search
HomeTechnology peripheralsAIDimensity 9300, the first generative AI mobile chip: capable of running large models with 33 billion parameters

AI drawing is generated in one second, and the large language model is 20 tokens per second.

2023 is the first year of generative AI, and the mobile devices in our hands are also accelerating into the era of large models.

On the evening of November 6th, MediaTek officially released the annual flagship SoC Dimensity 9300. This chip has a 4+4 full-core design and completely surpasses Android in terms of performance and energy consumption. Competing with Apple.

What’s more noteworthy is that this year’s flagship chip is defined as a 5G generative AI mobile chip, which provides outstanding features of high intelligence, high performance, high energy efficiency, and low power consumption that far exceed those of the past.

Dimensity 9300, the first generative AI mobile chip: capable of running large models with 33 billion parameters

Dimensity 9300 uses TSMC’s new generation 4nm process and has 22.7 billion transistors. In order to meet the computing power needs of the generative AI era, the Dimensity 9300 pioneered the use of the "all-large-core" CPU architecture, including 4 Cortex-X4 ultra-large cores with a maximum frequency of 3.25GHz and 4 main frequency The 2.0GHz Cortex-A720 large core has a peak performance that is 40% higher than the previous generation, and power consumption is saved by 33% with the same performance.

Dimensity 9300, the first generative AI mobile chip: capable of running large models with 33 billion parameters

This architecture ensures fast working speed and high efficiency while also having power-saving characteristics. It can reduce power consumption in both light-load and heavy-load application scenarios. consumption and extend battery life. MediaTek said that the Dimensity 9300 has been optimized for common task requirements such as video, live broadcast, and games, and is more adaptable to the multi-tasking mode of folding screen mobile phones than before.

With the improvement of chip manufacturing processes, transistors are continuously miniaturized, and various leakage problems have become a major obstacle to the development of Moore's Law. Leakage means a significant increase in energy consumption, and the chip will also face problems of overheating or even failure. In this case, the power consumption gap between the small core and the large core has become smaller and smaller.

According to reports, MediaTek began exploring the form of full-core chips as early as three years ago. Now, by letting the big cores process tasks quickly and then sleep for longer periods of time, we can make big-core processors counterintuitively more power-efficient than their smaller-core counterparts. On the other hand, MediaTek has also added an out-of-order strategy to further increase the efficiency of application execution. Dimensity 9300, the first generative AI mobile chip: capable of running large models with 33 billion parametersMediaTek believes that by next year, the design of all large cores will become the consensus in the industry.

In addition to conventional capabilities, MediaTek has focused on upgrading the AI ​​performance of the chip this time. Dimensity 9300 integrates MediaTek's seventh-generation AI processor APU 790, which is specially designed for generative AI. Its performance and energy efficiency have been significantly improved compared to the previous generation. The performance of integer operations and floating point operations is twice that of the previous generation. Power consumption is reduced by 45%.

Dimensity 9300, the first generative AI mobile chip: capable of running large models with 33 billion parametersAPU 790 has a built-in hardware-level generative AI engine, which can achieve faster and safer edge AI calculations. Compared with the previous generation, it specifically performs operator acceleration for Transformers commonly used in large language models. , the processing speed of large models is 8 times that of the previous generation.

The main core network architecture of contemporary popular large language models (LLM) is mostly composed of transformer blocks. Compared with the common CNN network in computer vision, the transformer network uses the Softmax LayerNorm operator as the core and has fewer convolution operators, so the acceleration mechanism of the original AI core is not applicable. On the seventh-generation APU processor, MediaTek focuses on optimizing the Softmax LayerNorm operator to increase computing power.

Quantification is currently one of the effective ways to optimize AI reasoning. Based on the characteristics of large language models with hundreds of millions of parameters, MediaTek has developed mixed-precision INT4 quantization technology, combined with its unique memory hardware compression technology NeuroPilot Compression, which can more efficiently utilize memory bandwidth and significantly reduce the occupation of terminal memory by large AI models.

MediaTek engineers said that although large models can bring better productivity, running 13B locally means that it takes up about 13GB of memory, plus Android’s own 4GB, and other APP’s 6GB has exceeded the 16G memory capacity of most mobile phones. The memory hardware compression technology used by Dimensity 9300 reduces the memory usage of large models to 5GB through quantization and compression. Only in this way can most users afford to run large model applications in daily life.

