MediaTek Tianji 9400 was officially released on October 9, 24. This is Tianji's flagship SoC, which was released in early October for the first time.
Fage's CPU and GPU strategy focuses on energy efficiency, which is the core reason why Tianji 9400 uses X925 super core+X4 super core+A720 super core. This generation of NPUs has added a large number of end-to-end AI capabilities, and even launched end-to-end training (LoRA training) and graphics video capabilities.
Tianji 9400 (MT6991) Brief specification:
TSMC N3E process, 29.1 billion transistors
One 3.62GHz X925 super core+three 3.3GHz X4 super cores+four 2.4GHz A720
GPU is Immortalis-G925 MC12, AI processor is NPU 890
First launch LPDDR5x 10.7Gbps memory supporting 10.7Gbps
It will be launched by vivo X200 series released on October 14, followed by OPPO Find X series
Brief specification table of flagship SoC ↑
Structure and changes
The latest Fage ⭐：
TSMC's third-generation 4nm → second-generation 3nm process (that is, the same N3E as Apple's A18/A18 Pro);
22.7 billion transistors → 29.1 billion transistors;
[CPU] is the second generation full core architecture. It claims that the single core performance is 35% higher than that of the previous generation, the multi-core performance is 28% higher, and the power consumption is 40% lower than that of the previous generation:
The first X925 architecture super core, the architecture code is Black Hawk, which was previously thought to be called X5 super core. 3.62GHz, 2MB L2 cache, IPC increased by 15% (under the same frequency, the higher the IPC, the stronger the performance);
Three X4 super cores, 3.3GHz, each with 1MB L2 cache;
Four A720s (also called energy efficient cores), 2.4GHz, 512KB L2 cache for each;
12MB L3 cache+10MB SLC system cache;
Compared with Tianji 9300, the L2 cache of each core of Tianji 9400 doubled, and the L3 cache also increased by 50% (Tianji 9300 is 8MB L3+10MB SLC). Large cache is good for energy efficiency, but it costs transistors and money ↓
Tianji 9400 maximum support 10.7Gbps LPDDR5X memory (Both GPU and AI are sensitive to memory bandwidth)
Samsung completed the verification of 10.7Gbps LPDDR5X memory with Tianji 9400 in July this year. Compared with the 8533Mbps version, the new memory performance is improved by 25%, and the power consumption under the same performance is reduced by 25%;
The LPDDR5T memory of Tianji 9300 is 9.6Gbps, which was first launched by the emperor version of X100/X100 Pro last year. Because of the cost and mass production, the later models are all LPDDR5X memory.
【GPU】 It is the first Immortalis-G925 MC12 1.612GHz, with a frequency increase of 23%.
Arm previously said that the G925 is the GPU with the highest performance and efficiency. Its shader cores have been increased from 10 to 16 in G720 to 10 to 24.
Its frequency is soaring - Tianji 9200 is Immortalis-G715MC11 981MHz → Tianji 9300 is Immortalis-G720 MC12 1.3GHz → Tianji 9400's Immortalis-G925 MC12 even mentions 1.612GHz, which is higher than A18 Pro GPU's 1.45GHz.
PS： 10 cores or above are called Immortalis-G925, and up to 24 cores (designed for laptops) can be selected. The lower models are called Mali-G725 with 6-9 cores, Mali-G625 with 5 cores and Mali-G625 with less than 6-9 cores.
Official promotion data of Tianji 9400
Official Tianji 9400 increase:
【CPU】
Compared with the previous generation, the single core performance of the CPU is 35% higher, the multi-core performance is 28% higher, and the power consumption in the same performance is 40% lower;
An'tu rabbit has a room temperature of 2.8 million and a laboratory environment of more than 3 million.
