xiaolong 8 Gen 3 and Tianji 9300 have the ability of text creation and cultural map, but their stacking materials and ecological scale are different.
The Tianji 9300 has once again made great efforts to fly. The APU 790 it carries is the first NPU to carry a hardware generative AI engine and support the expansion of generative AI end side skills. The hardware acceleration of the Transformer model operator is 8 times faster than that of the previous generation, and it can be used for a 1-second cultural map (the 512x512 image only takes about 0.5 seconds).
It supports such big models as Meta Llama2, Baidu Wenxin Yiyan, and Baichuan Intelligent Baichuan, and has the first "end skill expansion" function on mobile terminals. LoRA integration can be used to expand different functions of the basic big model.
More fierce, APU 7 90 Yes Hardware Memory compression, claimed 61% reduction The storage occupancy makes Tianji 9300 the first SoC that can run 33 billion parameter AI big voice model on mobile terminals.
The Tianji 9300 can compress the 13GB memory occupation of the 13 billion parameter model to 5GB (4GB memory can be reserved for the system and 6GB memory can be kept alive for the APP for mobile phones with 16GB memory), and the 33GB memory occupation of the 33 billion parameter model to 13GB (mobile phones with 24GB memory are required to run). and The large models with parameters of 1 billion, 7 billion and 13 billion are naturally worth mentioning.
As a comparison, when everyone has 77GB/s memory bandwidth, the Snapdragon 8 Gen 3 generates 20Tokens/s of 7 billion parameter models, and the Wensheng diagram is also less than 1 second. However, Qualcomm can only run 10 billion parameter models, and now only 6 billion parameter big language models have been implemented.
There is still a little threshold for using such AI tools as GPT and Wenshengtu in China. This wave of flagship chip generation AI is likely to be the "first contact" of most users.
Extraneous remarks: about 32-bit applications
This year's Tianji 9300 and Snapdragon 8 Gen 3 have pure 64 bit CPUs without native 32-bit support, but Qualcomm and MediaTek have both given their own compatible translation solutions.
MediaTek said that it has HBT technology, which can "translate" 32-bit applications into code that can run on 64 bit processors (the first cold start after installation will be about 1 second slower than the original), and claimed that the top 3000 popular 32-bit applications in the app store can run 100%. Qualcomm is rumored to have similar technologies, and the Xiaomi 14 series of the surging OS can indeed run most 32-bit applications. What's more funny is that now all cores can run together. Even with the loss of translation performance, the performance is better than that of the previous generation, which can only run with the A510 refresh small core.
The 32 bit ad free+lightweight version of online cloud suppression/color cloud weather and other old apps we have collected should not be doomed. I have a friend who said that those XX applications are smoother than the previous flagship. In the absence of great progress in the manufacturing process, Tianji 9300 showed an increase that usually takes two generations to achieve. Compared with the previous generation of CPU/GPU, which is often 40% to 50% higher, the AI performance doubled, and various new features, plus the exaggerated energy efficiency performance in the game, we can know how powerful the stack of Fage is in the semiconductor impossible triangle of "chip area (cost) - performance - energy efficiency". In a year, many things have changed. In the past two years, Fage was released in front of Qualcomm, but this year, Qualcomm's release was one month ahead of schedule - perhaps, if you know you will lose and release in front of your competitors, at least you will not lose at the starting line This year Fage has the confidence of taking his time, which comes from 22.7 billion transistors. There are also a lot of starters, but there is much less writing this year, such as the debut of Immortalis-G720, and even no mention. Fage also purposely added many performance restrictions in the test activity of Tianji 9300, such as the screen brightness of 260 to 280 nits, and the performance mode cannot be used. Naturally, air cooling or heat dissipation is not allowed. Maybe it is to leave the record and reputation of the strongest score run to ODM manufacturers, or to know that the Snapdragon 8 Gen 3 can't win them. In a word, save the flow: The Tianji 9300 doesn't need to run points to prove itself. Fage has tried his best. The rest depends on the actual performance of the vivo X100 of the Tianji 9300 on November 13.
Follow our Weibo @ Love Computer
Follow our WeChat official account: playphone
Of course, we also pay attention to our Bilibili account: love computer
