FAMLIES - Advanced Linear Algebra Framework

The Problem

For decades, the BLAS, LAPACK, and LAPACK derivatives have been part of the foundation of computational science. But their design enforces strict boundaries between layers—between single-node and multi-node computation, between matrix operations and tensor contractions, between CPU and GPU execution.

These boundaries were once a strength, but are now a liability. Memory movement is the central bottleneck of modern hardware, and one way to reduce it is to fuse operations across the stack. Existing libraries cannot do this.

The Solution: FAMLIES

FAMLIES can. By vertically integrating layers that are today rigidly separated—from BLAS-level kernel through LAPACK-equivalent routines to distributed tensor contractions—FAMLIES aims to reduce the performance ceilings that hold back both scientific computing and modern AI workloads.

This project lays the foundation for a new framework that overcomes challenges through vertical integration, the flexible control of algorithms and communication, and a consistent programming API across levels.

Algorithm Families, Not Hard-Coded Choices

FAMLIES allows the best algorithmic variants to be selected for the problem size, target hardware, and memory hierarchy level—replacing the rigid, siloed approach of LAPACK and its descendants.

Vertical Integration Across the Full Stack

A consistent API spanning from architecture micro-kernels up through distributed computation enables cross-layer fusion and communication reduction, while increasing expressiveness and flexibility.

Unified Linear and Tensor Algebra

Dense matrix operations and tensor contractions share deep mathematical structure. FAMLIES integrates both—using building blocks from BLIS and others to power higher-order tensor operations.