User Guide

IBM

Power

Power10 Quick Start Guide - P10 MMA Performance Guide

P10 Compute & MMA Architecture

Ø2x Bandwidth matched SIMD*

Ø 8 independent Fixed & Float SIMD engines per Core

Ø4 – 32x Matrix Math Acceleration*

Ø 4 512 bit engine per core = 2048b results / cycles

Ø Matrix math outer products of Single, Double & Reduced precision.

ØMMA Architecture support introduced in POWER ISA v3.1

ØSupports SP, DP, BF16, HP, Int-16, Int-8 & Int-4 precision levels.

P10 MMAApplications & Workload Integration

Ø ML & HPC applications with dense linear algebra computations, matrix

multiplications, convolutions, FFT can be accelerate with MMA

Ø GCC version >= 10 & LLVM version >=12 supports MMA through built-ins.

Ø OpenBLAS, IBM ESSL & Eigen Libraries already optimized with MMA instructions

for P10.

Ø Easy integration of MMA for enterprise applications, ML frameworks and Open

Community packages via above BLAS libraries.

PowerPC Matrix-Multiply Assist Built-in Functions

https://gcc.gnu.org/onlinedocs/gcc/PowerPC-Matrix-Multiply-Assist-Built-in-Functions.html

Matrix-Multiply Assist Best Practices Guide

https://www.redbooks.ibm.com/Redbooks.nsf/RedpieceAbstracts/redp5612.html?Open