Core Viewpoint - AMD has launched the MI350X and MI355X GPUs, claiming they outperform NVIDIA's B200 in AI model inference by 30% and offer significant improvements in performance and efficiency [1][4][24]. Group 1: Product Launch and Specifications - AMD introduced two new GPUs, MI350X and MI355X, built on a 3nm process with 185 billion transistors and equipped with HBM3E memory [2]. - The MI350 series boasts a fourfold increase in computing power and a 35-fold increase in inference speed compared to the previous MI300X [3]. - Both models feature the fourth-generation Instinct architecture (CDNA 4), with 288GB of HBM3E memory and 8TB/s memory bandwidth, which is 1.6 times that of NVIDIA's B200 [9]. Group 2: Performance Comparison - The MI350X and MI355X have peak FP64 performance of 72 TFLOPs and 78.6 TFLOPs, respectively, which is twice that of NVIDIA's equivalent [12]. - In lower precision formats (FP16, FP8, FP4), the MI350 series performs comparably or slightly better than NVIDIA [13]. - The MI355X can process 40% more tokens per dollar spent compared to the B200, indicating higher cost efficiency [5][28]. Group 3: Future Developments - AMD plans to release the MI400 series next year, developed in collaboration with OpenAI, which is expected to be 10 times faster than the MI300 series [7][41]. - The MI400 series will feature the next-generation CDNA architecture, with up to 432GB of HBM4 memory and 19.6TB/s memory bandwidth [42]. - AMD's roadmap includes the introduction of the MI500 series GPUs and Verono CPUs by 2027, aiming to further enhance performance and scalability [52].