SSBLAST

Project Overview

SSBLAST is a groundbreaking high-performance computing project that introduces the first open-source FP8 scaled linear solver for consumer NVIDIA GPUs. By utilizing FP8-precision matrix kernels with per-tile dynamic scaling and iterative refinement, it recovers full FP64 accuracy while running 2-3x faster than cuBLAS FP64. This tool unlocks enterprise-grade matrix calculations on hardware ranging from the GTX 10-series through the RTX 40-series.

Key Features

• 2-3x speedup over cuBLAS FP64 via FP8 tensor cores.
• Hardware-agnostic support for consumer NVIDIA GPUs.
• Per-tile dynamic scaling to prevent FP8 precision loss.
• Iterative refinement algorithm for full FP64 recovery.
• Easy integration via pip install and drop-in replacements.