OpenAI shouts out to NVIDIA that GPUs are seriously insufficient! Hardware demand is experiencing resonance; Google brings StyleDrop to explode the AI drawing circle; Microsoft will fully AI-ify Windows by the end of the year | Insight Research

One week highlights

Wall Street Insights and Wisdom Research Perspective:GPUs are in serious short supply, and the demand for optical modules and storage PCBs is resonant.

Overseas:

1. Google StyleDrop comes to the field of AI drawing again, with creativity and more controllable style;
2. Apple will launch the AIGC MR strategy, and AI XR will become the next generation mobile battlefield;
3. The NVIDIA team launches a large 3D video AI model, making virtual reality more realistic;
4. Microsoft will launch Teams 2.0 at the end of the year, and the operating system will launch a general AI attack;

国产

1. Policies in many places are favorable, and Beijing, Shanghai and Shenzhen have successively introduced AI development plans;
2. Tencent invests in large models for the first time! AI startups are at the forefront;
3. Alibaba Cloud AI assistant "Tongyi Listening" is in public beta, and the mobile version will be launched faster than expected;
4. Domestic fully self-developed database, Torsi Haibei has strong demand in the financial and government fields;
5. The Chinese AI large model initiates open source governance to complete the attack and defense of AI from "poisoning" and "detoxification";

Jianzhi Research Perspective

GPUs are in serious short supply, and the demand for optical modules and storage PCBs is resonant.

OpenAI and Supermicro shout out to NVIDIA that GPUs are not enough!

The biggest complaint currently faced by OpenAI customers is the reliability and speed of the API. OpenAI CEO Sam Altman admitted that the current shortage of GPUs has forced many short-term plans to be postponed. Fine-tuning APIs and dedicated capacity products are all limited by GPU availability. However, OpenAI will also provide dedicated capacity to provide customers with private copies of models, but to access this service, customers must pre-commit $1 million.

Liang Jianhou, founder and CEO of Supermicro, said: The market demand for AI is strong. The company is expanding its capacity in the United States, the Netherlands and other places. It also has server production bases in Malaysia and Japan. It is expected to be completed by the end of the year. The production capacity of 4,000 cabinets has been increased to 5,000 cabinets. He also told Huang Renxun that Nvidia should provide more chips, even if it is currently provided, it is not enough.

Jianzhi Research believes:

Driven by the demand for generative AI, GPU products will face continuous shortages and price increases. NVIDIA's current delivery cycle is still lengthening. It basically takes three months or more from one month before to now. It will take longer, and some orders may not be delivered until the end of the year.

In addition, NVIDIA also released the super powerful AI computing platform GH200, which is not only faster but also more cost-effective for large model training. Google Cloud, Meta, Microsoft and software have all announced that it will be used for generative AI work;

For the industry chain: it has become a general consensus that the usage of optical modules has increased, and at the same time, the growth in demand for storage and PCB has begun to slowly materialize. In the past two years, Nvidia's advanced GPUs have led to a sharp increase in demand for HBM memory chips, so HBM orders from the two major memory companies Samsung and Hynix have also grown rapidly

Overseas

1. StyleDrop, Google’s customization master, comes to the AI drawing circle again, which is both creative and more controllable.

StyleDrop is able to capture the nuances of texture, shading, and structure in various styles. With just one image as a reference, no matter how complex the art style is, it can be deconstructed and recreated. Even Nvidia scientists called it a "phenomenal" result.

Jianzhi Research believes:

Compared with MidJourney, the previously popular graphic tool, StyleDrop can better control the style of image generation, and the generated content will be closer to the needs of designers. The characteristic of MidJourney is that it avoids the simple daily camera effects and increases the overall sense of reality when generating ultra-clear images. In addition, it prefers the latter in terms of content and aesthetic preferences.

But the two are similar in that they can draw inspiration from other art media and painting styles and create.

2. Apple will launch the AIGC MR strategy, and AI XR will become the next generation mobile battlefield

The market has very high expectations for MR. Moreover, judging from the fact that Apple has always been a benchmark in consumer electronics, MR, which has been waiting for 7 years, may drop a blockbuster to the XR industry. Everyone is looking forward to MR whether it is in It is a new change brought about by technology and experience.

Jianzhi Research believes:

The rapid development of generative AI and the combination of MR will bring about a comprehensive upgrade of mobile products, especially in terms of application content innovation, which will break through the previous development methods and greatly improve the current XR hit game category with few categories. question.

This will also become an important factor in the sinking of the MR market. Previously, the difficulty in breaking through after XR game penetration reached the growth bottleneck was the nicheization of the application ecosystem, and the number of loyal fans in the Apple ecosystem is extremely large. Under the all-round high-quality integration of content terminal ecology, it will help the rapid sales of MR and drive a new development cycle of the XR industry chain. We once analyzed in AI Daily that VR market leader Meta announced the impact of the autumn release of Oculus 3 on the market in advance.

3. The NVIDIA team launched a large 3D video AI model, making virtual reality more realistic

NVIDIA Research has developed a new AI large model Neuralangelo, which is an AI model that uses neural networks to edit 3D videos for 2D reconstruction. The new model can transform videos from any device into detailed 3D structures.

