How to accelerate the popularization of AI PC? Nvidia looks to RTX as the ultimate answer

Over the past decade, artificial intelligence (AI) technology has moved from theoretical research and small-scale applications to a global technological revolution, completely changing the way we live and work. Whether it is voice assistants on smartphones or complex data analysis and automated production lines, the impact of AI is everywhere, and the efficiency improvements and cost reductions it brings are driving an unprecedented productivity revolution.



In order to seize the opportunity of AI productivity, many brands have quickly come up with their own AI hardware: some mobile phone brands use AI technology to eliminate motion smear when taking photos; Some use AI to remove unwanted elements from pictures, and some companies are even the first to publish the so-called "AI PC" standard.


In comparison, NVIDIA, which keeps a low profile and works quietly, seems a bit out of place. In order to change this "wine alley goes deeper" situation, NVIDIA has also been working hard to increase its exposure opportunities in the AI ​​field in recent years and find ways to make more users aware of NVIDIA's leading technology in addition to game graphics cards - 2024 On April 24, NVIDIA held an offline exchange meeting called RTX For AI in Shenzhen, allowing everyone to experience first-hand how NVIDIA "supports half of the AI ​​industry."



Image source: Lei Technology


Although NVIDIA is not the first company to propose the concept of AI, from a product and technology perspective Look, many historical nodes in computing, including AI, are more or less supported by NVIDIA: In 2008, NVIDIA released the GeForce 8800 GTX graphics card.


Obviously the performance of this graphics card is no longer worth mentioning now, but on this graphics card, NVIDIA proposed the concept of "CUDA" (Unified Computing Architecture). The emergence of CUDA allows the GPU to not only process graphics operations, but also to perform and accelerate general-purpose calculations based on CUDA, making the computer a true "universal tool."



Image source: Lei Technology


In addition to CUDA, NVIDIA further "refined" the computing power of GPUs in 2018. The concepts of RT Core and Tensor Core are introduced to make ray tracing and specialized ML calculations possible - Tensor Core significantly speeds up the training and execution of AI models by efficiently performing large-scale matrix operations. DLSS, which is deeply loved by NVIDIA users and can significantly improve game FPS, is implemented based on Tensor Core. It can be said to be the earliest "real AI" use case that gamers have come into contact with.


Computing power is the basis of all AI




Before the emergence of the AI ​​era, NVIDIA began to think of ways to use Tensor Core to implement AI functions and accelerate The arrival of the AI ​​era; so compared with itself 6 years ago, what kind of technological leap has NVIDIA achieved in the AI ​​field now?


According to NVIDIA, RTX AI currently covers 10 different AI scenarios, namely: AI painting, AI graphic design, AI video editing, AI 3D creation, and AI video experience , AI conference, AI document assistant, AI application development, AI game and AI game development.


Image source: Lei Technology


Although these ten scenarios are different, they all have a common need for computers: computing power. And excellent computing power is precisely the most well-known feature of RTX hardware.



Image source: Lei Technology


There is no doubt that compared with when the RTX graphics card was just released 6 years ago and Tensor Core was introduced, the performance It is the easiest improvement to see from NVIDIA in the field of AI. Taking the most common T2I use case as an example, friends who have tried to deploy models such as StableDiffusion on their own computers should know that most current models have more or less "low hit rates" problems, causing users to need Use the same set of keywords to repeatedly generate images, and use a method similar to the "drawing cards" of mobile games to generate the images you want.



Image source: Lei Technology


In view of this "card drawing" scenario, NVIDIA demonstrated its flagship consumption at the sharing meeting The powerful performance of the RTX 4090D graphics card: Based on the acceleration function of TensorRT, the RTX 4090D can achieve StableDiffusion image generation at the fastest 120fps.


Fine control is the hallmark of AI productivity



I don’t know if you have discovered a detail. In the top ten scenarios just mentioned, NVIDIA puts AI Painting and AI graphic design are distinguished.This is not because NVIDIA wants to use more use cases to support the scene, but because AI painting and AI graphic design actually mark two different stages of AI technology:


AI painting represented by Vincent Tu , due to the low hit rate, users need to continuously generate a large number of pictures to "draw cards" before they can get the finished product they want. And this "uncontrollability" means that the uses of these AIGC works are very limited: either for entertainment, to find inspiration for designers, or as materials for training AI.