APU 790 also supports the generative AI model end-side "skill expansion" technology NeuroPilot Fusion, which can continuously perform low-rank adaptation (LoRA, Low-Rank Adaptation) fusion on the end-side based on the basic large model, thereby empowering The basic large model has more comprehensive capabilities.

Based on such hardware and optimization, Dimensity 9300 can reach 2019 points on the latest version of the AI ​​Benchmark proposed by ETH Zurich, which is a new high for mobile chips.

Dimensity 9300, the first generative AI mobile chip: capable of running large models with 33 billion parameters

Using Dimensity 9300, end-side AI image generation such as Stable Diffusion can produce images within one second, and end-side inference of a 7 billion parameter large language model can be achieved per second 20token.

MediaTek said that in its cooperation with vivo, based on Dimensity 9300, it has taken the lead in realizing the inference of 7B and 13B large models on the mobile terminal. It is expected that such products will be launched on the terminal soon. In addition, in more extreme cases, MediaTek has also run through large models up to 33B.

In the communication meeting and on-site before the release, MediaTek demonstrated the ability to use the Dimensity 9300 engineering machine to realize LoRA text diagrams and large model text generation.

We can foresee that on the latest generation of flagship mobile phones, we can use smarter intelligent assistants to quickly reply to chats and right-click based on suggestions given by large models, and use AI-generated emoticons to fight pictures...

Dimensity 9300, the first generative AI mobile chip: capable of running large models with 33 billion parameters

What will be launched soon and everyone can experience it is the AI ​​assistant Lan Xin Xiao V in the OriginOS 4 system on vivo X100 series mobile phones.

Dimensity 9300, the first generative AI mobile chip: capable of running large models with 33 billion parameters

At the beginning of this month, vivo just introduced the blue heart model and its applications at the developer conference. This series of capabilities is obviously for mobile phones. . With the blessing of large models, Lanxin Xiao V has industry-leading wisdom. It can receive information for processing through voice, text, file drag and drop, etc. If you encounter simple questions, Little V will reply with text or pictures. For complex questions, it can also output answers in the form of a SWOT model or mind map.

In addition, the AI ​​capabilities of Dimensity 9300 also cover everything from search to shooting.

Dimensity 9300, the first generative AI mobile chip: capable of running large models with 33 billion parameters

MediaTek said that its AI development platform NeuroPilot has deeply adapted and implemented cutting-edge mainstream AI large models such as Android, LIama 2, Wen Xinyiyan, and Baichuan large models. Optimization, more applications are on the way.

Dimensity 9300, the first generative AI mobile chip: capable of running large models with 33 billion parameters

In terms of GPU, Dimensity 9300 uses the 12-core Immortalis-G720. Compared with the previous generation, the peak performance is increased by 46%, and the power consumption can be saved by 40% under the same performance. Dimensity 9300 is equipped with MediaTek’s second-generation hardware ray tracing engine, which supports 60FPS high-smooth ray tracing and brings game console-level global illumination effects.

Dimensity 9300, the first generative AI mobile chip: capable of running large models with 33 billion parameters

In addition, MediaTek's unique MAGT game adaptive control technology has been upgraded to "Star Speed ​​Engine", which will not only cooperate extensively with game applications, but also expand the ecosystem of more types of applications cooperate. At the event, MediaTek announced that the Dimensity 9300 platform will launch the ray-tracing version of "World of Sword and Sword" and will also achieve a stable 60-frame ray-tracing experience on the mobile game "Dark Zone Breakout".

Currently, according to many third-party reviews, the Dimensity 9300 engineering machine’s CPU and GPU absolute performance and power consumption ratio exceed those of Snapdragon 8Gen3 and A17Pro. It seems that the strategy of using all large cores is successful.

Dimensity 9300, the first generative AI mobile chip: capable of running large models with 33 billion parameters##                                                                                                                                                                                                                                      in in in in in,                 in IN IN IN OTHER APPROVED   · As for Imagiq 990, it supports AI semantic segmentation video engine, 16-layer image semantic segmentation, dual depth-of-field and light spot engines, full-pixel focus overlay 2x lossless zoom, OIS optical image stabilization core, 3-microphone high-dynamic recording noise reduction, and can filter 25km /h Wind noise above 99% of wind speed. The Dimensity 9300 also comes with a new secure boot chip, an isolated secure computing environment and Armv9’s memory marking extensions to help developers avoid memory exploits. In terms of network, Dimensity 9300 integrates a 5G modem that supports Sub-6GHz four-carrier aggregation (4CC-CA) and multi-standard dual-card dual-pass. It also improves the signal through AI algorithms and supports 5G situational awareness functions. Dimensity 9300 supports Wi-Fi 7 and 5G sub-6GHz frequency bands, with a downlink rate of 7Gbps. In terms of Bluetooth connection, Dimensity 9300 supports 3 Bluetooth antennas and unique dual-channel Bluetooth flash connection technology, which can bring an ultra-low latency Bluetooth audio experience.