【GPU】
GPU peak performance can be improved by 41%, GPU power consumption can be reduced by 44%, and ray tracing performance can be improved by 40%;
The third generation optical pursuit (30fps → 60fps → 90fps), The light tracking frame rate of Breaking through the Dark Zone is increased by 50%, while the power consumption is reduced by 10% (Note that it is at the same time);
Power consumption of King Glory reduced by 35%, "Miha Tour Three piece Set" The original power consumption of 60fps is reduced by 23%, the power consumption of 60fps satellite iron is reduced by 29%, and the zero power consumption of 60fps absolute zone is reduced by 35%.
[Baseband and connection]
The new 5G baseband supports R17, 5G power consumption reduced by 18% ；
WiFi/Bluetooth chip is upgraded from 6nm to 4nm, daily WiFi power consumption is reduced by 15% - 50%;
Declare that WiFi anti-interference ability is increased by 80% and transmission distance is increased by up to 30 meters (equivalent to two floors);
support WiFi 7 up to 7.3Gbps (Including 2x2 2.4G band+dual 5G band. One is 2x320M, and the other is 160M). Tri band WiFi supports MLO of 2.4G+5G, 2.4G+6G, 5G+5G, 5G+6G (6G band is not available in China), Hardware supports dual 5G MLO and dual 6G MLO (Fage said that he would release corresponding software according to ecological progress);
It supports dual Bluetooth connection, the latest "km" BLR Bluetooth protocol (Long Range), and Bluetooth audio of up to 12Mbps, 384kHz, 24bit.
X925 Super Core
The CPU part, of course, is the new X925 super core architecture. X925 is the new arm ⭐， It is the most advanced CPU architecture with the latest design features for mobile phones. In the official PPT of Arm ↓, X925 is the generation with the largest increase in IPC.
Tianji 9400 X925 super core of is Arm V9.2 instruction set, 10 width decoding unit, 6x128bit SMID/FP execution unit, SVE2/SME support (256 width)
Apple A18/A18 Pro The performance core of is Arm V9 instruction set, 10 width decoding unit, 6x128bit SMID/FP execution unit, and SVE/SME support (512 width)
Snapdragon 8 Gen 4 The self-developed performance core is the Arm V8.7 instruction set, 8-width decoding unit, 4x128bit SMID/FP execution unit, and the more "ancient" NEON (harm, the same origin as Snapdragon X Elite, and the disadvantage of "early design")
This year's Tianji 9400 and Apple A18/A18 Pro, last year's Tianji 9300 and Snapdragon 8 Gen 3 (the last two are both X4 super cores) belong to Arm V9, while Snapdragon 8 Gen 4 and Snapdragon X Elite belong to Arm V8.7.
The performance, energy efficiency and security of the Arm V9 will be better. The new V9 instruction set can bring 10% to 20% speed increase in the validation of MediaTek with Hongruan, Fasthand and other application manufacturers.
PS： Unlike the Arm V9, which completely removes 32-bit support, the Snapdragon 8 Gen 4, which theoretically uses the Arm V8.7, may be the only flagship chip that natively supports 32-bit apps (to be confirmed after actual measurement).
The front end decoding unit of the X925 is 10 wide like the X4 and Apple A18/A18 Pro 。
In contrast, the A17 Pro is a 9-width decoder, the AMD Zen 5 is a dual 4-width decoder, and the M1/M2/A15/A14 is 8-width. The X3 super core of Arm can be 6-8 width (the X4 a year ago, the core change is to increase the decoding width+shorten the pipeline length).
Arm's performance comparison between NEON and SVE 2 ↑
While NEON, SVE, and SVE 2 are three generations of progressive (single instruction and multiple data execution) instruction sets, SVE/SVE 2 has the advantage of supporting flexible vector length, while SVE 2 adds support for machine learning, DSP digital signal processing (5G and multimedia processing), which can improve the performance and energy efficiency of decryption, video decoding, machine vision, camera video streaming, games, baseband and other parts.
However, whether the advanced features of the Arm V9 and SVE 2 can be utilized depends on the follow-up of the application ecology. They are still in a future state (several years later, when the machines enter the middle and old age, their performance and power consumption gap may be widened). The addition of Apple this year should accelerate the promotion and popularization of the new instruction set.
