Published 8 times a year! AMD ROCm launches a general attack on CUDA: new features are pushed every 6 weeks

Kuai Technology reported on July 25 that AMD recently announced a major adjustment: the ROCm platform will strictly implement an iteration cycle of releasing an official version of new features every six weeks. There are 52 weeks in a year, which means about 8 to 9 official releases will be launched every year. Previously, the official version of X.Y with new features of ROCm was basically pushed on a quarterly basis. The adjusted six-week cycle means that users will receive new features, performance optimizations, and new hardware support significantly faster, with twice the frequency of iterations. The pace of catching up with the maturity of the CUDA ecosystem is also accelerating. The direct background of this adjustment is the release of ROCm 7.14, which is the first production-grade version based on TheRock build system. TheRock is AMD's new automated, open source build and release system. As the new build infrastructure is deployed, the pace of subsequent release releases will be significantly accelerated. Regarding the version number issue, the previous official version jumped directly from ROCm 7.2 to ROCm 7.14, and the middle 7.9 to 7.13 are all technical preview versions. It is difficult for users to sort out version positioning. Fortunately, AMD revealed that when it adapts to the six-week release cycle, the version number rules will also be adjusted simultaneously, but the specific new version number plan has not yet been announced. The next ROCm version is expected to be launched in August this year. The new version launched in this six-week iteration cycle will be officially equipped with the new ROCm.AI software suite, which uses AI agents to help developers automatically complete code adaptation, performance tuning and deployment optimization. For AMD GPU developers, this adjustment means that new hardware does not have to wait another quarter to receive official support. ROCm is launching a general attack on CUDA with faster iteration speed.