Skip to main content

Advancing Faster, More Reliable Cloud and Edge AI

Support from AMD and NVIDIA helps Daehyeok Kim’s lab build more efficient and practical computing systems and give students hands-on research experience.

Posted by Karen Davidson on Tuesday, October 27, 2026
Daehyeok Kim

As artificial intelligence workloads grow, the computing systems behind them must deliver faster, more reliable results while making better use of available hardware. Assistant Professor Daehyeok Kim is developing infrastructure to meet that challenge, helping shared cloud networks run more efficiently and edge AI applications respond when needed. 

Achieving these advances requires hands-on experimentation. Equipment support from AMD and NVIDIA gives Kim’s lab the tools to build and evaluate new infrastructure, turning research ideas into practical, open-source systems. It also allows Kim to invest more funding in junior researchers and give them experience with industry-grade hardware. In turn, equipment donors can benefit from new research insights and connections with students gaining experience on their hardware. 

Unlocking Cloud Efficiency with AMD 

In shared multi-tenant cloud data centers, multiple users draw from the same pool of hardware resources. This shared environment often leads to network congestion and unpredictable delays. To address these challenges, Kim’s research focuses on programmable Network Interface Cards (NICs), which handle network traffic directly at the server boundary.  

Kim’s lab uses hardware made available through the AMD University Program, including AMD U55 and V80 FPGA (Field-Programmable Gate Array) cards. FPGAs offer customizable compute resources, flexible memory and embedded processing units, making them ideal for prototyping new NIC architectures.  

Using these FPGA cards, Kim and his team are developing an operating system for programmable NICs, called NicOS, alongside related frameworks like Transcraft. This work enables cloud operators and tenants to deploy custom application logic directly onto shared network cards, allowing dynamic provisioning, elastic scaling, and strong resource isolation. By moving compute tasks onto the NIC itself, the project minimizes performance variability and helps multi-tenant clouds run far more efficiently.  

Improving Edge AI Performance with NVIDIA  

While cloud networks handle massive backend workloads, mobile and edge applications demand immediate response times. Edge computing brings AI inference close to end devices, powering latency-sensitive applications like real-time live video generation, immersive computing, and autonomous robotics systems.  

However, today’s edge computing deployments are prone to unpredictable performance. Measurements by Kim’s lab revealed that up to 47% of latency-critical edge AI requests fail to meet their timing deadlines. This degradation stems from a combination of wireless signal fluctuations and compute contention on server GPUs.  

To solve this, Kim received support from the NVIDIA Academic Grant Program for his project, GPU-Accelerated Digital Twins for Predictable 5G/6G Edge AI Systems. Using NVIDIA RTX PRO GPUs alongside his private 5G network testbed, his lab built an end-to-end digital twin that models both wireless radio access networks and GPU resource dynamics in real time.  

This integrated environment allows researchers to emulate thousands of concurrent inference requests and test cross-layer scheduling policies. The same hardware setup directly supports broader initiatives, such as the FarmBrain project, where complex AI computations from lightweight field devices, like agricultural robotics, are offloaded seamlessly to nearby GPU-equipped edge servers.  

A Catalyst for Education and Future Innovation 

The benefits of this support extend to the students developing these systems. In Kim’s lab, six student researchers work directly with the donated hardware, gaining hands-on experience in computer systems, machine learning, operating systems and networking. Students build, test and troubleshoot systems on real-world equipment, preparing them to tackle computing challenges beyond the lab. 

Purchasing high-performance GPUs and specialized accelerator cards can strain research budgets. Equipment donations allow faculty to direct more funding toward graduate student stipends and tuition, supporting the people who make these advances possible. 

The resulting prototypes also provide essential proof-of-concept data for competitive research grants. Early experiments conducted on donated AMD and NVIDIA hardware directly strengthened Kim’s successful proposals for national research awards, including his NSF CAREER Award. 

As UT Austin researchers develop open-source systems and publish their findings, external organizations can use these frameworks to improve their own data centers and AI pipelines. Support from AMD and NVIDIA helps make that broader impact possible, advancing computing research while preparing students to contribute to its future. 

UT Austin provides industry partners with access to cutting-edge research, top talent, and collaborative opportunities to drive innovation and advance AI technologies. Interested in partnering with UT Computer Science? Contact:  
 
Claire Wright 
Director of Industry & Research Relations 
clairew@cs.utexas.edu. 

 

News Categories

For media inquiries:
Mark Evans, Assistant Director of Communications
mark.evans@utexas.edu