Co-op Overview: During the spring semester of 2026, I took a gap semester to work at AMD as a Datacenter Failure Analysis Engineer. This role gave me experience with one of the largest chip design companies, and gave me insight into the workings of frontier AI infrastructure units, especially GPUs. My main goal was to own my own failure analysis cases as well as automate processes within the Failure Analysis team. The internship was thus focused on debugging and resolving customer product failures, and escalating and working with design engineers on issues that were consistently produced.

Failure Analysis Cases

My primary goal was to get experience with each category of failure mode in FA. Additionally, I was able to gain experience across all case escalation levels. Most of the cases I worked on fell into three broad categories.

• Firmware issues: A large share of cases came down to firmware not behaving as expected. These required debugging to determine where the behavior deviated from what was intended and whether a fix was needed on the firmware side.

• Hardware issues: Hardware failures such as shorts and open circuits were identified using thermal imaging and inspection of the ASIC. Thermal cameras helped locate abnormal heat signatures, which narrowed down where to look before moving to deeper inspection.

• Power issues: I worked with two types of power failures. Standard power failures were cases where the power indicator showed no power to the system. Transient failures were cases where the power indicator flickered, which pointed to more serious underlying problems and required more advanced debugging.

While debugging individual cases, it was important to identify trends across failures and flagging them to design teams, making the case for why certain failure modes needed to be addressed at the design level. Instead of treating each case as a one-off, I looked for patterns that pointed to recurring root causes, then communicated them so they could be fixed in future firmware and hardware revisions.

Additionally, cases moved through three escalation levels depending on complexity.

• L1: Cases where the solution could be determined through basic field health tests.
• L2: Cases that required further debugging with the team and more advanced debugging tests.
• L2+: Cases that required consulting design field expert specialists from other teams, along with repeated rounds of debugging and testing.

Every case required documentation, especially L2+ cases. This ensured that similar issues could be avoided in the future and that lessons were carried into firmware and hardware updates for the next design. Working across all three levels gave me a complete view of how a failure moves from the field to a design fix.

Test Automation

A significant part of my role was automating the test procedures used across failure analysis cases, which originally involved a lot of manual setup and input. I helped develop and configure hardware test systems and contributed to implementing a new failure analysis system that would be used across AMD.

The system was especially valuable for one test that normally requires extensive setup and a different testing configuration from the rest. To make that possible, I worked alongside teams at other AMD locations to develop the hardware testing approach and implement it on the systems themselves. Overall, the system reduced manual effort, made testing more consistent, and freed up engineers to focus on the harder debugging work.

System Allocation Dashboard

I built a Power BI System Allocation Dashboard that gave teams visibility into their testing infrastructure. It showed which tests were running on different GPUs and CPUs, which tasks were causing failures, and which systems were available for loading GPUs (on GPU test devices) and CPUs (on CPU test devices). This made it easier for teams to plan their testing and use available hardware efficiently.

Industry Exposure

I sat in on many meetings with senior leadership about the state of failure analysis, overall trends, and customer sentiment. Seeing which issues came up most often gave me a view into the broader AI infrastructure space, how AMD is positioned within it, and how customers interact with the company as a whole. It gave me context beyond my day-to-day cases and a better understanding of where the AI landscape is heading.