Intel® Liftoff Days returned for its third edition, running from October 21–25, 2024. This quarterly virtual sprint provided Intel® Liftoff members with a focused week of innovation centered on multimodal AI for early-stage startups. Participants explored Intel® Tiber™ AI Cloud on Intel® Gaudi® 2 AI accelerator, tackling cutting-edge AI projects across diverse domains.
What We Covered
The workshops and brainstorming sessions were focused on AI and designed to update participating members on the latest techniques and get people talking and sharing new ideas.
Here’s a snapshot of what we covered over the course of the week.
The Intel® Liftoff Days hackathon kicked off with a launch day, where participating startups pitched their ideas and shared plans to leverage Intel® Tiber AI Cloud and Intel® Gaudi hardware.
The agenda included two hands-on workshops designed to equip startups with the tools needed to optimize their projects:
- Fine-tuning and inference of GenAI models on Gaudi2
- Computer vision AI using Intel GETI
Throughout the week, startups received personalized guidance from Intel experts during individual tech mentoring sessions.
The event wrapped up with Demo Day, where all 6 participating startups showcased their final projects. This was a chance for participants to demonstrate the innovative applications they developed, putting the knowledge and skills gained during the week into practice.
Startup Gallery
The startups participating in the hackathon were Intel® Liftoff members - early-stage AI companies from diverse industries, selected for their potential to create impactful solutions. Their goals ranged from applying computer vision to optimize wind turbine maintenance to fine-tuning language models for specific, real-world applications.
Each team was paired with a dedicated mentor and given access to Intel® Tiber AI Cloud, providing the tools and expertise to bring their ideas to life. Beyond building their projects, startups exchanged insights, collaborated, and connected with Intel experts and fellow innovators, making the hackathon a platform for both technical progress and meaningful collaboration.
uses deep machine learning to forecast where the wind will blow, and using a power grid model to buy and sell utility-scale wholesale power.
|
a no-code Platform-as-a-Service enabling AI studios, startups, and enterprises to rapidly and affordably develop AI-powered Apps and Dapps through decentralized parallel training mechanisms. |
a machine learning organization dedicated to driving innovation across industries by creating visually stunning imagery through its Pixel ML Gen AI Studio. |
is digital asset lifecycle overlay platform for renewable energy asset managers that enables 24/7 oversight of renewable energy power plant performance, failures, and stoppages. |
is an award-winning, patent-pending system and approach for detecting and estimating people count information acquired by signal sensing and other technologies. |
focuses on text-to-image and LLM technology, addressing the challenge of adapting diffusion models and LLMs for Indic use cases. Their work involves enhancing the cultural context of model training and incorporating multilingual capabilities to better serve diverse linguistic and cultural needs.
|
Projects We Loved
First place was awarded to Amplify Renewables for their multimodal weather prediction model, integrated into their trading pipeline using Gaudi. They received one month of free Intel® Tiber AI Cloud credits on their chosen hardware.
Second place went to Pixel ML for their image-to-video model and flux inference workflow, while an additional prize was given to PeopleSense for their computer vision workflow and ridership prediction model on edge devices.
Amplify Renewables
Amplify Renewables focuses on grid forecasting for renewable energy sources like solar and wind. Their platform predicts energy generation and simulates grid conditions to identify and mitigate congestion points. By integrating weather forecasting data, such as wind speed, solar radiation, and temperature, they aim to optimize energy distribution, reduce curtailment, and improve grid efficiency.
Hackathon Objective
During the hackathon, Amplify Renewables focused on running large-scale weather forecasting models on Intel hardware, prioritizing inference over training. Their graph neural network (GNN)-based models were tested on Intel’s Gaudi architecture to evaluate performance and compatibility. Their long-term goal is to fine-tune these models using new data from recent weather events, enhancing forecasting accuracy and reducing grid congestion.
What Did They Achieve During Intel® Liftoff Days?
Amplify Renewables developed a multimodal weather prediction model for power grid forecasting. The model leveraged high-resolution weather data to predict renewable energy output, with a focus on wind patterns and grid congestion in regions like South Texas.
- Successfully implemented the model on Intel® Gaudi 2 for training and inference, achieving significant GPU utilization despite port configuration and profiling limitations.
- Reported Gaudi 2 as easy to configure and effective for high-load weather modeling tasks.
Next Steps
Amplify plans to continue optimizing Gaudi 2 for their high-frequency, bursty computational needs and aims to deploy inference on CPUs for improved operational efficiency.
Pixel ML
Pixel ML’s project centers on transforming text-based input into video content. Their platform uses large language models (LLMs) to analyze user-provided scripts, breaking them into scenes and generating corresponding visual elements with Flux models. This approach combines automated video creation with personalized visual elements, offering users an innovative way to produce videos with minimal manual effort.
