Snapdragon AI fragments inference tasks for 4x performance

A few minutes ago we were talking about the higher performance we found in the Hexagon NPU of the new Qualcomm Snapdragon 8 Gen 2, higher performance that even allows you to distinguish several elements in real time to process them separatelybut the truth is that the surprises do not end there, since the company has surprised us with another new technology in this SoC.

This is a feature called Micro Tile Inferencingsomething that allows the chip to split inference tasks into smaller layers, allowing to use the three modules of the NPU (Tensor, Scalar and Vectorial) to speed up the tasks that allow it performing what would effectively be processed in parallel, something that was previously impossible since these tasks could not be divided into smaller and independent ones.

This processing technique can be applied in principle to all the tasks that the NPU must carry out, because although we can see how semantic segmentation is carried out in real time, It has also been possible to see a demonstration where a voice-to-text transcription is made instantly in two languagessomething that was previously not possible.

In total, we find an improvement of more than 4 times in performance compared to other Qualcomm chips, an improvement by no means inconsiderable, and that, compared to the competition, certainly shows the importance of this technology.

End of Article. Tell us something in the Comments!