1. Performance upgrade
The MediaTek Dimensity 9300+, which was just released in the morning, if you follow the routine of previous years, then this is basically the overclocked version of the Dimensity 9300, but this time the situation is different.
It turns out that this MediaTek update focuses on AI, but as is customary, let's take a look at what upgrades are available in terms of performance.
First of all, the CPU super-core frequency has been increased from 3.25GHz to 3.4GHz, and the theoretical single-core performance will be increased by about 5%;
2. AI highlights
As before, the little performance upgrade obtained by overclocking has no bright spots, but this time MediaTek has followed the general trend to fully explore the field of AI, and has achieved the following series of achievements:
1. This time, the "speculative decoding acceleration" technology on the device side has been realized, thus achieving a breakthrough improvement in inference performance - the generation speed has been greatly increased by 120%!
2. This is the industry's first chip to achieve higher-speed Llama2 7B (this is a large language model launched by Meta) to run on the device side, with a running speed of up to 22Tokens/second!
3. This is also the industry's first chip to support generative AI device-side "dual LoRA fusion", which is an evolution of the Dimensity 9300's industry-first LoRA Fusion technology - this is the Dimensity AILoRA Fusion 2.0 technology.
Among them, LoRA is the abbreviation of Low-Rank Adaptation, which is a low-rank-based adapter algorithm that can be used for efficient fine-tuning on large language models, and with the empowerment of hybrid AI, it can also complete the integration of N functions on the device side.
Functions such as "Picture Generating GIFs" are examples of this in action.
This time, due to the introduction of "dual LoRA fusion", the actual generation efficiency has not only been doubled, but also directly reduced the memory space requirement in half! This is the source of confidence in "shaping the future application scalability".
4. It supports ExecuTorch, a cutting-edge mainstream AI framework, to accelerate the development of device-side generative AI applications, and also launches a new Dimensity AI development kit to provide a one-stop solution for terminal generative AI application development, as follows:
It includes four modules: GenAI Model Hub, which covers the world's mainstream large models, fast and efficient GenAI best practices, GenAI optimization technology that efficiently improves performance, and Neuron Studio's one-stop visual development environment.
In other words, based on the AI technologies developed by the previous Dimensity 9300 and the Dimensity 9300+, the full-stack greatly optimized Dimensity AI system has been officially built, whether it is at the platform level for developers or at the terminal level for consumers!
3. AI applications
MediaTek has now cooperated with large models such as Google, Meta, Baidu, and Baichuan Intelligence, while Dimensity chips will support large models such as Alibaba Cloud's Tongyi Qianwen and Zero One Everything.
In addition, MediaTek has also pulled in many partners, including Blue Factory, Green Factory, Tencent, and Alibaba, to jointly release the "Generative AI Mobile Phone Industry White Paper".
The goal is to jointly explore the path of device-side AI technology by leveraging large-scale, pre-trained generative AI models to achieve multimodal generation, contextual awareness, and further enhance human-like capabilities.
Since AI has made such great progress, it is natural that AI is now indispensable to focus on optimizing the game aspect, so this time we have launched a new and upgraded Star Speed Engine technology.
The new Star Speed Engine integrates the adaptive technology software and hardware-level ray tracing technology of the previous Dimensity 9300, combined with the newly upgraded powerful AI computing power, to develop key technologies such as accurate system monitoring and network quality monitoring system.
In addition, by introducing AI technology into game interaction and game content generation, developers can be assisted in implementing the application of large language models in games.
In practical applications, based on the accurate prediction of the Star Speed Engine network, its data usage can be greatly reduced by 25%, and the power consumption of full frame can be reduced by 10%.
In addition, in the "head chicken game", it can also achieve 90 frames of HDR image quality, which is an astonishing 20% reduction compared to Snapdragon 8 Gen3!
In the 60-frame fine image quality scene of "Against the Cold", after turning on the Star Engine Adaptive Technology, the full frame power consumption can be reduced by 10% compared with the non-enabled state.
END
Committed to the exploration and collation of mobile phone knowledge