A configurable experimental environment for large-scale edge to cloud research

This month, we have new H100 GPU nodes on KVM@TACC! Today, you can launch VM instances with 1 full H100 GPU. This hardware comes with a brand new workflow for reserving VMs. It’s important to note that this reservation workflow will be rolled out to the rest of KVM later in the summer. Additionally, we have refreshed our documentation. Lastly, CHI-in-a-box comes with a new image deploy tool for associate sites.
 

What's the secret ingredient that makes our new Chameleon images so much better? From automatic SSH configuration to built-in rclone support, these aren't your ordinary cloud images. Find out what makes them special.

View or contribute to the experiment packaging and style checklists (appendix A and B) on our GitHub repository here.

Download the report here.

We’re excited to announce the publication of the NSF-sponsored REPETO Report on Challenges of Practical Reproducibility for Systems and HPC Computer Science, a culmination of our Community Workshop on Practical Reproducibility in HPC, held in November 2024 in Atlanta, GA (reproduciblehpc.org).

Leo Han, a second-year Ph.D. student at Cornell Tech, conducted pioneering research on the fair attribution of cloud carbon emissions, resulting in the development of Fair-CO2. Enabled by the unique bare-metal capabilities and flexible environment of Chameleon Cloud, this work tackles the critical issue of accurately distributing responsibility for carbon emissions in cloud computing. This research underscores the potential of adaptable testbeds like Chameleon in advancing sustainability in technology.

High-resolution Vision-Language Models (VLMs) offer impressive accuracy but come with significant computational costs—processing thousands of tokens per image can consume 5GB of GPU memory and add 15 seconds of latency. The HiRED (High-Resolution Early Dropping) framework addresses this challenge by intelligently selecting only the most informative visual tokens based on attention patterns. By keeping just 20% of tokens, researchers achieved a 4.7× throughput increase and 78% latency reduction while maintaining accuracy across vision tasks. This research, conducted on Chameleon's infrastructure using RTX 6000 and A100 GPUs, demonstrates how thoughtful optimization can make advanced AI more accessible and affordable.