Click the blue research report sharing platform above and follow us to find us. Thank you!
On the evening of May 11, I met with Dr. Li Yuan, an alumnus from Tsinghua’s class of ’89. I have been following Dr. Li’s work on AI chips (see the link for breakthroughs in deep learning AI chips by Chinese innovators), and this meeting brought uplifting news: he has confirmed a specific direction for artificial intelligence chips, which is the large chip GPU. Yes, it’s a GPU that competes with Nvidia.
1. The GPU is to AI what the CPU is to PCs, holding a core position. Due to its parallel architecture, the GPU is particularly suitable for deep learning, facilitating the explosive growth of artificial intelligence and contributing to Nvidia’s tenfold stock price increase. Currently, AI main chips are divided into three paths: GPU, FPGA, and ASIC, each with its own advantages. Domestic AI newcomers like Cambricon and Bitmain focus on ASIC, while some national teams invest in costly FPGAs, perhaps believing Nvidia is too strong, holding 80% of the GPU market, making direct confrontation unwise. This has instead become an opportunity for Dr. Li to enter the GPU market.
2. First, let’s discuss whether GPUs still have competitiveness in AI. The answer is affirmative; AI has not yet been fully deployed in human society, and the most common applications currently are in security (face recognition) and healthcare, which are precisely the strengths of GPU image accelerators. Dr. Li mentioned that in many scenarios, only 30% of computations are deep learning (which can be handled by both GPUs and ASICs); the rest are non-deep learning computations that can be performed by GPUs. Relying solely on ASICs may hinder other supporting functions, thereby reducing AI efficiency. Moreover, AI algorithms are still evolving, which is a significant concern for chip developers. GPUs are general-purpose performance chips that can quickly adapt to millions of variations.
3. Next, how to compete with Nvidia? Nvidia GPUs emphasize performance, primarily used in cloud data center servers, where price sensitivity is low and power consumption is not a strict requirement. These two points create challenges for edge servers using Nvidia GPUs: first, massive terminals prefer lower prices, and second, power consumption metrics are a pain point. This is precisely where Dr. Li’s GPU enters. Utilizing the RPP reconfigurable parallel data flow architecture founded by Li, it can achieve high computing power and low power consumption at 28nm process technology without resorting to costly sub-10nm processes, thus significantly reducing costs, allowing prices to reach one-fifth of Nvidia’s, making it highly competitive.
4. Edge servers can be made into boxes, inserted into terminals without altering existing wiring, which is particularly suitable for breaking the current landscape in security, gaining favor from new AI forces (such as SenseTime and Megvii). The edge server’s GPU must have low power consumption and support passive cooling. This is the target solution for Li’s GPU. Fortunately, this GPU solution has been designed and verified through FPGA. Recently, it has also been recognized by major clients for the entire architecture design, and even before the chip is produced, clients have pre-approved it, which is a rare occurrence in China and a significant encouragement.
5. According to the plan, Li is about to initiate a second round of financing, seeking several million dollars, aiming for early next year to officially produce the GPU chip. In the current atmosphere of “everyone is making chips, and all are on the same chip,” I believe many investment institutions will actively come to negotiate.
6. Finally, let me introduce Li Yuan. He graduated from Tsinghua’s Physics Department in ’89, obtained his PhD in Japan and Canada, and worked for many years at major companies in the US like TI and Intel, accumulating 15 years of experience in communication base station chips and server main chip architecture design. He returned to China to start a business early last year, focusing on artificial intelligence chips, striving to create a domestically produced GPU and achieve the goal of self-sufficiency.
Here’s wishing my old classmate and his team great success in the national chip strategy development trend, persevering forward and achieving recognition.
Wishing all friends and investors a prosperous 2018!
Come on, follow this public account, and you will definitely feel it is worth it. Thank you!
This research report is merely a sharing of facts and content; all authors’ speculations, conjectures, and forecasts do not constitute investment advice. Readers are advised to think and make decisions independently.