Divya Mahajan
Georgia Institute of Technology, GA·Computer Science·Early-career
Training today's artificial intelligence (AI) models takes thousands of specialized chips, such as graphics processing units (GPUs), working together in data centers that draw hundreds of megawatts of power per site. These chips run hot, see power spikes beyond their rated limits, and slow themselves down automatically to avoid damage.
- Fields
Computer Science
Engineering
- Topics
AI infrastructure
data centers
distributed training
- Grants
- 1 known grant, through Jun 2031
What they work on
- Build a cross-platform framework that measures how thermal hotspots, frequency throttling, power spikes and energy limits restrict distributed AI training performance on GPUs and other accelerators.
- Develop runtime algorithms that use real-time hardware telemetry, meaning live temperature and power readings, to guide device placement, micro-batch scaling and kernel scheduling across mixed accelerators.
- Create a co-design method that jointly tunes accelerator architecture, memory hierarchy, network topology and runtime configuration, treating the whole data center as one coupled optimization problem under thermal and energy limits.
- Release open-source profiling tools and benchmarking results for AI systems research, and train students through hands-on experiments on shared accelerator clusters.
More about the research
Yet the software that runs AI training usually assumes every chip runs at full speed and behaves identically all the time. This project argues that the gap between those assumptions and physical reality will widen as models and data centers grow.
Why it matters
AI data centers consume enormous amounts of electricity and lose performance when chips overheat or throttle. Scheduling and hardware designs that account for temperature, power and cooling could let operators train large models faster on the same hardware with less wasted energy. Chip designers and data center operators could also use the planned open-source profiling tools and benchmarks to plan infrastructure that holds up under real physical conditions.
Known grants
From NIH and NSF research grants. Labs often have other funding we can't see.
Closest labs
Shivaram Venkataraman
University of Wisconsin-Madison, WI·Computer Science·Early-career
Characterize performance variability across many types of accelerator hardware, where identical applications run at different speeds on chips of the same type, and identify which applications are most affected (Fortuna).
Omiya Hassan
Boise State University, ID·Engineering
Co-design an electro-optical application-specific integrated circuit that pairs photonic (light-based) and electronic processing units, assigning AI training and inference to the best-suited unit.
Rachee Singh
Cornell University, NY·Computer Science
Measure how communication speed and delay vary between accelerators in cloud data centers, accounting for proprietary interconnects and hidden network paths that current tools ignore.
The school
Georgia Institute of Technology: 50 labs on PI Pages, $1.5B a year on research, 572 research doctorates a year.
Reach out
A start for your email. Fill in the brackets.