Simran Arora

Simran Arora

Computer Science PhD Student · Stanford University

Website
Share

Bio

Simran Arora is a Stanford computer science PhD student working on AI systems, GPU kernels, and efficient model execution. She is a co-author of KernelBench and related multi-GPU kernel research, and is associated with Together AI in recent AI-for-science and systems work.

Session (1)

Day 212:05pm-12:25pmExpo Stage 3 SW
Can LLMs write fast multi-GPU kernels? We built a benchmark to find out.

Additional Speaker Details

Using publicly available information we constructed an analysis to help you get a feel for this speaker before deciding to attend their session.

Simran Arora is a Stanford PhD researcher (advised by Chris Ré) working at the intersection of ML systems and model design. She is a core author/lead of hardware-efficient AI projects: ThunderKittens (a CUDA tile-primitive DSL for fast GPU kernels, ~3.5k stars), the BASED attention-free / linear-attention architecture, and KernelBench (an LLM-GPU-kernel benchmark). Her work pushes the Pareto frontier of model quality vs. compute efficiency, with linear-attention ideas adopted across newer architectures (Mamba-v2, RWKV-v5/v6, MiniMax).

Speaking style

Deeply technicalBattle-testedStoryteller

Recent talks (2)

Pareto-efficient AI systems—Simran Arora (Stanford)Paul G. Allen School Colloquium (University of Washington) · Apr 2025Panel Discussion: The AI Kernel Revolution: Rethinking Execution from the Ground UpAI Infra Summit at PyTorch Conference 2025 · Oct 2025

GitHub

@simran-arora

Recent writing (6)