Intel’s first generation Core Ultra is officially released, building a foundation for AI PC and capable of running large AI models without a network

When we use AI tools supported by large language models such as ChatGPT, Wenxinyiyan, and Tongyi Qianwen, there is a necessary condition: the Internet must be connected.

Because the computing of AI tools is in the cloud, not locally on the computer. However, from the perspective of upstream chip manufacturers and terminal manufacturers, large models in a networked environment are powerful, but they also need to be supplemented by large end-side models that can run without a network. The combination of end and cloud is the perfect solution.

So at Intel's first-generation Core Ultra official launch event on December 15, we saw that Alibaba Cloud Tongyi Qianwen's large model was successfully adapted to the Intel Core Ultra platform and can be deployed on PCs and other devices in the future.

Thanks to the first-generation Core Ultra platform (codenamed Meteor Lake), it has Intel's most significant processor architecture change in 40 years, namely a separated module architecture and the first integrated neural network processing unit (NPU), which is deployed on Intel Core Ultra platform devices The Tongyi Qianwen large model on the Internet can greatly shorten the model inference response time, with an average acceleration of up to 3 times.

In addition, Intel has also previously tested the good performance of the LLaMa2-7B large model from Meta in running without network on the device side. The concept of AI PC has become clear due to the gradual implementation of the large model on the device side.

Since the release of ChatGPT at the end of last year, AIGC applications supported by large language model technology have achieved great development. The main position of this AI wave is on PC, because PC is more suitable for most users in terms of interaction and presentation. AIGC applications, whether it is ChatGPT or MidJourney, need to be connected to the Internet. Once there is no network, we will be disconnected from AI.

Therefore, if a PC wants to truly evolve into an AI PC, not only AIGC applications can be easily run on it, but it itself also needs to have powerful AI processing capabilities.

Therefore, the news that Alibaba Cloud Tongyi Qianwen's large model Qwen-7B has successfully adapted to the Intel Core Ultra platform and can be deployed on PCs and other terminals in the future has two meanings, and a symbolic meaning.

The first level of meaning is, what does Intel rely on when it shouts the AI ​​PC slogan?

At the on Technology Innovation Conference held in Silicon Valley in September this year, Intel CEO Pat Gelsinger proposed the revolutionary concept of PC-AI PC.

If you want to efficiently perform AI calculations and run AI applications on a PC, the traditional CPU+GPU architecture is not the optimal solution. Therefore, on the first-generation Core Ultra platform, Intel innovatively adopted a separated module architecture, which is also the history of Intel The largest processor structural change in the world, the processor will be composed of four independent modules: computing module, SoC module, graphics module and IO module, and will be connected through Foveros 3D packaging technology:

Compute module (Compute Tile): adopts the latest generation of energy-efficiency core and performance core micro-architecture and enhanced functions. This module uses the new generation Intel 4 process technology and has achieved significant progress in energy consumption ratio.

SoC module (SoC Tile): Innovative low power island design (Low Power Island) integrates a neural network processing unit (NPU), bringing energy-efficient AI function performance to the PC, and is compatible with standardized program interfaces such as OpenVINO. Facilitate the development and popularization of AI applications. The new low-power energy efficiency core further optimizes the balance between energy saving and performance. The SoC module also integrates a memory controller, media codec processing and display unit, supporting 8K HDR and AV1 codecs as well as HDMI 2.1 and Display Port 2.1 standards. Wi-Fi and Bluetooth are also supported, including Wi-Fi 6E.

Graphics module (GPU Tile): This processor integrates the Intel Sharp graphics architecture, which can provide independent graphics card-level performance in the integrated graphics card and supports ray tracing and Intel XeSS. With a jump in graphics capabilities and improved energy efficiency, Meteor Lake delivers outstanding performance per watt.

IO module (IO Tile): Contains superior connectivity, integrating Thunderbolt 4 and PCIe Gen 5.0.

Taking the Core Ultra 7 165H processor (TDP 28W) as an example, Intel said its multi-threaded performance is more than three times that of Qualcomm Snapdragon 8cx Gen 3 and 1.11 times that of its direct competitor AMD 7840U.

At the same time, Intel said that with better multi-threading performance, the energy efficiency of the Core Ultra 7 165H in various scenarios is also better than that of the AMD 7840U.

In addition, the low-power island is a very interesting design. It has 2 LPE-Cores ultra-low-power cores, which can run tasks that do not require response speed, such as online playback, etc., further reducing power consumption. What makes this design interesting is that these two ultra-low-power cores are located in the SoC module, not the computing module.

