Intel oneDNN 3.7 Begins Tuning For Xe3 Graphics, Adds More Granite Rapids Optimizations
Intel software engineers on Tuesday released oneDNN 3.7 as the newest version of this oneAPI Deep Neural Network Library that is in turn used by software like ONNX, MATLAB, PaddlePaddle, Apache MXNet, and others as part of the building blocks for deep learning software.
The oneDNN software more recently falls under the UXL Foundation umbrella but continues to be heavily optimized for all Intel CPU and GPU wares plus to a lesser extent ARM, RISC-V, POWER, and other ISAs along with AMD/NVIDIA GPUs. With the new oneDNN 3.7 release, there are more optimizations for both current and upcoming Intel hardware.
On the Intel CPU side there are more performance improvements for Intel AMX instruction usage found since Xeon Sapphire Rapids, various AVX2 and AVX-512 performance improvements for INT8, and other mostly INT8 and BF16 optimizations. The AVX2/AVX-512 improvements should also end up paying off for AMD processors too.
Over on the Intel graphics side, oneDNN 3.7 has landed initial optimizations for Xe3 graphics that will first be found in integrated graphics form later this year with Panther Lake. The Intel graphics optimizations also include performance improvements for Xe2 with Lunar Lake integrated graphics and Battlemage discrete graphics.
The oneDNN 3.7 release also has BF16 performance improvements for AArch64 processors, improved MATMUL performance on NVIDIA GPUs with cuBLASLt, and other improvements.
Downloads and more details on the Intel oneDNN 3.7 release via GitHub. Updated oneDNN benchmarks of different CPUs soon.
The oneDNN software more recently falls under the UXL Foundation umbrella but continues to be heavily optimized for all Intel CPU and GPU wares plus to a lesser extent ARM, RISC-V, POWER, and other ISAs along with AMD/NVIDIA GPUs. With the new oneDNN 3.7 release, there are more optimizations for both current and upcoming Intel hardware.
On the Intel CPU side there are more performance improvements for Intel AMX instruction usage found since Xeon Sapphire Rapids, various AVX2 and AVX-512 performance improvements for INT8, and other mostly INT8 and BF16 optimizations. The AVX2/AVX-512 improvements should also end up paying off for AMD processors too.
Over on the Intel graphics side, oneDNN 3.7 has landed initial optimizations for Xe3 graphics that will first be found in integrated graphics form later this year with Panther Lake. The Intel graphics optimizations also include performance improvements for Xe2 with Lunar Lake integrated graphics and Battlemage discrete graphics.
The oneDNN 3.7 release also has BF16 performance improvements for AArch64 processors, improved MATMUL performance on NVIDIA GPUs with cuBLASLt, and other improvements.
Downloads and more details on the Intel oneDNN 3.7 release via GitHub. Updated oneDNN benchmarks of different CPUs soon.