Behind the ChatGPT: Exponential growth in demand for chips and semiconductors

Chat GPT (Chat Generative Pre-trained Transformer), a conversational AI model launched by OpenAI in December 2022, has received widespread attention since its launch, reaching 100 million monthly active users in January 2023, making it the fastest growing monthly active user in history. consumer app in history.

Chat GPT mainly focuses on Q&A category, but unlike other Q&A AI products, it has all the knowledge in the training set and has language generation capability to achieve anthropomorphic communication, not just the one-question-answer model like other AI products such as Tmall Genie and Xiao Ai classmates. On the basis of Q&A mode Chat GPT can reason, write code, create text and so on, such special advantages and user experience make the flow of application scenarios increase dramatically.

Chip demand = volume↑x price↑

1) Volume: new scenes brought by AIGC + significant increase in traffic of original scenes

① Technical principle perspective: Chat GPT is a conversational AI model developed based on GPT3.5 architecture, after GPT-1/2/3 iterations, after GPT3.5 model, it started to introduce code training and instruction fine-tuning, adding RLHF technology (Reinforcement Learning by Human Feedback) so as to achieve capability evolution.GPT, as a well-known NLP model, is based on Transformer technology, as the model keeps iterating, the number of layers increases, and the demand for arithmetic power becomes bigger and bigger.

② Operation condition perspective: Chat GPT has three conditions for perfect operation: training data + model algorithm + arithmetic power. Among them, training data has a wide market and low technical barriers, which can be obtained after investing sufficient human and financial resources; the basic model and model tuning have a low demand for arithmetic power, but obtaining Chat GPT functions requires large-scale pre-training on the basic model and the ability to store knowledge from 175 billion parameters, which requires a lot of arithmetic power. Therefore, arithmetic power is the key to the operation of Chat GPT.

2) Price: demand for high-end chips will drive the average price of chips

The cost of procuring a top-of-the-line Nvidia GPU is $80,000, and GPU servers typically cost more than $400,000. For Chat GPT, at least tens of thousands of Nvidia GPU A100s are needed to support its arithmetic infrastructure, with model training costing over $12 million.

From the chip market perspective, the rapid increase in chip demand will further drive up the average chip price. At present, OpenAI has launched a $20/month subscription model, initially building a high-quality subscription business model, and the ability to continue expansion will be significantly increased in the future.

The “hero behind” system GPU or CPU + FPGA and other arithmetic power support

(1) GPU can support powerful arithmetic demand. Specifically, from the perspective of AI model construction: the first stage is to build a pre-training model with large computing power and data; the second stage is to conduct targeted training on the pre-training model. GPU is widely used now because of its parallel computing capability and compatibility with training and inference. At least 10,000 Nvidia GPUs have been imported into Chat GPT training model (the once popular AlphaGO only needs 8 GPUs), and the inference part uses Microsoft’s Azure cloud service, which also needs GPU for operation. So, the hot rise of Chat GPT demand for GPU can be seen.

2) CPU+FPGA swabbing. From the perspective of deep learning, although GPU is the most suitable chip for deep learning applications, but CPU and FPGA can not be ignored. FPGA chip as a programmable chip, can be extended for specific functions, in the second phase of AI model building has a certain space to play. And FPGA wants to realize deep learning function, it needs to combine with CPU and apply to deep learning model together, which can also realize huge arithmetic power demand.

(3) Cloud computing relies on optical modules to realize device interconnection. AI models develop towards large-scale language models led by Chat GPT, driving data transmission volume and arithmetic power enhancement. Along with the growth of data transmission volume, the demand for optical modules as a carrier for interconnecting devices in data centers grows. In addition, along with the growth of energy consumption for arithmetic power enhancement, vendors seek to reduce energy consumption solutions and drive the development of optical modules with low energy consumption.

Conclusion: Chat GPT, as an emerging super-intelligent conversational AI product, needs strong arithmetic power as support, both from the perspective of technical principles and operating conditions, thus driving a significant increase in scene traffic. In addition, the increased demand for high-end chips in Chat GPT will also drive the average price of chips, and the rise in volume and price will lead to a surge in chip demand; in the face of exponential growth in Faced with the exponential growth of arithmetic power and data transmission demand, GPU or CPU+FPGA chip manufacturers and optical module manufacturers will soon usher in the blue ocean market.

Leave a Comment

Your email address will not be published. Required fields are marked *

Scroll to Top