Performance measurement:
CPU single core increased by 23%/multi-core increased by 15%, and GPU increased by more than 30%
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.
The engineering machine of MediaTek has been an old friend since Tianji 9000. It has a directly exposed card slot and an extremely advanced 3.5mm headphone connector.
This year's configuration is Tianji 9400, with 16GB LPDDR5X 10667Mbps imperial memory+1TB UFS 4.0 flash memory.
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.
During the run in test, the room temperature is between 21 ℃ and 23 ℃, and the body temperature is fixed between 21 ℃ and 22 ℃. Turn on the performance mode, and do not use fans, heat dissipation clips and other tools.
*If the online cooling clip/refrigerator scores, the multi-core score will be several hundred points higher.
SPEC 2017 cannot be run at the test site. This time, the test items are GeekBench 6, GFXBench and 3DMark, as well as the "Miha Tour Three Piece Set" of Yuanshen, Xingtie and Jujube Zero.
GeekBench of Tianji 9400 ran 6 points ↑
Comparison with other SoCs ↑
GeekBench 6 (two columns on the left) and GeekBench 5 (two columns on the right) are two sets of test systems.
[GeekBench 6] is a multi-core single load, which is closer to daily applications than GeekBench 5, and is more beneficial to Apple with fewer cores. 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.
In the previous [GeekBench 5] test, it can call local libraries, and the multi-core load is higher, so the performance of the Android camp will be better. and GeekBench 6.3 has introduced the SVE/SME support mentioned above, so Tianji 9400 chips that support SVE/SVE2 will be greatly improved . However, there are not many apps supporting SVE/SVE2 in the application ecosystem, so it is actually The achievements of GeekBench 5 may be closer to the current daily use experience.
(The second column is the ratio ↑ of "GeekBench 6 score and super core frequency" to compare the performance of different chips at the same frequency)
This is a generation of CPU single core performance that has soared collectively. The single core performance of Tianji 9400 has left the previous generation without any doubt, 23% stronger than Snapdragon 8 Gen 3, and 23.2% stronger than Tianji 9300.
The multi-core CPU performance of Tianji 9400 is 22.2% better than Snapdragon 8 Gen 3 and 15.4% better than Tianji 9300.
Tianji 9400 should be one of the flagship chips of the same generation, with the largest increase in the same frequency performance. This may be the core reason why Fage can achieve "40% lower power consumption than the previous generation under the same performance":
Calculated at 3.62GHz, Tianji 9400's same frequency performance in GeekBench 6 is 16.6% better than Tianji 9300, 18% better than Snapdragon 8 Gen 3, and 9% better than Snapdragon 8 Gen 4's leakage score;
If calculated according to the 3.78GHz recorded by GeekBench in the background (the frequency is 4.4% higher than 3.62GHz), the GeekBench 6 same frequency performance of Tianji 9400 is 11% better than Tianji 9300, 12% better than Snapdragon 8 Gen 3, and the leakage score of Snapdragon 8 Gen 4 is 4% better.
GFXBench Test ↑
3DMark Steel Nomad Light Test ↑
In the GPU part, the rendering pipeline and API used for GFXBench testing are closer to traditional mobile games, and are suitable for testing the performance of traditional GPUs.
The new Steel Nomad Light test launched by 3DMark has updated features and is closer to the rendering pipeline of modern 3A games, which is suitable for viewing the new flagship GPU. However, it has high requirements for mobile devices and requires 8GB of memory, which makes A16 unable to run this test.
GPU side, Tianji 9400 in GFXBench It increased by about 35% in the high pressure test.
stay 3DMark Steel Nomad Light directly slaughtered the list, and the score of 2524 was increased by nearly 30% , leaving behind about 2200 A18 Pro, less than 2000 Tianji 9300+and less than 1800 Snapdragon 8 Gen 3.
According to the existing explosives, This year's GPU of Snapdragon 8 Gen 4 may not win Tianji 9400, and Jige will probably continue to defend the GPU championship 。
Finally, let's look at the entertainment rabbit, The total score of Anthare is not reliable, but its single CPU and GPU test has certain reference value.