Dimensity 9300, the first generative AI mobile chip: capable of running large models with 33 billion parameters

According to reports, the first mobile phones using Dimensity 9300 chip include vivo, OPPO, Xiaomi, Transsion, etc. After MediaTek’s launch event, vivo has announced that it will be the first to carry a new flagship chip in the X100 series released on November 13, and will be the first to implement LPDDR5T-9600 memory.

Dimensity 9300, the first generative AI mobile chip: capable of running large models with 33 billion parameters

We look forward to the advent of a new generation of products.

The above is the detailed content of Dimensity 9300, the first generative AI mobile chip: capable of running large models with 33 billion parameters. For more information, please follow other related articles on the PHP Chinese website!

Statement
This article is reproduced at:机器之心. If there is any infringement, please contact admin@php.cn delete
DSA如何弯道超车NVIDIA GPU?DSA如何弯道超车NVIDIA GPU?Sep 20, 2023 pm 06:09 PM

你可能听过以下犀利的观点:1.跟着NVIDIA的技术路线,可能永远也追不上NVIDIA的脚步。2.DSA或许有机会追赶上NVIDIA,但目前的状况是DSA濒临消亡,看不到任何希望另一方面,我们都知道现在大模型正处于风口位置,业界很多人想做大模型芯片,也有很多人想投大模型芯片。但是,大模型芯片的设计关键在哪,大带宽大内存的重要性好像大家都知道,但做出来的芯片跟NVIDIA相比,又有何不同?带着问题,本文尝试给大家一点启发。纯粹以观点为主的文章往往显得形式主义,我们可以通过一个架构的例子来说明Sam

阿里云通义千问14B模型开源!性能超越Llama2等同等尺寸模型阿里云通义千问14B模型开源!性能超越Llama2等同等尺寸模型Sep 25, 2023 pm 10:25 PM

2021年9月25日,阿里云发布了开源项目通义千问140亿参数模型Qwen-14B以及其对话模型Qwen-14B-Chat,并且可以免费商用。Qwen-14B在多个权威评测中表现出色,超过了同等规模的模型,甚至有些指标接近Llama2-70B。此前,阿里云还开源了70亿参数模型Qwen-7B,仅一个多月的时间下载量就突破了100万,成为开源社区的热门项目Qwen-14B是一款支持多种语言的高性能开源模型,相比同类模型使用了更多的高质量数据,整体训练数据超过3万亿Token,使得模型具备更强大的推

ICCV 2023揭晓:ControlNet、SAM等热门论文斩获奖项ICCV 2023揭晓:ControlNet、SAM等热门论文斩获奖项Oct 04, 2023 pm 09:37 PM

在法国巴黎举行了国际计算机视觉大会ICCV(InternationalConferenceonComputerVision)本周开幕作为全球计算机视觉领域顶级的学术会议,ICCV每两年召开一次。ICCV的热度一直以来都与CVPR不相上下,屡创新高在今天的开幕式上,ICCV官方公布了今年的论文数据:本届ICCV共有8068篇投稿,其中有2160篇被接收,录用率为26.8%,略高于上一届ICCV2021的录用率25.9%在论文主题方面,官方也公布了相关数据:多视角和传感器的3D技术热度最高在今天的开

复旦大学团队发布中文智慧法律系统DISC-LawLLM,构建司法评测基准,开源30万微调数据复旦大学团队发布中文智慧法律系统DISC-LawLLM,构建司法评测基准,开源30万微调数据Sep 29, 2023 pm 01:17 PM

随着智慧司法的兴起,智能化方法驱动的智能法律系统有望惠及不同群体。例如,为法律专业人员减轻文书工作,为普通民众提供法律咨询服务,为法学学生提供学习和考试辅导。由于法律知识的独特性和司法任务的多样性,此前的智慧司法研究方面主要着眼于为特定任务设计自动化算法,难以满足对司法领域提供支撑性服务的需求,离应用落地有不小的距离。而大型语言模型(LLMs)在不同的传统任务上展示出强大的能力,为智能法律系统的进一步发展带来希望。近日,复旦大学数据智能与社会计算实验室(FudanDISC)发布大语言模型驱动的中

