This was due to the 3X performance increase in Snapdragon over the previous generation, making AI a very important part of the SoC and getting a lot of the spotlight. There is no dedicated AI chip, it is the combination of CPU, GPU and DSP. The DSP has a new Tensor core accelerator. Gary said that Snapdragon 855 can perform seven plus trillion operations per second, and target different AI workloads. This gives three times of times performance compared to the previous generation.
Gary shared that AI helps agriculture to use less pesticide, can help the health industry detect diseases and find cures faster - the global economy could get infused with as much as $16 trillion because of AI.
There are new tensor accelerators matching the 4X vector eXtensors, optimize scalar in 32, 16 and INT8 neural network. The key element that you need to take on board is that AI is not a one chip or a one block solution.
Depending on the neural network you want to run - for example image recognition or voice recognition - different neural networks are needed and it is necessary to save as much power as possible while using the most optimized solution. Voice recognition and voice activation can, for example, sit on DSP. Its digital signal processor is very power efficient and Google has shipped more than 10 billion of its DSPs over the years.
Kryo 485 cores based on Cortex A76 with optimized pipeline for performance perform a better job with AI as ARM is aware that AI is a great way to get you to high quality pictures.
Gary showed a great demo from Elevoc that wil extract your voice from background noise using only one microphone. This will work at stadiums, sports event, and music festivals - you will get exceptional voice over IP quality. This was one great demo and I am sure that these guys will get a lot of phone calls after this presentation. It works well, based on the video we saw at the keynote.
Qualcomm partners use Neural Networks to process images getting better bokeh on pictures already taken. What Qualcomm wants to underline is that 2019 will be crucial for AI user cases and different neural networks and it can do a better job than its competition.