AMD detailed the Helios-oriented MI450X and scientific-computing MI430X. It says MI450X offers 288GB of HBM4 while retaining MI455X compute, memory bandwidth and connectivity. MI430X provides native FP64, 432GB of HBM4 and up to 23.3TB/s of bandwidth. In its own DeepSeek-R1 MLPerf offline test, AMD reported 21,333 tokens per second per MI455X GPU and claimed up to a 30% lead over a competing platform. Infrastructure decisions should compare model size, precision requirements and operating costs rather than extrapolate from that single benchmark.