MediaTek Tianji 9300 AI’s performance explodes and produces high-quality pictures within 1 second.
Recently, MediaTek unveiled its new generation flagship 5G generative AI mobile chip Tianji 9300. The innovative all-core architecture design combined with the new generation AI processor APU and MediaTek’s unique cutting-edge technology, its powerful performance provided support for generative AI applications and created a vivid and substantial generative AI experience. At the same time, MediaTek shared resources with many AI companies in the industry, and jointly created an adequate AI ecosystem on the mobile side.
The brand-new seventh-generation AI processor APU 790 is born for generative AI.
With the increasing demand of users for generative AI applications, the advantages of end-side generative AI such as convenience and security are highlighted. Of course, deploying the end-side AI language model requires strong AI computing power support.
Tianji 9300 is equipped with MediaTek’s seventh generation AI processor APU 790, which is designed for generative AI. It has a hardware-level generative AI engine, which can realize higher-speed and safer edge AI calculation, and deeply adapt to Transformer model for operator acceleration, with a speed eight times higher than that of the previous generation.
At the same time, the performance and energy efficiency of APU 790 have been significantly improved, and the integer operation and floating-point operation capabilities have been improved to twice that of the previous generation. Zurich ETHZv5.1 AI-Benchmark Mobile Soc scored 2109 points, and the AI performance successfully dominated the list, and the power consumption was reduced by 45%. With the support of powerful AI performance, images can be generated within 1 second. Tianji 9300′ s powerful AI computing power, innovative all-core CPU architecture and Immortalis-G720 GPU all lay a solid performance foundation for end-to-end running of generative AI.
At the same time, based on the characteristics of the large language model with hundreds of millions of parameters, MediaTek has developed the mixed-precision INT4 quantization technology, which, combined with the memory hardware compression technology of MediaTek, can make more efficient use of memory bandwidth, greatly reduce the terminal memory occupied by the large AI model, break through the memory limit of mobile phones for running the large AI language model end-to-end, and help the larger parameter model land on the end-side.
Based on the above, Tianqi 9300 landed a 7 billion-parameter AI language model on the side of vivo flagship mobile phone for the first time, with a processing speed of 20 Tokens per second. Not only that, MediaTek has broken through the industry limit, and has successfully run a large language model with 13 billion parameters on the end side with vivo. Even, Tianji 9300 has taken the lead in successfully running the AI language model with 33 billion parameters on mobile chips, leading the industry.
Tianji 9300 also supports multi-modal generative AI model, creating rich and interesting end-to-end experiences such as "Wen Sheng Poetry", "Wen Sheng Picture" and "Wen Sheng Interesting Picture".
It can be seen that Tianqi 9300′ s AI computing power and end-to-end generative AI capability are ahead of the industry, which is enough for users to be full of AI creativity anytime and anywhere.
The end-side skill expansion of generative AI model brings comprehensive and rich end-side generative AI experience.
Different from the cloud-side generative AI solution, due to the difference of hardware environment, the deployment of end-side generative AI also needs to consider factors such as mobile phone memory, storage capacity, and upper load limit. Therefore, MediaTek took the lead in proposing advanced solutions.
APU 790 supports NeuroPilot Fusion, an end-side skill expansion technology of generative AI model. It can continuously carry out LoRA (Low-Rank Adaptation) fusion on the end-side based on the basic big model. With the empowerment of hybrid AI, it can complete the fusion of n functions on the end-side based on one basic big model through cloud training, giving the basic big model more comprehensive and richer generative AI application capabilities.
For example, based on the "drawing GIF animation" function of the AI model end-side skill expansion technology, users can change different styles or even expressions according to a photo, and play an expression pack with a personal style, and the second expression pack is small.
NeuroPilot, an AI development platform, accelerates the end-side generative AI ecological layout.
Based on powerful AI computing power, advanced memory hardware compression technology and AI model end-side skill expansion technology, the APU 790 of Tianji 9300 has raised the speed and breadth of end-side generative AI to a new level. At the same time, in order to accelerate the deployment and popularization of generative AI in the end-side, MediaTek has also built a rich AI ecosystem with its AI development platform NeuroPilot, from the underlying hardware to the tool chain, model center and development ecosystem, helping the ecosystem to deploy end-side generative AI applications quickly and efficiently.
NeuroPilot, the AI development platform, supports leading-edge mainstream AI models such as Android, Meta LIama 2, Baidu ERNIE Bot Model and Hundred Rivers Intelligent Hundred Rivers Model.
More importantly, NeuroPilot has a complete and advanced tool chain, including low-rank adaptive fusion of NeuroPilot Compression, Speculative Decoding acceleration and model optimization and transformation technology.
MediaTek Tianji Developer Center can also provide one-stop developer resources for end-side generative AI landing, and share end-side model deployment cases to improve development efficiency. At present, more than 20 generative AI partners have joined the ecological co-construction.
MediaTek also works with industry contract partners to create a wonderful generative AI application experience. ArcSoft’s generative AI super-resolution technology is based on the edge computing power of Tianji 9300 APU, and its performance can be improved by 30% compared with the previous generation. When shooting at a magnification of 25 times, using the generative AI super-resolution technology, you can shoot more realistic image effects.
The generative AI semantic search technology of Extreme Sense Technology is also based on the edge computing ability of Tianji 9300 APU. Compared with the previous generation, the performance can be improved by 260%. For example, searching for photos in the photo album of a mobile phone and describing the content of the photos can accurately find the corresponding photos in milliseconds. Moreover, it can also be searched when the network is disconnected, and privacy will not be revealed.
Morpho’s video call real-time digital avatar generation technology is also based on the edge computing ability of Tianji 9300 APU, and its performance is improved by 26% compared with the previous generation. The general virtual portrait generator needs to manually select the appearance style, which takes time. However, based on the video call real-time digital avatar generation technology, the user can operate easily, and the digital avatar can be generated instantly only by opening a single frame of the camera.
Based on the edge computing ability of Tianji 9300 APU, the performance of Hui carp’s generative AI anti-glare technology can be improved by 60%. With the help of this technology, whether outdoors or indoors, if there is glare when shooting, you can eliminate glare interference as long as you relax and darken it.
It can be seen that under the trend of AI-side cloud convergence, Tianqi 9300 has shown comprehensive advantages in AI computing power, generative AI user experience and ecology, setting a new benchmark for a new generation of flagship end-to-end generative AI experience, and powerful generative AI will use Tianqi.
In addition, MediaTek and other pioneers of generative AI are constantly carrying out technological innovation and ecological layout, vigorously promoting hybrid AI computing, providing a unique and efficient solution for end-to-end generative AI deployment, and devoting themselves to promoting the widespread application of generative AI at the end-side, so that more users can experience personalized end-to-end AI applications and build a new panoramic intelligent experience, thus enabling technology to better serve the public.