Research & Systems Engineering
Foundational research papers, open-source measurement libraries, and systems-engineering analyses on distributed TPU/GPU training efficiency, ML Productivity Goodput, and accelerator hardware co-design.
Beyond Peak TFLOPs: Why Communication-Compute Overlap Dictates ML Productivity Goodput
A systems-engineering analysis responding to Cursor's Mixture-of-Kittens (MoK) paper. Analyzes how unified megakernels, pull-based dispatch, and hardware co-design redefine MoE training goodput across NVIDIA Blackwell and Google TPU architectures.
Machine Learning Fleet Efficiency: Analyzing and Optimizing Large-Scale Google TPU Systems with ML Productivity Goodput
Introduces a systematic framework (MPG) to measure and optimize the efficiency of warehouse-scale ML infrastructures, analyzing system-wide dependencies across Google's production TPU fleet.
Goodput Metric as a Measure of ML Productivity
Introducing the ML Productivity Goodput framework on Google Cloud, decomposing training efficiency into Scheduling, Runtime, and Program metrics to pinpoint and eliminate compute waste.
ML Goodput Measurement Library (AI-Hypercomputer)
A Python-based measurement library inside Google's AI-Hypercomputer repository to instrument ML models and automatically log productive training time and failure recovery overhead.
Orbital Data Centers: High-Performance Edge Computing in Low Earth Orbit
Evaluates FPGAs, GPUs, and TPUs for edge compute in Low Earth Orbit (LEO), analyzing radiative cooling thermodynamics, single-event upset (SEU) radiation hazards, and optical inter-satellite links.