← Work
Nine Carnegie Mellon Racing members standing behind the team's car outside Margaret Morrison Carnegie Hall.
Photo: Carnegie Mellon Racing

Carnegie Mellon Racing

Firmware, GPU profiling and live telemetry for a Formula SAE race car.

Role
Firmware, controls and data acquisition engineer
Dates
Aug 2025 – present
Carnegie Mellon Racing (Formula SAE)
Tools
C
Python
SPI
UART
Particle Boron
Raspberry Pi
CUDA
Thrust
NVTX
Nsight Systems
TensorRT
Jetson AGX Orin

Carnegie Mellon Racing is CMU's Premier FSAE Formula Student Team. Every year, a team of over 150+ CMU students built an entire racecar from scratch. I joined Carnegie Mellon Racing in August 2025 of my freshman year. My work spans the car's firmware, its driverless software and its data. I started in aerodynamics (August 2025 to May 2026) and have worked on controls and firmware since August 2026.

Where the frame time goes

On the driverless side, I profiled the perception pipeline (TensorRT YOLO on a Jetson AGX Orin) in NVIDIA Nsight Systems. Using NVTX ranges and GPU projection analysis, I broke the 6.5 ms per-frame latency budget down across preprocessing, inference and memory staging.

One frame in Nsight Systems. The CUDA kernel lane sits empty while the image uploads.

About 10% of each frame was GPU idle time, spent waiting on image uploads from pageable memory that moved at roughly 4% of the available DRAM bandwidth. I reviewed the diagnosis with the perception lead, and it's logged as a pipeline optimization item.

This experience was my first exposure to CUDA, Nsight Systems, and NVTX, and I found it fascinating to see how actual hardware components run based on the lower level code.

per-frame latency budget
6.5 ms
of frame time with the GPU idle
~10%
of DRAM bandwidth used by the upload
~4%

Live telemetry

I co-lead the live telemetry module, which talks to the outside world through a Particle Boron (LTE). I'm moving the module's link to the Boron from UART to SPI, with messages encoded as CBOR and framed with COBS. That frees the Boron's only UART to feed a Movella GNSS/INS for more precise GPS. This is in progress.

Moving the Boron link from UART to SPI frees a UART for the GNSS/INS.

Timing the controller

For the driverless MPPI controller, I instrumented the CUDA/Thrust pipeline with reusable CUDA events and NVTX ranges across noise generation, trajectory rollout and reduction, and added per-stage nanosecond timing logs to find the GPU bottlenecks.

Aero sensor validation

Earlier, I validated MS4525DO differential pressure sensors for the pitot-tube aero sensing module using a Raspberry Pi, checking signal integrity and accuracy before hardware integration. I designed CAD test rigs and worked with the aero and vehicle dynamics teams to bring the sensors into the car's telemetry pipeline.

The aero pitot-tube board, a Raspberry Pi and a pressure sensor, taped to the car's carbon bodywork.
The aero pitot-tube board in action.