MediaTek Dimensity 9300 was released, the first all-large core architecture, and the 13 billion parameter large model ran locally
Smart stuff
Author | Jiahui
Edit | Yun Peng
Zhidong reported on November 6 that at seven o'clock tonight, MediaTek's new Dimensity flagship chip launch conference started on time.
At the press conference, MediaTek officially released a new generation of flagship 5G generative AI mobile chip Dimensity 9300. And according to MediaTek's official introduction, this chip is the world's first smartphone chip with a full nuclear architecture.
According to reports, Dimensity 9300 is 4 super large cores + 4 large core architecture, the CPU contains 4 Cortex-X4 super large cores, and 4 Cortex-A720 large cores with a main frequency of 2.0GHz, using TSMC's new generation 4nm process, with 22.7 billion transistors, 8MB L3 cache and 10MB system cache. Compared with the Dimensity 9200, the peak performance of the Dimensity 9300 is increased by 40%, and the energy consumption is reduced by 33% at the same performance.
In terms of GPU, the Dimensity 9300 uses a 12-core Immortalis-G720 MC12 1300MHz, which improves performance by 46% with the same energy consumption and reduces the energy consumption by 40% with the same performance compared to the Dimensity 9200.
1. Integrate the seventh-generation AI processor with a built-in generative AI engine
The Dimensity 9300 integrates MediaTek's 7th generation AI processor APU 790. The APU 790 has a built-in generative AI engine that adapts to the Transformer model for operator acceleration, with a processing speed 8 times that of the Dimensity 9200, and can generate images in less than 1 second.
Combined with NeuroPilot Compression, a memory hardware compression technology, the Dimensity 9300 supports running large AI models with 1 billion, 7 billion, and 13 billion parameters on the terminal. According to MediaTek officials, the Dimensity 9300 successfully runs a large AI model with 33 billion parameters.
The APU 790 also supports NeuroPilot Fusion, a generative AI model technology that continuously performs Low-Rank Adaptation (LoRA) fusion on the device side based on the basic large model to expand the capabilities of the basic large model.
MediaTek's AI development platform NeuroPilot supports mainstream AI models such as Android, Meta Llama 2, Baidu Wenxin Yiyan Model, and Baichuan Intelligent Baichuan Model, allowing developers to deploy multi-modal generative AI applications on the device side to achieve device-side generative AI innovation in text, images, music, and other aspects.
2. AI semantic segmentation video engine, frame-by-frame optimization, optical image stabilization
The Dimensity 9300's ISP image processor is the Imagiq 990, which supports AI semantic segmentation video engine, which can perform 16-layer image semantic segmentation, and optimize the captured image color, texture, noise and brightness in real time on a frame-by-frame basis, improving the brightness, sharpness and detail richness of video recording.
The Dimensity 9300 also supports dual engines of depth of field and flare, OIS optical image stabilization core, and all-pixel focus superposition of 2x lossless zoom, which improves the speed of image stabilization and film yield, and the effect of light and shadow to varying degrees.
3. Dynamic radio noise separation, AI depth of field image quality engine, audio-visual upgrade
In terms of mobile display, the Dimensity 9300 integrates MediaTek's MiraVision 990 mobile display processor, which supports 180Hz WQHD and 120Hz 4K display, as well as dual-screen display in the form of a folding screen. Equipped with an AI depth-of-field quality engine, it realizes real-time detection of major objects and background images.
At the same time, the Dimensity 9300 is equipped with MediaTek's second-generation hardware ray tracing engine, which supports 60FPS high-smooth ray tracing, and global lighting effects achieve console level. MediaTek's MAGT game adaptive control technology has also been upgraded to a "star speed engine".
In terms of audio noise reduction, the Dimensity 9300 supports 3 microphones for high dynamic recording noise reduction, has dynamic audio recording capabilities, and uses noise separation technology to filter environmental noise such as wind noise.
In addition, in terms of network communication and data security, the Dimensity 9300 is also different from the Dimensity 9200.
In terms of network communication, the Dimensity 9300 integrates a 5G modem, supports sub-6GHz quad-carrier aggregation (4CC-CA) and multi-standard dual-SIM dual-pass, and is equipped with MediaTek 5G UltraSave 3.0+ power-saving technology to reduce 5G communication power consumption. and integrated with AI to support situational awareness.
In terms of data security, the Dimensity 9300 is equipped with a new secure boot chip, an isolated secure computing environment, and Armv9 memory marker extensions, which can help developers avoid memory vulnerabilities and achieve a secure user use environment and development environment.
Conclusion: Improving performance and optimizing user experience is the key to AI chip scuffle
With the continuous development of artificial intelligence technology, more and more enterprises have begun to get involved in the field of AI chips, the AI chip market is constantly developing and innovating, and new chip architectures and technologies are also emerging, forming a fierce "AI chip melee".
Judging from the new flagship chip released by MediaTek, the Dimensity 9300 adopts an all-large core architecture design, whether it is the development platform that supports a variety of mainstream large models, or supports AI semantic segmentation video engine, AI depth of field image quality engine, Dimensity 9300 is inseparable from AI in all aspects of optimization, and it is also inseparable from the user experience upgrade brought by AI functions.
It can be seen that an excellent chip not only needs to have high performance, but also needs to make users feel simple to operate, easy to use, and feel the experience changes brought about by performance improvement in a variety of scenarios of daily use. In the race for AI chips, improving performance and optimizing the user experience are key.