ROCm 6.3 adds several new features including a Fortran compiler, and SGLang

AMD
(Image credit: AMD)

AMD has announced ROCm version 6.3, which adds many new updates to the ROCm ecosystem. The latest iteration of the open-source driver stack features several additions, including SGLang, FlashAttention-2, and a Fortran Compiler.

SGLang is a new runtime in ROCm 6.3 that purportedly improves latency, throughput, and resource utilization by optimizing "cutting-edge" generative AI models on AMD's homebrewed Instinct GPUs. SGLang purportedly achieves up to 6X higher performance on large language model inferencing and comes with pre-configured Docker containers that use Python to accelerate AI, multimodal workflows, and scalable cloud backends.

Aaron Klotz
Contributing Writer

Aaron Klotz is a contributing writer for Tom’s Hardware, covering news related to computer hardware such as CPUs, and graphics cards.

  • GenericUsername109
    What about full support of consumer GPUs? Like what CUDA has offered since forever?
    Reply
  • Kurt Lust
    ROCm has had a Fortran compiler for a long time. So far it was based on "classic flang", LLVM backend with a front-end based on code donated by the PGI (now part of NVIDIA). What has changed is that AMD is now also making its version of the new flang compiler available to those who want to try it out, as an alternative to classic flang, which it will ultimately replace. The new compiler offers OpenMP offload and hence more GPU support than the classic compiler.
    Reply
  • Kurt Lust
    GenericUsername109 said:
    What about full support of consumer GPUs? Like what CUDA has offered since forever?
    The officially supported list is rather short. Basically some high-end RDN3 cards and then some RDN2 cards in the pro series. One GCN5.1/gfx906 card is in deprecated mode. That architecture served as the basis for CDNA later on.

    However, there is nothing that makes those RDNA3 or RDNA2 cards special. Except that they have more compute units and more memory than other cards. So those other cards may also work, but are not officially supported, also because AMD considers their memory capacity a bit low. In fact, elsewhere in the documentation of ROCm I found a reference based only on architectures and 24GB of GPU RAM recommended with 8 GB as a minimum.

    The trouble for AMD is that their rendering cards (RDNA) differ more from their compute cards (CDNA architecture) than is the case between NVIDIA rendering and compute GPUs. So let's see what happens when UDNA comes out, likely later in 2026.
    Reply