百度文心一言全面向全社会开放,率先迈出重要一步百度文心一言全面向全社会开放,率先迈出重要一步Aug 31, 2023 pm 01:33 PM

8月31日,文心一言首次向全社会全面开放。用户可以在应用商店下载“文心一言APP”或登录“文心一言官网”(https://yiyan.baidu.com)进行体验据报道,百度计划推出一系列经过全新重构的AI原生应用,以便让用户充分体验生成式AI的理解、生成、逻辑和记忆等四大核心能力今年3月16日,文心一言开启邀测。作为全球大厂中首个发布的生成式AI产品,文心一言的基础模型文心大模型早在2019年就在国内率先发布,近期升级的文心大模型3.5也持续在十余个国内外权威测评中位居第一。李彦宏表示,当文心

AI技术在蚂蚁集团保险业务中的应用:革新保险服务,带来全新体验AI技术在蚂蚁集团保险业务中的应用:革新保险服务,带来全新体验Sep 20, 2023 pm 10:45 PM

保险行业对于社会民生和国民经济的重要性不言而喻。作为风险管理工具,保险为人民群众提供保障和福利,推动经济的稳定和可持续发展。在新的时代背景下,保险行业面临着新的机遇和挑战,需要不断创新和转型,以适应社会需求的变化和经济结构的调整近年来,中国的保险科技蓬勃发展。通过创新的商业模式和先进的技术手段,积极推动保险行业实现数字化和智能化转型。保险科技的目标是提升保险服务的便利性、个性化和智能化水平,以前所未有的速度改变传统保险业的面貌。这一发展趋势为保险行业注入了新的活力,使保险产品更贴近人民群众的实际

致敬TempleOS,有开发者创建了启动Llama 2的操作系统,网友:8G内存老电脑就能跑致敬TempleOS,有开发者创建了启动Llama 2的操作系统,网友:8G内存老电脑就能跑Oct 07, 2023 pm 10:09 PM

不得不说,Llama2的「二创」项目越来越硬核、有趣了。自Meta发布开源大模型Llama2以来,围绕着该模型的「二创」项目便多了起来。此前7月,特斯拉前AI总监、重回OpenAI的AndrejKarpathy利用周末时间,做了一个关于Llama2的有趣项目llama2.c,让用户在PyTorch中训练一个babyLlama2模型,然后使用近500行纯C、无任何依赖性的文件进行推理。今天,在Karpathyllama2.c项目的基础上,又有开发者创建了一个启动Llama2的演示操作系统,以及一个

快手黑科技“子弹时间”赋能亚运转播,打造智慧观赛新体验快手黑科技“子弹时间”赋能亚运转播,打造智慧观赛新体验Oct 11, 2023 am 11:21 AM

杭州第19届亚运会不仅是国际顶级体育盛会,更是一场精彩绝伦的中国科技盛宴。本届亚运会中,快手StreamLake与杭州电信深度合作,联合打造智慧观赛新体验,在击剑赛事的转播中,全面应用了快手StreamLake六自由度技术,其中“子弹时间”也是首次应用于击剑项目国际顶级赛事。中国电信杭州分公司智能亚运专班组长芮杰表示,依托快手StreamLake自研的4K3D虚拟运镜视频技术和中国电信5G/全光网,通过赛场内部署的4K专业摄像机阵列实时采集的高清竞赛视频,

See all articles

Hot AI Tools

Undresser.AI Undress

Undresser.AI Undress

AI-powered app for creating realistic nude photos

AI Clothes Remover

AI Clothes Remover

Online AI tool for removing clothes from photos.

Undress AI Tool

Undress AI Tool

Undress images for free

Clothoff.io

Clothoff.io

AI clothes remover

AI Hentai Generator

AI Hentai Generator

Generate AI Hentai for free.

Hot Tools

SublimeText3 Mac version

SublimeText3 Mac version

God-level code editing software (SublimeText3)

Dreamweaver Mac version

Dreamweaver Mac version

Visual web development tools

SublimeText3 Chinese version

SublimeText3 Chinese version

Chinese version, very easy to use

VSCode Windows 64-bit Download

VSCode Windows 64-bit Download

A free and powerful IDE editor launched by Microsoft

SublimeText3 Linux new version

SublimeText3 Linux new version

SublimeText3 Linux latest version