Having downloaded CUDA Developer SDK more than a year ago for a few experiments with an excellent HuggingFace framework, I've left my email on Nvidia Developer network. A couple of weeks ago, I've noticed an invitation to Nvidia GTC 2022. Topic of federated learning and analytics, applied to data science calculations across commodity hardware and not too many not too expensive video cards bothered me since hitting issues outlined in this NoteInWeb.com blog article. GTC's list of sessions had a few very promising in this regard titles (neglecting the equipment price tags, of course). So I decided to benchmark my understanding of the deep learning technology state and get the news from the first hands (of the newsmakers themselves, that is).
There were several types of events - ranging from full-day educational classes from DLI (Nvidia Deep Learning Institute) to panel discussions to product presentations to best practices/fresh experiences from the field reviews.
I could attend all of the events that raised my interest after business hours in CEST time zone. Obviously, not all of the attended events were about various aspects of federated learning. Other topics of interest included:
Needless to mention, that I realized bias for all of the offered solutions to evolve around Nvidia technologies and offerings, thus often GPU centric (well understandable and welcome).
Session points also include references to the original NVIDIA session announcements with spearker names and titles, quick summary and links to video and often presentations. So I won't copy this details.
5 professionals from the field of AI talk about their career paths, collaboration with academia, and their day to day contribution in their respective industries - healthcare, automotive, augmented/virtual reality, climate/energy, and manufacturing.
4 DL engineers answer chat questions. This is a continuation of the previous event "5 Paths to a Career in AI". As panel member Ke Chen, senior DL scientist in autonomous vehicles at NVidia, outlines, DL Engineers work on DL projects. DL projects often comprise of
DL Engineer's focus may shift depending on project stage:
Great session about high-level differences about regular and multi-threaded CPU programming and GPU programming. I've got an impression that author focuses on igniting deep-level understanding of GPU programming specifics in the audience - SM, thread blocks, struggle for occupancy, use efficiency measurement.
That was an exciting session on transfer learning, I actually started considering attending GTC fall 2022 in the first place. Skip learning zoo of models, selecting large data sets - use TAO for declarative programming based on pre-trained models (with GPU optimied software for Computer Vision, Speech AI, Data Science and Recommendation). TAO now includes support for latest transformer-based models. Its being evaluated for use with Note Web for recommendations. An article dedicated to NVidia TAO would follow if TAO is deemed to suit Note Web needs better than currently used DeepLearning4J.
Its a brief intro of Nvidia Inception along with presentation of DeepStream and OmniVerse Replicator. Nvidia Inception unites 11000+ startups, has grown 59% in 2021, helped to raise $89B so far. I'd have to point out that fairly detailed session parts on DeepStream and OmniVerse target those types of startups that build their business around computer vision, like autonomous vehicles, conversational AI.
Nvidia Inception supports startups in improving these aspects:
Of 6 NVidia Inception nominated startups who presented, I personally found interesting parts about Omniverse from Metaphysic presentation and overall impression of how GPU powered AI proliferates various aspects of business today.
Jay Puri talks about computers write software for computers (ca 9:50). Bernd Leukert mentions monitoring financial streams for signs of fraudulent activity (ca 16:25).
Dask and Apache Spark enjoy continued and accelerated month-to-month adoption.
Data Scientists and ML Engineers spend most of their timie - waiting. Accelerated Data Science - breaks this trend. Data Analytics Trends discussed:
RAPIDS brings open-source GPU acceleration to data science and data engineering.
It is easy to stand up a POC of AI model + software solution in a small demo labs. Rolling that out across the enterprise, addressing all use cases, satisfying all requirements of data governance, RBAC, SSO, teams/projects sharing, cluster-wide telemetery, utilization and resourcemanagement - to name a few. NVidia suggests to start with accelerated datacenter (DGX), based on Large Language Models (both training and inference).
Trends for most companies in AI and ML adoption (per Adept's Erich Elsen):
Nvidia Inception is Nvidia's program for startups. It aims to help startups evolve faster through access to newest technology and Nvidia expert networks, opportunities to connect with venture capitalists, and co-marketing support to improve startup's visibility.
NoteInWeb.com would rely on AWS Nvidia GPU-powered EC2 instances for the couple of streamlined ML/HPC use cases aimed for release and announcement in 2023. So Note In Web, Inc.'s journey with Nvidia Inception is worth a series of articles of its own.
For now, I'd just mention that registration process is straightforward, benefits are attractive and direct partnership costs are at 0.