TLDR · PROGRAMMING, DESIGN & DATA SCIENCE

BEWARE OF MISLEADING GPU VS CPU BENCHMARKS

pythonspeed.com4 min read

Summary from the TLDR newsletter

Nvidia's RAPIDS project claims that cuDF, a GPU-based Pandas replacement, is up to 80 times faster than Pandas. While these speed-ups are impressive, they're also misleading. The benchmarks compare against Pandas running on a single CPU core - while the Pandas library is single-threaded, fixing this doesn't require using a GPU. It also ignores costs - switching to a GPU may make things faster, but they are significantly more expensive. GPUs may still be an appropriate solution for some tasks, but it will require careful experimentation to determine if they are right for the use case and budget.