AMD has released ROCm version 6.3, introducing significant enhancements to its ecosystem. Key features include SGLang, which optimizes generative AI models on AMD’s Instinct GPUs, achieving up to 6X performance improvement in large language model inferencing.
Also introduced is FlashAttention-2, offering up to 3X speedup for Transformer AI models, reducing memory and compute requirements during training. Additionally, a Fortran compiler is now part of ROCm 6.3, allowing for the execution of legacy Fortran applications on modern GPUs. This compiler supports GPU offloading via OpenMP and ensures backward compatibility for the continued development of existing applications.
Read More: ROCm 6.3 adds several new features including a Fortran compiler, and SGLang