Hackathon Objective
During the hackathon, Sean focused on solving a key challenge: maintaining character consistency in videos. The platform currently struggles with ensuring the same character appears visually consistent across multiple scenes. To address this, the startup team proposed enabling users to upload custom character images and train personalized models using Flux. These models would ensure consistent character appearances throughout the generated video. The project also leveraged Intel’s Gaudi hardware to accelerate the training process and improve model performance.
What Did They Achieve During Intel® Liftoff Days?
Pixel ML advanced their image-to-video model with Flux inference, focusing on achieving consistent frames for high-quality video generation.
- Successfully utilized Intel® Gaudi 2 for inference, achieving partial functionality but encountering slower-than-expected iteration times.
- Identified areas for improvement, including optimizing Intel® Gaudi 2's compatibility with their model and enhancing frame consistency in video outputs.
- Planned further testing with other multimodal AI models like COG Video X and experiments to maximize GPU utilization on Intel® Gaudi 2.
The team highlighted the responsiveness of Intel® Tiber AI Cloud’s support team, whose assistance was instrumental in addressing inference bottlenecks.
PeopleSense.AI
PeopleSense.AI specializes in real-time crowd occupancy monitoring for transportation systems, focusing on intercity train corridors. Their platform provides dashboards with live occupancy updates, heatmaps, and crowd behavior trends to help optimize transportation services. The system is already in use across multiple train networks in Northern California, including the Capitol Corridor and Joaquin Corridor, and also supports crowd management for special events like games at Levi's Stadium.
Hackathon Objective
During the hackathon, PeopleSense.AI aimed to enhance their platform by integrating machine learning-based prediction modeling into their existing crowd-monitoring dashboards. They also explored incorporating security camera data for crowd analysis during 49ers football games and leveraged Intel® OpenVINO toolkit to detect cracks on bike racks frequently used by commuters. Additionally, the team worked on improving real-time occupancy models by adding machine learning capabilities to more accurately predict future crowd densities.
What Did They Achieve During Intel® Liftoff Days?
PeopleSense.AI developed a multimodal AI solution to monitor train ridership and bike rack usage, leveraging Intel Geti and predictive modeling tools.
- Achieved high accuracy in predicting bike rack occupancy by training a computer vision model to detect empty racks, tires, and seats.
- Utilized Intel Geti’s labeling features to quickly refine and adjust the model.
Next Steps: The team plans to explore Intel® GETI’s SDK further and incorporate more consistent camera angles to enhance accuracy in real-world deployments.
Experience with Intel Technology: PeopleSense highlighted GETI’s ease of use, which enabled them to adapt quickly to complex image datasets and streamline model development.
Lessons We Learned
The hackathon offered valuable insights into the capabilities and applications of Intel’s technologies, demonstrating their impact across diverse use cases:
Gaudi Architecture
Startups praised the Gaudi ASICs for their excellent performance on multimodal models, emphasizing their cost-effectiveness for high-load tasks compared to traditional GPUs:
- Amplify Renewables successfully deployed a Swin Transformer-based model for weather prediction—the first of its kind on Intel® Gaudi hardware. They are preparing for production as soon as Gaudi 3 VMs become available.
- Pixel ML achieved up to 4x cost savings in video processing, demonstrating Gaudi’s efficiency for handling complex workloads.
Intel GETI
Participants highlighted Intel GETI for its user-friendly setup and fast model deployment capabilities:
- Reama AI and PeopleSense.AI reported significant reductions in model training time and improved data labeling efficiency, accelerating their development processes.
What Our Participants Had to Say
We were very impressed with how easy it is to use Intel Gaudi VMs with our Pytorch models. Profiling, optimizing, inference - everything worked out of the box.
– Rachit Singh, CTO, Amplify Renewables.
The recent Intel Liftoff Days during the week of 21 to 25 October 2024 was an exceptionally executed hackathon. The combination of hands-on coding challenges, knowledgeable mentors, and collaborative atmosphere created an ideal learning environment for participating startups like ours. The Reama AI team has gained valuable skills and connections that will benefit our development journey long after the hackathon.
– Tebogo Mohlahlana, CEO and Co-Founder, Reama AI.
About Intel® Liftoff
Intel® Liftoff for startups is open to early-stage AI and machine learning startups. This free virtual program helps you innovate and scale, no matter where you are in your entrepreneurial journey.
Resources used during Intel® Liftoff Days
Intel® Tiber™ AI Cloud - Cloud platform for AI development and deployment
Intel® Distribution of OpenVINO™ toolkit (Powered by oneAPI) - Framework for optimizing deep learning inference
Intel® Extension for PyTorch* - Optimizations for PyTorch on Intel hardware
Intel® Gaudi® 2 AI accelerator - High-performance AI training processor designed for deep learning workload
As a Software Tools Ecosystem Specialist at Intel, I’ve had the privilege of working on the dynamic GenAI initiative. My focus is on driving engagement with the software developer audience. I'm a proud team member of the Intel® Liftoff for Startups and the Developer Engagement, Relations, and Studio team.