On the rabbit side, The CPU of Tianji 9400 increased by 22.9%, the GPU increased by 37.3%, and the memory increased by 35.7%:
Tianji 9400 engineering machine, 2.84 million points in total, 623000 CPU, 1.274 million GPU, 530000 memory, UX 410000 (Angu rabbit V10.2.6)
Tianji 9300 engineering machine, 2.13 million in total, 507000 CPU, GPU 928000, memory 389000, UX 312000 (Anthare V10.06-OB6)
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, 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 have a total score of 1.497 million, 368000 CPU, 526000 GPU, 264000 memory, 339000 UX (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
Ben Yanzu is more concerned about energy efficiency and temperature rise in the process of score running than Anthare: the same room temperature test is about 22 degrees. This year's temperature rise is 8 degrees (10% power off), and last year's temperature rise is 10 degrees (8% power off). But it is an engineering machine after all, and the battery capacity of the machine is unknown, so the data is only for reference.
According to the conservation of energy, electric energy is converted into heat energy in the near time, and the temperature rise and power failure should be positively correlated. However, the actual trend is opposite. It may be that the chips/cooling configurations of the two generations of engineering machines are different, or the batteries of the engineering machines are different
Game Measurement and Energy Efficiency Analysis: Miha Tour Three piece Set
The most concerned part of energy efficiency is that there is no mass production machine yet, so we can only conduct side testing through games for the time being.
The room temperature is 21 to 23 degrees, the screen brightness is 200 nits, the brightness does not decrease throughout the whole process, and the pressure is arranged from low to high:
Primordial god 15 minutes, 60fps maximum image quality+off dynamic blur, Xumi City running map, average frame rate 60.1 frames, average power consumption 4.36W, maximum temperature 37.1 degrees.
Stellate iron 30 minutes, 60fps, the highest picture quality, Xingcha Sea Center, Xuanye Avenue in the first 15 minutes+garden turn back in the next 15 minutes, the average frame rate is 59.8 frames, the average power consumption is 4.72W, and the maximum temperature of the fuselage is 39.5 degrees.
Absolute zone zero 30 minutes, the highest picture quality of 60fps, the path around the police station in the full high light square, the average frame rate is 59.9 frames, the average power consumption is 4.84W, and the highest body temperature is 39.7 degrees (the actual running time is 35 minutes, taking the middle 30 minutes).
Interestingly, CPU scheduling ↑. In the high load scenario of large games, Tianji 9400 regards X925 super core as the main computing core. This may be because the X925 is strong enough in performance and has energy efficiency advantages. In addition, focusing on the setting of a single core can reduce scheduling loss and significantly improve the use experience of many scenarios 。
[Energy efficiency] Tianji 9400 uses the TSMC N3E process at the same level as A18 Pro. MediaTek said that X4 and A720 have made secondary optimization for the 3nm process, and the L2 cache of X4 and A720 has doubled. In theory, large cache will increase the rest power consumption, but will reduce the power consumption of middle and high frequencies, thus improving the overall energy efficiency performance.
Results Tianji 9400 focused on the new X925 super core. It shows that the X925 is a more power saving or efficient solution at this time, which makes Ben Yanzu can't help but have high expectations for the medium and low frequency energy efficiency of the X925.
The United Development Division mentioned energy efficiency many times in the communication meeting. When asked why "the newly released A725 was not used", MediaTek said that the Tianji 9400 is "the optimal solution for energy efficiency" - their design goal is to achieve more than 30% light and heavy load power optimization. After the test, they finally chose to continue to use X4 and A720 with the X925.
According to the test, the single core CPU of Tianji 9400 is 23% stronger, the multi-core CPU is 15.4% stronger, and the GPU is improved by more than 30%.
The performance of 30% to 40% GPUs of Tianji has soared for three consecutive generations. In addition to the improvement of the two generations of CPUs (X4, X925, and the full core design to increase the proportion of large cores), the flagship chip of last year has killed all mobile games in terms of absolute performance (what actually restricts them is power consumption). The generation of Tianji 9400 has become what Fage called "the post full frame era", and it needs to pursue a higher dimension.
