Nan Fang Du Shi Bao·2025-11-13 04:53Core Insights - Baidu announced its Kunlun chip iteration plan at the Baidu World Conference on November 13, with the M100 chip set to launch in early 2026, optimized for large-scale inference scenarios, and the M300 chip expected in early 2027, aimed at ultra-large-scale multimodal model training and inference [2][5]. Group 1: Kunlun Chip Development - Kunlun chip is Baidu's self-developed chip project, established in 2011, and has been independently financed since 2021 [5]. - The current flagship product is the P800 chip, which is set to launch in 2024, and has already been validated internally with a deployment of a 30,000-card cluster [5]. - The P800 chip is utilized for most inference tasks within Baidu and has been used to train a multimodal model efficiently [5]. Group 2: Client Applications and Market Reach - Kunlun chip products are not only used internally by Baidu but also deployed in various industries such as internet finance, energy, manufacturing, transportation, and education [5]. - Notable clients include state-owned enterprises like China Merchants Bank and Southern Power Grid, along with an unnamed major internet company [5]. Group 3: Future Product Roadmap - Baidu plans to launch the "Tianchi 256 Super Node" in the first half of 2026, which will support 256 interconnected cards, enhancing total interconnect bandwidth by four times and improving performance by over 50% compared to the previous model [5][6]. - The "Tianchi 512 Super Node" is expected in the second half of 2026, supporting 512 interconnected cards, with a further 1x increase in interconnect bandwidth and capable of training trillion-parameter models [6]. - From the second half of 2027, Kunlun chip will introduce super nodes with 1,000 and 4,000 cards, while continuing to optimize the synergy between hardware and software [6]. - A new generation of N series chips is projected for launch in 2029, with plans to light up a million-card Kunlun chip single cluster by 2030 [6].