Power Profiling in Accel-Sim

August 7, 2025 | 1:18 AM ET

I'll be starting my PhD in a very nice university in 2-3 weeks. It's very exciting!!! I'm looking forward to it a lot. Although, about a month ago, I was asked by my new lab to help them run some simulations on a GPU simulator named Accel-Sim. None of the lab members knew how to use this simulator, so I felt really cool and awesome because I'm their special savior who knows how to use it. Fufufufu... I have the knowledge... I am like a witch from a manga with witches in it like Soul Eater or Witch Hat Atelier or Dorohedoro or Little Witch Academia, where each of the witches are really unique and specialized and they know a very particular topic, but my Witch Specialization is in using the Accel-Sim GPU simulator to run tests... it was cool to be desired in this way!!

My task for the project was to organize the execution of a bunch of different GPU simulator runs for large language models. (I'll say later about my thoughts on LLMs, so if you have complicated feelings on their research in regards to computer architecture, then save those thoughts for later! We likely share the same opinions.) They wanted to show that their studied approach works for different large language models, so I had to (1) collect hardware traces executing large language models, (2) simulate them in the simulator to get per-cycle power traces, (3) simulate other non-LLM programs which are made-up but extenuate the effects of our approach to describe theoretical limits and get those power traces also, and (4) process these results to determine whether more experiments needed to be run.

In particular, that part about running the made-up programs was particularly difficult; I needed to run a whole bunch GPU programs, collect traces for them, and simulate them, with each one taking upwards 2 hours to trace-collect and another 4 hours to simulate... and there were 60 programs to run!!! With only a week until the deadline, it was quite the time crunch!! Thankfully, I do like the hustle of research and the last-minute time crunches, they can be very exciting and rewarding if you succeed! All data was collected in time, with even a few days to spare.

I spent the last few days awake until the early morning, getting a few hours of sleep until I needed to work my software engineering job, spending the night processing data and proofreading our paper. To make matters worse, I needed to proofread another paper: a paper which was rejected from another conference in May, which we re-submitted to this same conference... so I needed to help polish *two* papers!! The second paper also had a really time-consuming figure needing to be made, which took 7 hours to make!! (4 hours for one draft, 3 hours for the final draft...)

The first paper was submitted nicely and proof-read to perfection, while the second had some formatting issues last-minute. We decided to ignore the compile warnings for the paper, which complained about margin-sizes 100 millimeters too large, and just submit it anyways. (Hopefully the reviewers are kind!) Both papers got submitted successfully, and I'm very happy how they turned out.

I was invited to another project on Monday, which sounds very exciting too. We'll submit that paper in November, so there's lots of time to work on it. More details on that to come!!

-Sophie

(P.S.: this blog post was automatically generated by a markdown-to-Python converter I made. I'll be updating the rest of the website soon!)