Fage believes that there are three directions in the post full frame era:
Reduce power consumption : Tianji 9400 is using X925 super core and larger GPU to run low frequency, so as to reduce power consumption.
Cross end porting PC masterpiece : In this part, the main progress is to improve the absolute performance and optical pursuit performance. The light tracking performance of Tianji 9400 has been improved by 50%, and the light tracking technology of PC 3A masterpieces like Black Myth: Wukong has been introduced (the OMM light tracking engine it carries claims that by adding pre-processing parts, it will filter the parts that do not need to be calculated, leaving the computing power to the edge of the object)
AI generation rendering+AI interaction (Coach interaction and NPC): Fage is now officially announcing the end side AI language coach of the first episode of Glory of the King and the AI NPC of Waste Land.
In the super score comparison of the test site, the upper part is the high resolution video, and the lower part is the super score video ↑ (the smaller the difference, the better)
Among the "star speed engine MAGT 3.0 (reduced touch/network latency), frame doubling 2.0 (the same source of inserted frames on TV), and the first star speed engine super division technology" released together, MAGT 3.0 is a power consumption and delay control technology that is updated every year; The last two are superframe and super division, which are the work contents of individual display chips.
Fage also had a prototype display on the scene, and the effect was certainly good, especially the superframe in the GPU. Compared with the "single chip", it will have natural advantages, so we can see how to use the mass production machine.
There is another special point this year, Fage focused on background scheduling for the first time - Tianji scheduling engine, and released its own APP startup and touch response data.
Google's native scheduling can't circle all background threads of an application, so we dare not boldly tilt computing power.
However, MediaTek said that its "VIP engine" has a more accurate frame method, and it does not need a manual white list, and can automatically judge+real-time perception. The calculation force for the main thread and the small thread is tilted more, and the CPU/SLC, WiFi Bluetooth and other components can be notified at the same time.
The official test is to run a background app that can eat up CPU when running Yuanshen. In the test, Tianji 9400 can still maintain 59fps, while Tianji 9300 is 29fsp, and Snapdragon 8 Gen 3 is 20fps.
The principle of the chip manufacturer's software scheduling sounds a bit like the "unfair scheduling" ↑ of vivo OriginOS. Rounding it up is also an effort to "tombstone mechanism".
MediaTek said that from the generation of Tianji 9300, it began to systematically study the startup and response speed of the APP under light load (in addition to fast startup, the speed of each startup should be stable and controllable to reduce the impact of occasional stuttering on the "sense of sophistication"), focusing on confirming that the implementation pipeline would not be interrupted:
In terms of APP startup, Tianji 9400 claimed that the response time was less than 100ms (135ms for iPhone 15 Pro Max, and almost all iPhones after iPhone 11), and the response error was less than 12ms (13ms for iPhone 15 Pro Max).
Sliding response<110ms (iPhone 15 Pro Max is<127ms, the mainstream is 150 to 170ms), sliding point dwell time<40ms (iPhone 15 Pro Max is within 50ms).
I don't know why, but I have a strong premonition that the relevant content will probably appear in the terminal conference later.
Congenital AI Holy Sacrament
At present, almost all head chip and mobile phone manufacturers are engaged in AI, but Qualcomm and Apple pay less attention to AI than MediaTek.
The NPU890 of Tianji 9400 claims Diffusion performance increased by 100%, LLM performance increased by 80%, and energy efficiency increased by 45% 。
Published by the United Development Section Tianji 9400 Zurich test (ETHZ AIBenchmark) scores more than 6700 points, a huge improvement compared with the previous generation : The score of Tianji 9300 tester is 3145 points, that of X100 Pro is 3410 points, that of Snapdragon 8 Gen 3 Find X7 Ultra is 3348 points, that of Snapdragon 8 Gen 2 for Galaxy S23 Ultra is about 2500 points, that of Snapdragon 8+is more than 1800 points, that of Tianji 9200 is about 1600 points, and that of Tianji 9000 is more than 1000 points.
