Developers have always been crucial contributors to and beneficiaries of Intel’s ecosystem of hardware and software resources. A vast collection of our latest hardware (including CPUs, GPUs and Intel® Gaudi® AI accelerators), and software development tools, libraries and frameworks for open AI and accelerated computing are available at our one-stop-shop called the Intel® Tiber™ AI Cloud platform. We supported 13+ collegiate hackathons in 2024, allowing over 3.5K developers (students, educators, startups and other tech enthusiasts) to learn about our latest oneAPI powered AI tools and frameworks, and providing the participants with free access to our cloud platform. In a webinar held back in Dec’24, members from four winning teams at the 2024 hackathons presented their hackathon projects, explaining their real-world GenAI solutions and demonstrating how they took advantage of our AI developer resources on Intel Tiber AI Cloud.
This blog will give you highlights of the webinar and hence the projects presented during the session.
|
Intel’s Developer-focused Efforts and AI Resources
We support a number of developer-oriented events such as hackathons, webinars, workshops and tech conferences throughout the year. At each collegiate hackathon, we have a dedicated track called the ‘Best Use of Intel AI’ where participants compete for leveraging our AI software and hardware on Intel Tiber AI Cloud at their best according to the real-world problem they try to address through their project.
Our team conducts insightful workshops at the hackathons, educating the attendees on GenAI essentials, opimizing AI/GenAI solutions, our cloud services, Intel Gaudi AI accelerators, oneAPI-powered software, and much more. We also introduce our educational initiatives (Intel® Student® Ambassador Program and Educator Program), and Intel® Liftoff Program for AI startups to the attendees.
From [00:08:05] in the webinar recording, the presenter discusses about Intel’s AI software stack as shown in the figure below:
Fig.1: Intel’s AI Software Stack
Now let’s take a look at the GenAI projects presented by the hackathon winners at the webinar.
Project #1: Robotic Registers
The Robotic Registers project, a winner at the Georgia Institute of Technology’s Hacklytics’24 hackathon, was presented by Ajay Yadav from Arizona state University.
Robotic Registers is a Graph Generative Adversarial Network (GAN) model, which is used to create logic circuits within a PyTorch GPU environment on Intel Tiber AI Cloud. This project is instrumental in producing datasets essential for deep learning models in the logic circuit field. It enhances the training data for machine learning algorithms, thereby improving the robustness and variety of datasets in electronic design automation (EDA). By training models on these generated logic circuit graphs, it becomes possible to identify anomalies in real circuits more effectively. Additionally, the synthetic logic circuits act as valuable benchmark datasets for evaluating the performance of various EDA algorithms, aiding in thorough exploratory data analysis to discover patterns, trends, and irregularities in logic circuit design.
By creating standard logic circuit graphs, the model aids in anomaly detection by spotting irregular patterns in actual circuits, helping to identify faulty designs. These generated circuit graphs can be used in optimization algorithms to enhance the performance, power efficiency, and area utilization of FPGA and ASIC designs. Additionally, synthetic circuit graphs provide a secure option for sharing data in research or industry environments where privacy issues restrict the use of real circuit data. The team trained the model on 4th Gen Intel® Xeon® Scalable Processor.
|
Project #2: Easy Deep Learning
The Easy Deep Learning project was also a winner at the Hacklytics’24 hackathon. It was presented by Pranav Devarinti from Georgia Tech.
The team created a groundbreaking no-code platform designed to streamline the fine-tuning of deep learning models. This user-friendly platform allows users to begin with just a handful of images and employs synthetic data generation through stable diffusion and various data augmentations to rapidly build and fine-tune models. The project features a unique architecture that combines a locally developed frontend with Firebase for real-time data synchronization. It uses a stable diffusion Jupyter Notebook on Intel Tiber AI Cloud as the backend engine to develop an advanced synthetic data pipeline, followed by a custom training loop within the same notebook environment.
The team used Intel® Extension for PyTorch* on GPU, resulting in high throughput for large generative models. Using Intel Extension for PyTorch in conjunction with the Intel® oneAPI Math Kernel Library (oneMKL) and the Intel® oneAPI Deep Neural Network Library (oneDNN) helped reduce latency for small classifier models.
|
Project #3: Mind River AI
The Mind River AI project presented by Millie Wu from Columbia University won the University of Virginia’s HooHacks’24 hackathon.
Mind River is a cutting-edge GenAI-driven meditation and emotional support assistant that emphasizes user privacy and customization. This web application generates personalized conversations and meditations for each user, securely storing all interactions and personal data in a private database. Utilizing the LangChain* framework, VectorDB package (powered by Intel’s embeddings), and a fine-tuned open-source LLM model on a medical dataset on Intel Tiber AI Cloud, it provides a distinctive experience enriched by neuro-inspired art aimed at evoking specific emotions.
Apart from leveraging the cloud platform for fine-tuning LLM, the team used Intel’s QuantizedBiEncoder embeddings to accelerate vector search, allowing for faster and more efficient retrieval of personalized data. These embeddings enhance the performance of AI models that process private information securely and effectively. By incorporating Intel’s advanced embeddings, the project ensures enhanced privacy and speed in vector-based data storage and search functionalities, vital for managing and retrieving private wellness information in real-time applications.
|
Project #4: Scribe
The Scribe project presented by Wendon Song from the University of California, Los Angeles (UCLA) and the University of California, Davis (UC Davis) was a winner at UCLA’s LA Hacks’24 hackathon.
Scribe is an GenAI tool designed to enable doctors to make data-driven, precise diagnoses of patients' health conditions. Its goal is to prevent issues like medication errors and misdiagnosis. The team utilized the computational power of Intel Tiber AI Cloud to fine-tune an open-source LLM.
The team leveraged oneDNN and Intel® Distribution of OpenVINO™ Toolkit for faster training and inference with LLM.
|
What’s Next?
Check out the full webinar recording and get started with our AI tools and frameworks for open, high-performance, accelerated AI or GenAI development. Sign up to Intel Tiber AI Cloud and experiment with our latest hardware and software developer resources.
We also encourage you to explore other oneAPI powered toolkits for accelerated, heterogeneous, cross-vendor supported parallel computing.
Useful resources
- oneAPI Programming Model
- Jump-start AI Development
- GenAI Developer Resources
- Become A Top AI Developer
- Developer Events and Initiatives 2024 Recap Video
- Future Developer Events Schedule (Webinars, Workshops, AI Summits, and more)
- Tech Conferences and Events
- Past Developer Events Recordings and Training Materials (Tech Articles, Case Studies, Code Samples, and more)
- Intel Student Ambassador Program
- Educator Program
- Intel Liftoff Program for AI Startups
Technical Software Product Marketing Engineer, Intel