Of course, the most important and significant update is that the new NPU module will be mainly responsible for AI computing. Because it is a special computing unit for artificial intelligence acceleration, the advantage of NPU models in AI computing is high performance and low power consumption. , In addition, low-latency and high-response CPUs and high-performance and high-throughput GPUs will also bear the demand for AI computing power. The cooperation of the three forms a powerful AI computing power on the terminal.

Still compared to the AMD 7840U processor next door, Intel said that the overall AI computing power of the Core Ultra 7 165H not only beats its previous generation product, the Core i7-1370P, but also crushes the AMD 7840U.

To give an example of practical application, Intel and ByteDance's video editing tool Jianying worked together to optimize the "smart keying" function that users frequently use, and switched the computing requirements of this function to the first-generation Core Ultra products. After NPU, we found that processing video material not only takes less time, but also reduces power consumption. At this time, the free CPU and GPU computing power can be used for other editing needs.

Therefore, the so-called AI PC can not only run AI functions better, but also improve efficiency and reduce power consumption.

The second meaning is, what can a large model like Tongyi Qianwen do?

Now we can perform a variety of AI operations on PC, including but not limited to Wenshengwen, Wenshengtu, Wenshengmusic, as well as AI cutout, gesture recognition, motion capture, and in more detail, AI dubbing and voice changing, and video conferencing. Eye correction, etc.

Some of them are brand-new experiences based on large language models, and some have been implemented on PCs for a long time. For example, the 10th generation Core was the first to introduce AI performance support, which can bring automatic image enhancement and video resolution enhancement to PCs. , intelligent noise reduction and other AI functions.

Large language model technology is considered to be the road to general artificial intelligence (AGI, Artificial General Intelligence). AGI refers to an artificial intelligence system with a wide range of intelligence levels and abilities similar to humans. In comparison, currently large language models Most AI systems focus on a specific task or domain.

Under this premise, Alibaba Cloud's newly open sourced Tongyi Qianwen 72 billion parameter model (Qwen-72B) can be called the most powerful open source large model at present. Qwen-72B has achieved the best results among open source models in 10 authoritative benchmark evaluations. The performance exceeds the open source benchmark Llama 2-70B and most commercial closed source models, and it already has certain general capabilities.

Based on Alibaba Cloud's open source Qwen-72B, Qwen-14B, Qwen-7B, and Qwen-1.8B four large language models, as well as the large visual understanding model Qwen-VL and the large audio understanding model Qwen-Audio, Alibaba Cloud has achieved "full Size, full modality" open source.

This means that for developers from all walks of life, these open source large models can have a variety of packages to choose from, including PCs, mobile phones, and even IoT devices that are not familiar with the devices.

Tongyi Qianwen's large model is deployed on the Intel Core Ultra platform, and it is one of the possibilities that it can be deployed on terminals such as PCs in the future.

In addition, the XD Lab team of East China University of Science and Technology has developed a number of large industry models based on the Tongyi Qianwen open source models Qwen-7B and Qwen-14B, including the large mental health model MindChat (Man Tan), the large medical health model Sunsimiao (Sun Simiao), and the education model /Exam large model GradChat (Koi) and so on.

Zhejiang University and Higher Education Press developed the Zhihai-Sanle education vertical model based on Qwen-7B. It has been applied in 12 universities across the country and can provide intelligent question and answer, test question generation, learning navigation, teaching evaluation and other capabilities. The model has been used in Alibaba The Yunlingji platform provides external services, which can be called with one line of code.

Breaking away from the bit world, in the real world, Zhejiang Youlu Robot Technology Co., Ltd. has integrated Qwen-7B into the road cleaning robot, allowing the robot to interact with users in real time in natural language, understand the needs put forward by users, and carry out the user's high-level instructions. Analyze and dismantle, do high-level logical analysis and task planning, and complete cleaning tasks.

In short, the stage of AI is not just about PCs, and PCs are not just about AI applications, but they are beginning to have huge intersections, creating scenarios that will change our PC usage habits.

When drawing, should you use a digital tablet and a pressure-sensitive pen to create, or should you use a Vincentian drawing?

When writing, do you use a keyboard to type out words one by one, or do you use Chinese to write?

These problems have already emerged, and every user with an AI PC will gradually face changes in choices and usage habits.

In the plastic greenhouse of fate, every cabbage that has been sprayed with too many pesticides once had a dream of becoming a pollution-free organic vegetable.

# Welcome to follow the official WeChat public account of aifaner: aifaner (WeChat ID: ifanr). More exciting content will be provided to you as soon as possible.

Ai Faner | Original link · View comments · Sina Weibo