In addition to AI performance, Fage's AI feature is also the most active. The development trend of AI is from camera → big language model → Wensheng map → Tusheng video. Tianji 9400 generation directly supports the end side Tusheng video:
Tianji 9300 is a mobile terminal The first large model equipped with a hardware generative AI engine, the first to support the expansion of generative AI end-to-end skills, the most supported models, hardware memory compression, and the ability to run 33 billion parameters SoC of;
And Tianji 9400 is First to support graphic animation (Stable Diffusion), Temporal Tensor has time domain tensor hardware instruction acceleration, DiT (Diffusion Transformer) at the first sending end First end LoRA training The first mobile SoC to run the end side hybrid expert MoE model (all running on NPU), and the mobile SoC to support the end side running the 1B and 3B models of LLAma 3.2 just released by Meta (11B or cloud);
MediaTek said that Tianji 9400 supports multi-mode 50Token/s (Google Tensor G4 is 45Token/s, A18 Pro is 30Token/s), and its speed is close to GPT 4o (you can understand pictures at the end, like directly reading English menus and giving suggestions);
The Token of the big language model has soared from 2K to 32K, an increase of 15 times;
This year, the hardware low bit wide KV cache is compressed, and the memory occupation is reduced by 50%. The 7.68GB memory required by the traditional 7B model is directly halved to 3.65GB;
MediaTek said that its cooperation with Xiaohongshu in the end side SDXL large map generation has doubled the speed, faster than the cloud solution. AI learning in cooperation with Hongruan can repair the pasted portrait with other pictures locally.
In addition to graphics video, another trend of AI is end side operation. After all, end-to-end training can keep records without privacy problems. MediaTek said that AI training was conducted with NPU under the condition of user sleeping+power connection.
(General understanding: AI is an advanced repeater that is extremely good at finding rules. Their process of learning with materials is called "training", and their output content with the model generated by training is called "reasoning").
Finally, in the PPT of the "AI Pioneer Plan" of MediaTek, in addition to the old friend Vivo OPPO、 Xiaomi and Chuanyin also showed glory. I wonder if you will see the glorious flagship of Tianji 9400?
summary
Now is indeed the "post full frame era" as MediaTek said. Whether it is CPU or GPU, medium and low frequency energy efficiency is the key, and the performance that can be actually used is yours.
The output performance has practical value only when the power consumption of the mobile phone is below 6W. No matter how high the score of the extreme refrigerator is, it is meaningless to reduce the frequency at normal temperature.
Fage still does not recommend using air cooling or heat dissipation to score points this year. It still leaves the name of the strongest score to ODM manufacturers: Tianji 9400 engineering machine's 2.84 million rabbit score is still significantly lower than the 2.97 million normal temperature score of mass production machine and the 3.03 million refrigerator score.
The performance of the Tianji 9400 and Snapdragon 8 Gen 4 generation has been significantly improved, but their power consumption at low and medium frequencies is still a mystery. In a word, this year, either super god or super ghost, the party is in a dilemma for those who want to get on the bus and those who want to get off the bus.
Semiconductor has an "impossible triangle" - the chip cannot meet the three requirements of "area (cost) - performance - energy efficiency" at the same time.
Fage's answer this year is to use money to ensure performance and energy efficiency: 29.1 billion transistors are stacked on the stack, making the larger X925 super core and GPU run at a more sweet frequency.
The Tianji 9300 last year, with its full core CPU design, has higher energy efficiency than the Snapdragon 8 Gen 3. The X925's Tianji 9400 claims that the peak performance of the GPU is 41% higher than that of the previous generation, and the power consumption is 44% lower. In the actual game, the X925 is the absolute main operation, which also makes us full of expectations for its energy efficiency. Let's wait for the vivo X200 series released on October 14 to reveal the answer.
Follow our Weibo @ Love Computer
Follow our WeChat official account: playphone
Of course, we also pay attention to our Bilibili account: love computer
