The Summer After Freshman Year

Feels like summer went by too fast. We're back at school again, but this time I have my own apartment!

The main thing this summer was my Software/ML internship at TetraMem, but other than that, hanging out with my friends and going to Aruba were definitely my most fond memories.

TetraMem

I remember being so nervous on the first day. Maxwell greeted me at the door and we talked about onboarding stuff over breakfast. I think I was a little too anxious to get started right away, since I sped through the onboarding docs almost instantly. As I was getting a tour of the building, I almost missed my first standup meeting!

I was just about to order McDonalds for lunch when I though, "hmm, maybe they might take me out to lunch on the first day." Luckily, I stayed, since 10 mins later the whole software team told me we'd get lunch. It was nice meeting the other interns and engineers while we ate Sizzling Lunch. But we had to hurry back for Jeremy's big presentation.

After we got back and I finished reading some papers, it was time to choose what direction I'd focus on for the summer. I had two choices: a) Train a neural beamforming model and quantize to <4M params, and b) Optimize Depthwise Convolutions in the compiler. To be honest, I think the first choice would have been more relevant, but I didn't fully understand the paper he gave me and made up some excuse about "not understanding time domain very well"?? What... makes no sense.

So instead, I spent the first few days studying papers on optimizing CNNs for in-memory compute architectures. I also looked into autoschedulers such as Halide and TVM to get a better idea of how to optimize overall dataflow. Who knows, maybe this was a good decision as it introduced me to ML systems and compiles, but also—would it have been better to train AI models?

Fun Times Before Everyone Left

The best part of the internship was when all the OG interns were still there. It was nice to talk to others a bit older than me and see what it's like at other universities. It also felt like "work" during this time period, where we'd attend a bunch of meetings every week, and even presenting our work in our intern midterm.

This was also the time period where Fuey was still taking classes in San Jose, so we'd occasionally meet up for lunch or to get boba, which was really fun.

Eventually, everyone on the semester system left, leaving just me and only one or two other interns. At least we ate roast lamb that day.

The Grind

After the initial excitement began to wear off, I was exhausted. There was so much I didn't know, and working in a codebase as large as our ML Compiler was no easy task, even with the latest AI tools.

Eventually, I got to a point where I could demonstrate via simulation that my algorithm I've been studying could achieve a 16x speedup for depthwise convolutions. Implementing this algorithm would be my biggest contribution, from documenting its specific details to verifying correctness in analog accuracy and scheduling.

And after weeks of head bashing and experiments, I finally wrote an initial implementation in Rust for the frontend and middle-end stages of the compiler. Looking back, I think my progress seemed slower than I would've liked, but mostly because I didn't know anything about compilers or Rust.

The Final Weeks

My last few weeks were spent on polishing my MRs (merge requests) and demonstrating the viability of the algorithm through analog accuracy experiments. The thing is, even though we get a massive speedup, the NPU may struggle to get accurate results due to sparse weight mappings (i.e. it's hard to program zero-valued weights onto the crossbar mapping).

I have more details on my implementation somewhere else, but this is what ended up in my final presentation:

Essentially, by duplicating weights, we can utilize more of our NPU and get a lower cycle count compared to the naive method.

Last Day

This is it. My whole summer of work and this is the day I get to show the whole company what I've been working on.

Every intern does a final presentation in front of the whole company. I'd spent the last few days polishing my presentation and practicing for every possible scenario... except—for being drunk?

Funnily enough, we had important guests come over that day, so we had delicious food and drinks prepared for the whole company. I drank a whole cup of the fruity-looking drink before I realized—whoops, it was alcoholic. That's why everyone looks so happy today.

Being Asian, my face instantly bloomed red. And I definitely felt it for at least 10 mins. But luckily my presentation wasn't for another 30 mins, so I had time to calm down. After drinking lots of water and tea, I was mostly good. Time for my presentation (with extra confidence)!

When I first started, very few people were in the audience, but more people eventually joined and even the CTO showed up. The adrenaline killed my nervousness (as it always does during presentations), so that was great. Even though I probably didn't do a great job explaining everything, I think I demonstrated that my work could have a significant impact on several models planned for future demos. And I got a lot of Q&A afterwards, which means people actually listened?

I was sad to leave to building for the last time. After saying goodbyes to everyone, I was out.

Aruba

Now with the boring work stuff out of the way, let's talk about the fun stuff.

My dad's 50th birthday was in August, and to celebrate, the whole fam went to Aruba for 10 days. In total, we had 17 people, 4 different families, and one giant mansion.

The AirBnB host was nice enough to leave us groceries, drinks (including wine), and even a giant cake for my dad! We spent the first day just chilling in the pool and playing board games. I think this is where my Clash Royale addition started.

Over the week, we did a lot of really fun activities, like going on the "sofa of death" attached to a boat, eating yummy food (including private chef), snorkeling and seeing a sunken ship, drinks on a boat, and even going to a butterfly and donkey farm.

My only complaint was that it was really windy, so you can tell in all the pictures everyone's hair is messed up. I still had a really good time, and can't wait until the next family vacation (hopefully in Japan next, stay tuned...).

Life

Other than that, it was just really nice to be home. There's just so many small things that I realized I spent a whole year deprived of, like great Asian food, clean and hot showers, a nice gym (with steam room), and a giant couch.

Friends

It was also fun to catch up with old friends from high school that I hadn't seen in a while. Apart from seeing my close friends almost every day (whether for poker, basketball, or pickleball), I enjoyed seeing most of my old high school friends at a house party a few weeks into the summer.

We also played basketball every week in our OG "Mmm Cheese" squad.

Cousins

Also, my cousins were back! It's kind of crazy how Kevin's going to JLS and Evelyn's going to Gunn. I'm feeling old now...

After grinding to Mythic in Brawl Stars, watching K-Pop Demon Hunter for the 5th time, and eating lots of steak (paid for with my intern money), I think the summer was pretty fun! I really enjoyed the little moments where we could hang out and not worry about school or anything and just play Smash Bros.

And I'm slowly turning Kevin into a child prodigy, taking him to play basketball everyday and teaching him math + coding. Hopefully we'll get somewhere with him.

Back to School

As I'm writing this, I'm almost two weeks into school. I've been making a better effort to meet new people and join clubs, so we'll see that that goes. I think I'll be in SBHacks and TASA soon.

Here's me moving in with my parents.

The most fun part about having my own apartment is cooking. This is a collection of the things I've been making so far, including Ribeye, Katsu Chicken, and honey-glazed Salmon. Once I get my Apple TV, life will be so great.

Other than that, classes are going well. I should be talking with a PhD student sometime this week to discuss a possible research project related to compute-use AI agents. I was able to find him at a PhD research BBQ (where I somehow managed to fit in), and he said we could be looking at a publication here—it's not for sure obviously, but I'm going to work really hard to try getting it!

Other than that, internship applications for 2026 are starting off slow, but Nvidia Ignite just opened today, which is the one I'm hoping to get most. Time to stop writing this post and start writing some essays!