Date: November 6, 2026
Speaker: Tanmaey Gupta, Ph.D. Student, Cornell Bowers
Title: Making Every Bit Count: Quantization and Better Scaling for Low-Precision ML
Host: Chris De Sa

 A color photo of a man.

Abstract: Quantization is a key technique for improving the efficiency of modern machine learning models by reducing computation and memory costs through low-precision representations. This talk will first introduce the fundamentals of quantization, common approaches, and emerging microscaling formats supported by modern GPUs.

I will then present ScaleSearch, which improves microscaling Block Floating Point (BFP) quantization by searching for scale factors that better minimize quantization error than conventional max-based scaling. I will show how ScaleSearch can improve post-training quantization and low-precision attention, including ScaleSearchAttention, an accelerated NVFP4 attention algorithm that closely matches baseline model accuracy.

Bio: Tanmaey Gupta is a Ph.D. student working with Prof. Chris De Sa and Prof. Udit Gupta. His research focuses on cross-stack optimizations for ML systems for large-scale and resource constrained settings.