AMD says MMseqs2 and HIPSW++ speed protein search on Instinct GPUs

News

AMD says its ROCm-enabled MMseqs2 and HIPSW++ tools move evolutionary sequence search onto Instinct GPUs, with up to 97 TCUPS in internal tests.

AMD said it has built two open-source GPU paths for evolutionary sequence search on its Instinct hardware, positioning protein homology search and multiple sequence alignment on the same infrastructure used for downstream folding workloads. In a technical article published Oct. 14, 2026, the company said the tools are designed to accelerate an upstream step that often determines how much biological signal is available to AI models.

Evidence and context

According to AMD, HIPSW++ is a GPU implementation of exact Smith adWaterman alignment with affine gap penalties, while ROCm-enabled MMseqs2-GPU is intended for high-throughput homology search and MSA generation. AMD said the two approaches address different parts of the workflow: HIPSW++ retains the exact dynamic-programming algorithm, and MMseqs2 uses filtering to remove unlikely matches before more expensive alignment stages. The company said this matters because modern databases can contain hundreds of millions of proteins, making full pairwise alignment at scale computationally expensive.

AMD described HIPSW++ as using wavefront parallelism on AMD GPUs, with intermediate values kept in thread-local registers where possible and boundary values exchanged through wavefront shuffle instructions. It said substitution matrices are staged in Local Data Share to reduce bank conflicts. For MMseqs2, AMD said the ROCm-enabled implementation accelerates the gapless alignment filter, where each dynamic-programming cell depends only on its upper-left diagonal neighbor, enabling row-wise GPU execution. The company also said the implementation supports multi-GPU execution by splitting the sequence database across multiple AMD GPUs on the same server.

In the testing AMD cited in a footnote, conducted Sept. 21, 2026, ROCm-enabled MMseqs2 scanning UniRef50 with a single query sequence scaled across one, two, four and eight GPUs. AMD said the configuration reached up to 97 TCUPS on eight Instinct MI325X GPUs. The company said MI325X exceeded the corresponding NVIDIA H100 result in every tested 1-, 2-, 4- and 8-GPU configuration, while MI300X was comparable to or higher than H100. AMD also said the GPU implementations delivered substantially higher throughput than CPU-based MMseqs2 and JackHMMER. The footnote said the results came from internal testing, that data loading, initialization and preprocessing time were included, and that database build time was excluded.

AMD’s source material also ties the software work to broader protein-structure workflows. The company said evolutionary search sits upstream of AI models that predict folds, annotate genes or guide protein design, and that keeping search and folding on the same accelerator can reduce data movement and simplify deployment. Separate background material supplied with this report notes that MMseqs2, or Many-against-Many sequence searching, is a C++ suite for large protein and nucleotide datasets and that its GPU implementation was previously described in Nature Methods on Sept. 18, 2025.

The article is framed as a technical announcement rather than a market event. No pricing, product availability changes or revenue figures were provided in the source material, and the performance claims are limited to AMD’s cited internal benchmarks. AMD also noted in the footnote that system configurations may vary and results may differ.