SVP of Research at Together AI, working on model efficiency and performance. She is an incoming Caltech assistant professor, with the appointment beginning in 2027.[4][1][2]
Start with Frontier LLMs can’t write fast multi-GPU kernels (yet) →
Education
BSE, Computer Science; Jerome Fisher Management and Technology program
2015 – 2019
[2]Links
Website CV
X · simran_s_arora Advisor · Cartesia AI
Advisor · Engram
Advisor · Pearl
Advisor · Fearn AI
SVP of Research · Together AI
Career
Contributions
Test communication as well as arithmetic
ParallelKernelBench asks models to replace distributed PyTorch and NCCL reference implementations with custom multi-GPU kernels. Its problems specify both computation and device topology. Evaluation separates correctness from wall-clock speed and compares performance with the communication limits of the hardware. This catches a gap that single-GPU tests miss: fast arithmetic is insufficient if transfers between devices dominate the task. Arora coauthored the benchmark article with Willy Chan, Nathan Paek, Simon Guo and Daniel Fu.[5]
Choose how work moves between devices
Her kernel-programming talk compares hardware transfer mechanisms and parallelism strategies. Copy engines, device-initiated communication and computation compete for different resources; overlapping them requires understanding the workload and interconnect. The examples connect a profiler’s idle periods with decisions about data movement rather than treating a generated kernel’s syntax as evidence of speed.[6]
Tools, writing & talks
Writing, research & talks
Start here
The coauthored ParallelKernelBench article explains its workload coverage, correctness tests and communication-aware performance evaluation.
Start with the benchmark article’s reference task and evaluation diagram. Familiarity with GPU kernels and distributed training helps; distinguish a correct kernel from one that improves measured runtime.
Coauthor[5]Interconnect scaling has fallen far behind tensor-core throughput, leaving PyTorch plus NCCL under half its communication-aware roofline. ParallelKittens closes that with a dozen lines per kernel; ParallelKernelBench then measures how badly frontier models still fail at writing them.
Catalog workOriginal source- LLMs Write Fast Multi-GPU Kernelsai.engineer
A talk explaining hardware communication mechanisms, distributed workload strategies and kernel evaluation.
Speaker[6]