Shulou(Shulou.com)12/24 Report--

Since the beginning of this year, the hot circle of chatGPT has set off a new wave of global artificial intelligence for generative AI. AI is becoming a new driving force for the transformation of thousands of industries, as well as a common bet on the future of major technology companies.

For example, AMD recently launched the data center AI chip AMD Instinct MI300X GPU in the Advancing AI event, as well as the MI300A accelerated processing unit APU which combines the latest AMD CDNA 3 architecture and "Zen 4" CPU, and so on, which has attracted wide attention.

In the exploration of "AI changing the world", there is actually a company that has launched its layout early, that is, Intel. In 2018, Intel proposed to introduce AI on PC and launched the "AI on PC Developer Program" AI PC developer program. Since then, Intel has continued to integrate AI capabilities into its core processor products. Since the 10th generation of Core-X, Intel has added AI and deep learning-related acceleration instructions to its CPU, including improving the performance of AI at the architectural level, building Intel GNA into SoC to accelerate the application of low-power AI on PC, and introducing AI acceleration units into GPU of Xe and ARC architecture.

Intel's years of exploration will also usher in a centralized release in the near future. On December 15, Intel will officially release the Core Ultra processor based on the new Meteor Lake architecture in China. Among the Meteor Lake processors, Intel's most important measure is to introduce Al into the client PC and integrate independent NPU units in the Meteor Lake processor architecture, bringing independent low-power AI acceleration capabilities.

As for the integrated NPU unit added to Meteor Lake, it realizes more efficient AI computing and includes two neural computing engines, which can better support contents including generative AI, computer vision, image enhancement and collaborative AI. Moreover, this NPU is not a single island architecture. Apart from NPU,CPU and GPU, it can also perform AI operations. Different AI units will be used to deal with it in different scenarios and coordinate with each other. As a result, its overall energy consumption can be up to 8 times higher than that of the previous generation.

When the basic version of generative AI is defined as the era of AI 2.0, Intel has also made a lot of efforts to make AIGC run better on the local side of PC.

In our traditional understanding, to run a large language model like ChatGPT, there must be large graphics memory support, such as the Instinct MI300X GPU launched by AMD, which we mentioned earlier, but this is indeed a little far away from the majority of consumers, and Intel built the BigDL-LLM library in order to make the consumer-oriented 12th and 13th generation Kerry platform also run various large language models smoothly and provide a smooth experience. This library is specially designed for low-bit quantization of Intel hardware, and supports a variety of low-bit data accuracy, such as INT3, INT4, INT5, INT8, etc., with better performance and less memory consumption.

Through this library, Intel optimizes and supports a variety of large language models, including some open source models that can run locally. The library can even run a large language model with up to 16 billion parameters on an Intel lightweight machine with 16GB memory. In addition, many large language models such as LLaMA / LLaMA2, ChatGLM / ChatGLM2 and so on are supported.

Not to mention the upcoming Core Ultra series, a variety of Intel client chips represented by the 12th and 13th generation Intel Core processors and Intel Ruixian A-Series graphics cards can provide strong performance to meet the high-computing needs of generative AI. CTOnews.com has also done a practical test on this.

In the test, the editor chose a lightweight version certified by Intel's Evo platform: Asustek dawning Air, which carries Intel 13th-generation Core i7-1355U processor and 16GB LPDDR5 memory.

The editor installed Intel's big language model Demo on this Asustek dawning Air. This Demo integrates three major language models, including ChatGLM2, LLaMA2, and StarCoder. They are all optimized through Intel's corpus.

During the test, the editor first asked the big model Demo to help me make an introduction to the company's annual meeting in the story creation mode, which quickly presented a complete and decent opening copy, and the only First Latency of the whole process was 1249.8ms. If you are thinking and editing by yourself, it will take a long time to use the AI model on PC and it will be done in minutes.

When writing the copy in the big language model, the editor took a look at the scheduling of Asustek dawn Air performance resources. The 13th generation Core i7-1355U processor occupies 100%, the memory footprint reaches 9.7GB (62%), and the Xe core occupies 39%. It seems that the operation is indeed carried out locally. With the continuous optimization of Intel and the improvement of the computing power of the 13th generation Cooley processor, it is indeed possible to achieve the landing of AIGC on a thin and lightweight book.

Then the editor tests another question, asking it to extract the core information of a news article, and it can also quickly and accurately give the news content to the "summary". This is very useful for us to query information, organize reports, etc., and can greatly improve the efficiency of our work.

Finally, the editor asked the big model to help him write a syllabus of Zhu Ziqing's "back", which also quickly listed a set of logical and complete syllabus with detailed content. For those who need to refine and write an outline, such as teachers, AI can be used to assist teaching even if there is no network.

In addition to CPU, Intel also pays great attention to the optimization of GPU kernel display performance, so that GPU can also play a more important role in terminal-side AIGC tasks. For example, Intel enabled OpenVINO acceleration for the well-known open source image generation model Stable Diffusion. They developed an AI framework that can speed up the PyTorch model through a single line of code installation. With Stable Diffusion's WebUI, you can run Stable Diffusion Automatic1111 on both the Rui Torch integrated graphics card and the Arc stand-alone graphics card.

Through the actual test, we can see the performance effect of Stable Diffusion on the integrated graphics card on the Asustek dawning Air lightweight book. The powerful computing power of the 96EU version of Intel Ruiju Xe graphics card can support FP16 precision models running on Stable Diffusion software to quickly generate high-quality pictures. The editor asked it to generate a picture of "the Man watching TV". On Asustek's dawn Air, it only took more than a minute to make a "smooth release".

During the generation process, CTOnews.com also saw through the performance Explorer that the occupancy of GPU was 100%, while that of CPU was 15%, which shows that this image is indeed rendered locally using GPU.

In the past, it was hard to imagine that lightweight could have this kind of performance, but with the improvement in performance and power consumption ratio of the 13th generation Core processor, and the significant improvement in FP16 and FP32 floating-point performance of Ruiju Xe Graphics (96EU), as well as the addition of INT8 integer computing power, all these have greatly enhanced the overall AI graphics computing power of GPU. This is an important reason why a slim book like Asustek dawning Air can run Stable Diffusion well on the local side.

And in the Intel Meteor Lake processor we mentioned at the beginning, the performance of GPU kernel will be further improved, including 8 Xe GPU cores, 128 rendering engines, 8 hardware light tracking units, asynchronous copy of Arc graphics cards, out-of-order sampling and other functions, and DX12U has been optimized.

From the perspective of AI changing the world, Intel's efforts to widely introduce AI into PC and lead hundreds of millions of PC into the AI era are of great significance, because at least in the foreseeable future, PC is one of the most important productivity tools of mankind. These innovative technologies of Intel enable AIGC to be stably and smoothly deployed on the terminal side of PC, which is a fundamental enabling from the bottom. So that the productivity attribute of PC can be revolutionized, while the change of personal computing will further evolve into the productivity change of the whole society.

All these can fully prove Intel's leading position in the field of AIGC. Their continuous innovation provides users with a more intelligent and efficient computing experience and promotes the development and application of artificial intelligence technology. I believe that with the continuous progress and improvement of technology, we can look forward to seeing more and stronger end-cloud AI applications and solutions from Intel in the future, so that we can more quickly enter the era of productivity liberation driven by AI.

Welcome to subscribe "Shulou Technology Information " to get latest news, interesting things and hot topics in the IT industry, and controls the hottest and latest Internet news, technology news and IT industry trends.