Intel has released version 1.2 of DNNL, the company’s optimized library for high-performance deep learning operations on CPUs and its GPUs. The new version introduces support for int8 on GPUs, which indicates that this will be a standard feature of the Xe architecture.
Phoronix has reported on the release of DNNL 1.2, previously called MKL-DNN: the variant of the company’s math kernel library (MKL) for deep neural networks (DNN). New with this release is support for int8 on GPUs, according to the site (we could not find the patch notes).
The current DNNL guide on data types still says that int8 is not supports on GPUs, with following noteworthy (general) explanation:
“Considering that performance is the main purpose of the low precision data types support, DNNL implements this functionality only for the platforms that have hardware acceleration for these data types.”
In other words, Intel did not previously support int8 on GPUs because its integrated graphics does not natively support the data type so far. This hence provides evidence that Intel is preparing int8 support in anticipation of its upcoming Xe architecture.
Intel has disclosed that Ponte Vecchio, which uses the Xe HPC architecture, has a matrix engine (similar to Nvidia's tensor cores) with 32x int8 acceleration. Intel adding int8 support now already therefor provides a strong indication that DG1’s Xe LP architecture also has int8 support already, even if it may not have the full-blown matrix engine from the Xe HPC architecture.
Among general optimizations, DNNL 1.2 includes supports for AVX-512 and Cascade Lake’s DLBoost, which improved performance of int8 operations (as those are often used in deep learning inference applications). In DNNL 1.2, int8 performance on pre-AVX-512 hardware and for 3D spatial data has been improved. Performance for 1D backward convolutions is also increased, as well as introducing a variety of primitives.
Intel at CES: DLBoost in Tiger Lake
In hindsight, evidence of int8 support was likely already provided at CES, as the company seems to be planning to bring DLBoost to its integrated graphics.
During Intel's keynote, Lisa Pierce said: "So gaming is just a start. We also have advancements in media and display capabilities as well as AI improvements built in."
She then gave a demo of an application using Intel's OpenVINO toolkit for AI, and Intel's vice president of the PC Client Group, Gregory Bryant, remarked: "That's all done using AI acceleration on the GPU which is great."