Work / Portfolio optimizer, built twice

GPU Portfolio & Risk Decision Engine

A portfolio optimizer built twice, on CPU and GPU, with a parity suite proving both give the same answer.

Bar chart of QP solve time on CPU and GPU at 50, 500 and 3,000 assets, for two covariance models. The CPU solve is faster at 50 and 500 assets; the GPU solve is 2.48 to 2.66 times faster at 3,000.

Role: Solo build
Year: Jul/Aug 2026–present
Status: Built; CPU and GPU paths committed
Stack: Python, RAPIDS cuDF and cuML, NVIDIA cuOpt, CVXPY and Clarabel, pandas, Docker
Links: Repo

Problem

GPU speedups are easy to inflate with a slow baseline. I wanted to know where a GPU actually pays off in portfolio optimization.

What I built

A mean-variance optimizer built twice, once on CPU and once on GPU, sharing one interface. A parity suite proves both paths give the same answer before any timing is trusted, and a benchmark harness times each stage separately.

It includes three risk models, lot rounding with transaction costs, and a backtest with no lookahead.

Result

The QP solve is 2.48–2.66× faster on GPU at 3,000 assets (5.96 s to 2.41 s), with the crossover between 500 and 3,000 assets and parity to 1e-6. Lot rounding beat the greedy baseline in 18 of 18 cases. Only the solve stage wins on GPU; I don't claim an end-to-end speedup.

Next Case Closed