Matters:
On December 7, Google launched Gemini 1.0, the native multi-modal model, which is Google's largest and most powerful large-scale model to date; at the same time, it also launched a new TPU v5P for cloud-based AI acceleration, which is Google's most powerful and expandable AI acceleration chip so far.
Guoxin Electronics Opinion: 1) Google launched the latest AI acceleration chip, TPU v5p, and deployed OCS (optical circuit switch) on a large scale. Its supercomputer can easily dynamically reconfigure the connections between chips through the OCS switch, which helps avoid problems and make real-time adjustments to improve performance.
We believe that the future of OCS switches will bring significant changes to data center network architectures, thereby reducing power consumption and costs.
2) The input and output ports of the Google OCS switch are two optical fiber collimator arrays. When light enters the OCS switch through the optical fiber, it passes through two 2D MEMS arrays. Each array contains 136 flat mirrors to accurately adjust the direction of light transmission. A MEMS optical switch is a micromirror constructed on a semiconductor substrate based on semiconductor microprocessing technology. That is, electricity, machinery, and light are integrated into a single chip, which can complete arbitrary switching of multiple light wave routes. The core of Google's OCS switch is the MEMS reflector component, and we expect MEMS optical switches to be in strong demand.
3) Saiwei Electronics is in a leading position in the MEMS foundry field. The silicon photonic chip manufacturing technology of the company's overseas MEMS production line is quite mature. It already has experience in process development and small-batch production, and has been supplying well-known European and American manufacturers for a long time. It is expected that the company will expand widely and develop rapidly in the fields of MEMS optical switches and silicon optical chips.
4) Considering the global leading position of the company's Swedish production line and the large-scale production advantages of the Beijing production line, combined with the financial data performance for the third quarter, and the mass production progress of new products such as the company's BAW filters and lidar MEMS galvanometers, we raised the company's performance expectations. It is estimated that 2023-2025 operating income is 13.22/16.88/2,072 billion yuan (previous value: 9.90/13.96/1,785 million yuan), with net profit of 0.37/1.27/206 million yuan (previous value: 0.37/0.7/142 million) Yuan). The current stock price corresponding to PB is 3.57/3.48/3.35, respectively, maintaining the “buy” rating.
Commentary:
Google released the large multi-mode model Gemini with leading performance, and launched the new AI acceleration chip TPU v5p to support the model. On December 7, Google launched the native multi-modal model Gemini 1.0. It is Google's largest and most powerful large-scale model to date, and is available in three magnitude versions: the most capable Gemini Ultra, Gemini Pro for multi-tasking, and Gemini Nano for specific tasks and ends. As a multi-modal large model, Gemini can generalize and seamlessly understand, manipulate, and combine different types of information, including text, code, audio, images, and video.
The Gemini Ultra is the largest and most powerful model, designed for highly complex tasks. According to Google's DeepMind team tests, Gemini Ultra achieved leading results in a series of multi-modal benchmarks. Among them, the score rate was as high as 90.0% in the MMLU (Massive Multitask Language Understanding) test, becoming the first big model to surpass human experts (human expert score rate 89.8%) and surpass the GPT-4 model developed by OpenAI (GPT-4 score rate 86.4%) 。