Mainstream storage continues the trend of price increases.
Affected by the overall surge in peripheral AI themes, A-share post holiday related themes have also rebounded significantly. Just yesterday, there was another big news about the AI topic. A startup company named Groq has become popular in the AI industry. Its LPU inference engine has a 10 times faster inference speed than NVIDIA GPUs, and its cost is only 1/10 of that.
Under the influence of this news, relevant individual stocks surged in the late trading session yesterday, with West Test testing reaching the 20cm daily limit and Beijing Junzheng rising by over 17% at one point.
AI chips that are 10 times faster than Nvidia have emerged
At present, the AI chip market is mainly dominated by NVIDIA, and the H100 is also recognized as the most needed GPU for training large language models. Against the backdrop of the rapid development of AI, Nvidia's performance and stock price have both skyrocketed. Wall Street expects Nvidia's fourth quarter revenue to exceed $20 billion, more than three times the same period last year's $6.05 billion, and its results will also be announced on February 22 Beijing time. In terms of stock price, despite a surge of 239% last year, Nvidia's stock price has surged by over 40% again this year, with its latest market value exceeding $1.7 trillion, surpassing Google and Amazon to become the third largest company in the US stock market.
Having a positive outlook on the vast space of the AI market, giants such as Microsoft, Google, AMD, and Intel are also investing in their AI chip business. According to reports, technology giant Microsoft is developing a new network card to enhance the performance of its artificial intelligence chip Maia, ultimately reducing reliance on Nvidia products. In addition, at the end of last year, AMD, as Nvidia's strongest challenger, released the "strongest computing power" Instact MI300X.
The recent news from startup Groq has caught the attention even more. The company is a startup founded by the Google TPU team, who have launched a new type of self-developed chip - LPU (Language Processing Unit), which is used for accelerating large model inference. Its chip inference speed is 10 times faster than Nvidia GPU, and the cost is only 1/10 of it; The generation speed of the running large model is close to 500 tokens per second, and the speed of crushing ChatGPT-3.5 is about 40 tokens per second.
Revealing the truth behind the scenes
According to k_zeros, an investor closely related to Groq on Twitter, the working principle of LPU is completely different from that of GPU. It adopts the Temporal Instruction Set Computer architecture, which means it does not need to load data from memory as frequently as GPUs using high bandwidth memory (HBM). This feature not only helps to avoid the shortage of HBM, but also effectively reduces costs. Unlike Nvidia GPUs that rely on high-speed data transmission, Groq's LPU does not use High Bandwidth Memory (HBM) in its system. It uses SRAM, which is about 20 times faster than the memory used by GPUs.
However, the truth may not be simple. Industry guru Jia Yang cleared an account because Groq's pitifully small memory capacity (230MB) requires 305 Groq cards to run the Llama-2 70b model, while using H100 only requires 8 cards. From the current price, this means that at the same throughput, Groq's hardware cost is 40 times that of H100, and its energy cost is 10 times.
However, this does not affect the A-share market's pursuit of SRAM concept stocks. Yesterday at the end of trading, Beijing Jun was pushing up from the green limit to the 20cm daily limit in just one hour, while the Western Test raised the level by 14% in less than 10 minutes. This morning, some concept stocks continued to rise, with the Western Test testing a 20cm daily limit up, and Beijing Junzheng rose more than 17% at one point.
Three shares of net profit are expected to skyrocket by over 18 times
Industry insiders point out that the above-mentioned concept stocks have seen significant increases. On the one hand, they have indeed been assisted by the SRAM concept, and on the other hand, the core theme of memory itself has also benefited from the great development of the AI era. At the same time, the recovery of storage is also beneficial for the entire sector. According to a research report by Guotou Securities, mainstream storage manufacturers in the market have seen a turning point in performance since the third quarter of 2023, benefiting from the surge in demand for AI servers. High bandwidth storage chip HBM has been highly sought after, becoming an important increment in the development of the storage chip industry. According to TrendForce, DRAM contract prices are expected to increase by approximately 13% to 18% quarterly in the first quarter of this year, while NAND Flash prices are expected to increase by 18% to 23%. This trend will continue throughout the year, and by the end of 2024, DRAM and NAND Flash prices will rise by approximately 60%.
Some companies are also expected to see a surge in performance. According to consistent predictions from institutions, three companies including Langke Technology, Deming Li, and Hengshuo Group all predict a net profit growth rate of over 18 times this year, while individual companies such as Puran Group, Lanqi Technology, Jiangbolong, and Juchen Group all predict a net profit growth rate of over 100% this year. Contrary to performance, most individual stocks have performed poorly, with stocks such as Baiwei Storage, Hengshuo Shares, Dawei Shares, and Zhongdian Xingfa all falling by over 30% this year.
Disclaimer: All information provided by Databao does not constitute investment advice. The stock market carries risks and investment should be cautious.