In March of last year (2024), I started an internship at Weaviate (weaviate.io), a vector DB startup. While I originally started at the excellent developer growth team, the applied research team let me join them after I opened a performance improvement PR on the database repo.

I initially worked on speeding up some vector distance calculations (e.g. dot product) using SIMD. At the Berlin Buzzwords fair, I saw a talk by an Nvidia engineer presenting cuvs, a library that runs the vector search algorithm on Nvidia GPUs. I knew that I had to talk to him and approached him afterwards about integrating the library into Weaviate. They were very interested in supporting me, and Weaviate was interested in this feature as well, as they had considered an integration earlier already.

My internship was almost over, but I continued the integration as a student job next to my studies. What followed were a few months in which I got quite familiar with CGO, CUDA and performance optimization, as I contributed a Go wrapper library upstream and added another vector index type to Weaviate. In November, I got asked if I were interested to present my work at GTC, the yearly Nvidia developer conference in San Jose. I was very excited and obviously said yes.

After a few more months of polishing the integration, running benchmarks, and fine-tuning the presentation, I finally flew to SF on 13th of March. After getting to know some people in SF for a few days, I went to the south bay to attend the SemiAnalysis GPU hackathon. GTC itself was a really interesting experience, and the talk went well despite initial doubts.

(Weaviate and NVIDIA teams at GTC)