Power ISA v3.01 has introduced a new set of architecture capabilities to accelerate matrix math operations called MMA (Matrix Multiply Assist). This will improve the performance of key compute kernels such as Matrix Multiplication, Convolution, Fourier Transform etc. which are fundamental building blocks for ML/DL workloads. This talk will introduce the audience to the POWER ISA MMA capabilities and detail the following: MMA’s architectural features, new instructions and lower/mixed precision data type support. The talk will cover programming a simple matrix multiplication function using MMA code and differentiate it from the traditional VSX (Vector Scalar Extension) based code sequence. The talk will delve into details of exploiting the full capability of MMA with assembly-level optimal code sequences for different data types. This talk will also cover compiler built-ins from GCC-10 and discuss exploitation of MMA capability in ML/DL frameworks using MMA-enabled OpenBLAS library.
Continue the conversation in Slack
Open Zoom