Jianzhi Research believes:

Although 3D generation technology has long existed, it is worth noting that Neuralangelo, a large AI model, has significantly surpassed all previous methods in terms of its ability to convert 2D videos into 3D objects. The model will select pictures taken from different angles from the 2D video to obtain the details of the 3D object representation, and finally render them to improve detail clarity. The characteristic of this model is that NVIDIA's solution can better structure the video details, making the content look clearer, and it can be applied well whether it is for small statues or large constructions.

Pay special attention to areas that can be widely used in the future: such as virtual reality, digital twins, robot development, industrial digital and other large-scale scenes built using 3D objects.

4. Microsoft will launch Teams 2.0 at the end of the year, and the operating system will launch a general AI attack

Microsoft plans to start using Teams 2.0 version by default on Win10 and Win11 platforms before the end of 2023; launch the Teams 2.0 preview version to Mac, VDI and web users, and further promote it to other customer groups such as education and government.

The new version of Teams promises 3x faster installation, 2x faster startup time, and 1.7x faster switching between chats and channels. Joining meetings should also be 2x faster; memory resource usage should be reduced by 50% and disk space by 70%.

Jianzhi Research believes:

Teams 2.0 is embedded in Windows, and its impact on the operating system will be earth-shaking. This will greatly accelerate the process of PC-side AI, including the convenience of video conferencing, AI chat assistant, Office365 and many other tools. and intelligence will completely change users’ usage habits. What is particularly noteworthy is that the upgraded Teams 2.0 takes up less memory and is faster, so that multi-threaded and high-frequency use will not be particularly laggy.

domestic AI

1. Many places have favorable policies, and Beijing, Shanghai and Shenzhen have successively introduced AI development plans

Jianzhi Research believes:

Various local governments will successively introduce policies to encourage the development of the AI industry. From the construction of computing power of underlying hardware to the research and development of application-side embodied intelligent robots, all will enter a policy dividend period in order to create a better and more open industry. The environment drives the rapid development of the AI industry. Yesterday, Beijing and Shanghai also introduced new policy plans for AI; they include implementing a computing power partnership program, strengthening cooperation with cloud vendors, and providing diversified high-quality inclusive computing power; supporting private investment in major projects, and participating in data, computing power, etc. Artificial intelligence infrastructure construction and other contents.

In general, the current domestic progress in large model research and development is also very fast. Open source large models also have secure databases. The development of AI application side such as media IP, games and other content are being implemented rapidly. We will focus on it in the future. Pay attention to the development of the embodied intelligence track. This field is still in a relatively early stage, and innovative development opportunities are worth looking forward to.

2. Tencent invests in a large model for the first time! Minimax was revealed to have completed US$250 million in new financing

Jianzhi Research believes:

Due to the wave of large model and AI development driven by ChatGPT, many star startups are emerging one after another. MiniMax has become the most profitable star player in the venture capital field in just over a year and a half since its establishment. The virtual chat software product Glow was released in November last year; the generative dialogue AI assistant Inspo was launched in March this year; it also launched an API open platform for enterprise users to support service calls of text and speech models. Since its inception, MiniMax has expanded rapidly and is now valued at more than $1.2 billion. As Tencent invests in large-scale start-ups for the first time, it is foreseeable that with the recognition and pursuit of capital, the AI entrepreneurial atmosphere will become more active.

3. Alibaba Cloud AI assistant "Tongyi Listening" is in public beta, and the application implementation speed will exceed expectations

Jianzhi Research believes:

The implementation of domestic large models in the application field is progressing very rapidly. Tongyi Listening is mainly used in the audio and video fields, bringing users a new experience in recording and reading audio and video content. The user stickiness of traditional software will soon be broken. It is worth noting that in terms of content summary, Feishu Miaoji can only provide keywords; while for the speeches of different guests, Tingwu can provide corresponding speech summaries. At the same time, attention should be paid to the use of large speech models on the mobile terminal. Application progress, such as smart speakers, is a very good port.

4. Torsi released a purely self-developed database—Haibei Search Database V10

Jianzhi Research believes:

Haibei is a purely domestic search engine database that is completely self-developed from the underlying word segmentation algorithm to the core engine and upper-layer system. It has the characteristics of higher-level security, compatibility and high-performance retrieval. It can not only achieve full-field The index supports combined queries in any dimension, and the efficiency of data query and analysis is higher than other big data management systems; it can also realize automatic partitioning of hot and cold data and support the mixed use of multiple storages.

At the application level, especially for fields with strong specificity and high security, such as banking, government affairs, military industry, etc., it shows very strong competitiveness.

5. The first anti-discrimination confrontational open source project of Chinese AI large model: everyone asks 100 "poisonous" questions

Jianzhi Research believes:

Data annotation is a crucial step in the large model process. Only by using the annotated "safe data set" for model training can we obtain ideal training results. However, data standards have always been accompanied by subjective, religious, and personal preference characteristics. Therefore, if you use foreign data sets for training, you will be "acclimated" to a certain extent, and it is very important to build a local training data set. Chinese AI’s first anti-discrimination confrontation project has gathered many industry experts and will become one of the high-standard data sets for domestic open source large model training.

Follow next week

Apple WWDC conference, can MR live up to expectations and lead the XR industry into a new era.

The above is the detailed content of OpenAI shouts out to NVIDIA that GPUs are seriously insufficient! Hardware demand is experiencing resonance; Google brings StyleDrop to explode the AI drawing circle; Microsoft will fully AI-ify Windows by the end of the year | Insight Research. For more information, please follow other related articles on the PHP Chinese website!