v2.5.0

Summary of Key Changes in AMGX (v2.4.0 → v2.5.0)

CUDA Upgrades

• Blackwell (B200/GB200/RTX Pro 6000) support
• CUDA 13 support
• Minimum CUDA version raised from 10.0 to 12.0, tested up to 13.0
• Dropped support for older GPU architectures: SM20, SM35, SM52, SM60 removed
• New minimum: Volta (SM70)+, with support for SM75, SM80, SM86, SM89, SM90, SM100, SM120
  • Tested on Hopper, Blackwell (incl. RTX Pro 6000)
• Consolidated and removed repetitive / redundant architecture-specific code

Build System Changes

• Deprecated CUDA_ARCH in favor of standard CMAKE_CUDA_ARCHITECTURES
• Removed Thrust submodule dependency - now uses system/CUDA-bundled Thrust
• Removed OpenMP dependency entirely
• Removed NVTX linking in favor of NVTX3

cuSPARSE API Updates

• Removed mixed-precision support (DISABLE_MIXED_PRECISION removed)
• Consolidated to use only generic cuSPARSE SpGEMM interfaces and added multiple flags
  -- use_cusparse_spgemm, cusparse_spgemm_alg, etc.
• Removed legacy cusparseCsrgemm2 wrapper implementations

Error Handling Improvements

• New AMGX_CHECK_API_ERROR_NORSRC macro for resource-independent error checks
• Improved error handling throughout

Memory Management

• Added runtime detection for cudaMallocAsync support via cudaDevAttrMemoryPoolsSupported
• Fallback behavior when async memory pools aren't supported by the device

Perf optimizations

• Optimized hash_set insertion
• Fixed perf bug with fill_A_kernel_1x1

Bug fixes

• MPI comm dup bug
• Convergence check for absolute testing against relative
• Block size handling in distributed_arranger resize of A
• Block size handling in renumbering and reordering components
• Bug in scaled norm factor calculation
• Fixed output performance for matrix writer
• Integer overflow in dense LU
• Fixed handling of latency hiding to use global row count

MPI Example Enhancements (e.g., amgx_mpi_capi.c)

• Added -cd flag for diagonal dominance checking
• Added -om flag for matrix output/writing
• Added -r flag for performance repeat runs