Posted on November 6, 2026 by Anastasia Cisneros-Lunsford
A research gift from Tenstorrent is supporting Founding Dean Jinjun Xiong and his student researchers as they explore ways to make AI computing more efficient, affordable and accessible.
A new research gift from Tenstorrent is helping researchers in the College of AI, Cyber and Computing at UT San Antonio explore ways to make AI computing more efficient, affordable and accessible.
Tenstorrent, an AI computing company that develops hardware and software for artificial intelligence, provided Founding Dean Jinjun Xiong with an unrestricted gift grant and technical assistance to support his research on AI models’ acceleration and system software optimization.
The gift includes both financial support and two TT-QuietBox® Blackhole®, liquid-cooled desktop workstations designed for AI model inference and other high-performance computing work. Xiong and his student researchers will use the systems to test and evaluate their work.
Xiong explained that AI computing involves two major processes: AI training and AI inference. Training involves building a model with large amounts of data and allowing it to learn patterns and relationships. Inference is the process that occurs after training is complete, when the model uses what it has learned to respond to information.
“The AI model needs to be trained and it takes a lot of computing time, but it’s a one-time effort,” he said. “Once it’s trained, the model can be deployed for inference to support many people’s queries at scale.”
Why AI hardware matters
Graphics processing units (GPUs) currently handle much of the computing needed for AI training and inference, but the cost of this hardware and the increasing demand for AI computing have encouraged more companies to develop alternative hardware architectures for AI workloads.
Researchers need ways to determine how these different systems compare in terms of important measures, such as speed, energy use, cost and capacity. As models grow, these systems must be able to handle the increasing size of AI models while delivering enough computing performance without consuming excessive energy.
“In terms of power performance, less power means you consume less energy,” Xiong said. “Performance means, can you do it fast enough?”
Part of Xiong’s research focuses on creating fair and systematic benchmarking methodologies for different AI hardware platforms, particularly for inference. The goal is to examine not only raw performance but also the cost-benefit of different systems and their potential for further improvement.
Xiong described Tenstorrent’s technology as a promising option for inference and said that its architecture differs from traditional GPU systems.
“The Tenstorrent’s systems are an exciting alternative in the market to scale AI models’ inference. It’s fortunate that we are able to obtain access to their hardware systems so that we can evaluate this kind of special hardware, understand their performance cost benefit, and provide guidance for anyone who wants to adopt this hardware for inference,” he said.
Optimizing software across systems
Another challenge with AI, Xiong said, is that different hardware architectures often require customized software to be optimized differently, requiring additional non-trivia engineering efforts. Xiong’s research seeks to reduce that burden through a higher-level programming approach. Having access to hardware from multiple vendors is important because Xiong and his student team need to test whether these optimization methods truly work across different hardware architectures.
The team also plans to make its findings openly available, allowing others to learn from this research and better utilize their systems.
Xiong’s research is part of a broader AI ecosystem at UT San Antonio. Faculty across the college and university are working on applications, AI models, training, hardware and other parts of what he describes as the “full stack” of artificial intelligence. His research may also provide insights that will help guide the design of future AI hardware.
“All of this is tightly integrated with the university’s AI strategy to advance AI and make it more affordable and accessible to people.”
— Anastasia Cisneros-Lunsford
Pictures:
Top left: From left, Cristian Rivera, Sri Bhargava Bhamidi, Utkarsh Kumar, Founding Dean Jinjun Xiong and Aashish Dhakal are pictured with two Tenstorrent TT-QuietBox® Blackhole® workstations.
Middle right: Two Tenstorrent TT-QuietBox® Blackhole® workstations.
Bottom left: Utkarsh Kumar and Aashish Dhakal talk to Dean Xiong in their lab.
