There might be some interesting tidbits from this Qualcomm video filmed a few days ago for anyone who has a spare hour up their sleeve.
Show Notes:
0:00:00 - Countdown
0:04:54 - Teaser & Introductions
0:06:35 - About Qualcomm AI Research
0:12:20 - How Has AI Already Impacted Mobile User Experiences?
0:15:40 - How To Encourage More Natural Voice Assistant Interactions?
0:16:40 - Differences And Optimizations Of Generative AI Models
0:18:01 - Challenges Of Bringing AI To Low Power/Edge Devices
0:23:37 - Neurons? Quantization? Basics of AI Lingo
0:28:12 - How Is Accuracy Improved Despite Using Int8/Int4?
0:33:21 - What Mobile Apps Are Already Using Int8?
0:35:30 - Why Shouldn't FP Substitute For Int On Mobile?
0:39:07 - NN Model Optimization Visual
0:39:48 - Compilation Optimization For Various Hardware
0:41:33 - What Considerations Determine Where A Workload Runs?
0:43:23 - Int4 Super Resolution QSRNet Gaming Demo
0:50:00 - How Is Qualcomm Approaching And Researching AI Ethics?
0:52:55 - Stable Diffusion On Mobile Live Demo
1:00:43 - Will Super Resolution, Etc. Make It To Mid-Tier Phones?
1:02:46 - What Aspects Of AI Research Personally Excite You?
1:07:12 - Wrap-Up And Thanks For Stopping By!
Two And A Half Geeks LIVE With Qualcomm Director Of AI Research
by
Chris Goetting — Thursday, June 01, 2023, 11:36 AM EDT
AI has arrived and it seems to be everywhere. But what makes AI tick on ultra-low power devices
like smartphones? How can we expect to harness not only recommendation engines and Google maps, but also, generative AI and true machine learning on devices that consume
only a few watts of power? Let's find out and learn what market innovators like Qualcomm are doing to shape the future of personal AI assistance.
On a very special edition of our Two And A Half Geeks livestream, Dave, Marco, and Chris were joined by Dr. Joseph Soriaga, Senior Director of Technology for Qualcomm AI Research, for a discussion about all things artificial intelligence and the future of AI in your pocket and at the edge.
QSRNet Int4 Super Resolution Demo
Dr. Soriaga explained the role of Qualcomm's AI Research team, provided a crash course on AI terminology like quantization, and covered many of the ways that AI models are being tuned to run at the edge. He also shared a couple live demos with us that ran directly on a mobile phone. The first is a super resolution upscaler (QSRNet) for gaming which leverages Int4 to be incredibly efficient as it converts a 540p input to a 4K frame. Then, he showed us a Stable Diffusion model running locally on the handset, all while taking questions from the audience.