But AIGC, which is really used for "productivity", cannot tolerate this kind of "uncertainty". After all, no one wants to use AIGC to show customers When it comes to fashion effects, AI generates three hands on clothes; or when designers use AI to explain interior decoration styles to customers, AI paints the roof of a duplex mansion into a basement.


In other words, the biggest difference between "entertainment AI" and "productivity AI" is whether fine control of AIGC can be achieved.





We know that “entertainment AI” is mainly used to improve user experience and interactivity. For example, in areas such as video games, social media, and online entertainment, AI is used to recommend content, generate music, simulate conversations, and more. The core goal of this type of AI is to enhance entertainment and engagement, with less focus on rigor and predictability of output. This type of AI-generated artwork or music does not need to meet strict commercial application standards, and its creativity and novelty are more important.


In contrast, "Productivity AI" is used in more rigorous and demanding commercial and industrial environments, such as manufacturing, medical care, financial analysis, etc. In these areas, AI is tasked with increasing efficiency, reducing costs and error rates, and providing reliable decision support. For example, AI is used in medical diagnosis to analyze images and identify disease patterns, which requires extremely high accuracy and reliability. In these applications, fine control is not only related to the effectiveness of the AI ​​system, but also to the direct impact of its decision-making quality on human life.



Image source: Lei Technology


At the sharing meeting, NVIDIA also demonstrated what a “productivity AI” should look like—that is, To AI. As an AI application for the field of architectural design, Zhizhi AI provides a variety of pre-trained AI models suitable for different architectural styles and scenarios. At the same time, based on the powerful performance of RTX hardware, Zhizhi AI can operate at a speed of nearly zero latency. AI-generated sketches or lines drawn by the designer are used to explain the exterior design and interior decoration style of the building to customers in a near-real-time manner.


The problems encountered by AI should be solved by AI



Of course, the use cases just mentioned are just the applications of NVIDIA RTX in the AI ​​field a small part. From the entertainment-oriented Wenshengtu and DLSS 3.5, to NVIDIA ACE and sound cloning that change the way of game interaction, to AI video editing that changes the creative mode, and Chat with RTX that changes the work mode, AI technology has already Penetrating into every aspect of our lives.



Picture source: Lei Technology


When sharing the specific application of AIGC in the video creation process, the famous video special effects team "Special Effects Studio" ” He also mentioned a very interesting point - using AI to solve problems encountered by AI. According to what they shared, when reconstructing the depth of field of AIGC images, they did not choose to use traditional manual marking of depth maps. Instead, they directly threw the images to AI, let AI draw the depth map of AIGC, and output the results to another AI model.



Image source: Lei Technology


This kind of "defeat magic with magic" solution, in my opinion, is not just for the AIGC industry , a sign of formalization, and also one of the future development directions of AI.


First of all, the training of AI models requires a lot of computing resources. Since the acquisition of high-quality data is often costly and difficult to achieve, the use of synthetic data generation techniques such as generative adversarial networks (GANs) can create a large number of realistic training data, which is very helpful to improve the training efficiency and effect of the AI ​​system. This technology can not only be used to generate image data, but can also be extended to the generation of text, audio and even virtual environments, greatly enriching data sources and providing more possibilities for AI training.


Secondly, the interpretability of AI models is also an important technical challenge, because many efficient models such as deep neural networks are often like black boxes, and it is difficult to understand their internal decision-making logic.By developing interpretive AI technology, the decision-making process of the model can be made more transparent, increasing user trust, and making it easier for developers to find and improve model deficiencies.



Image source: Lei Technology


From a long-term perspective, solving these technical challenges not only requires more advanced algorithms and model design, but also It is necessary to find a balance between data processing, model training and practical application, which will be the key to promoting the future development of AI technology. We expect AI to bring more convenience, and we also expect it to help us solve old problems in new ways.


And when AI truly and completely liberates human productivity, creators and AI with endless imagination will surely make more unbridled ideas come true.

The above is the detailed content of How to accelerate the popularization of AI PC? Nvidia looks to RTX as the ultimate answer. For more information, please follow other related articles on the PHP Chinese website!