Snap, the social media company behind Snapchat, has cut daily data processing costs by 76% after switching its experimental testing infrastructure from traditional processors to graphics processing units (GPUs).
The company runs thousands of controlled experiments each month to refine features for Snapchat's 940 million monthly active users, measuring nearly 6,000 metrics and processing more than 10 petabytes of data within a three-hour morning window.
Snap moved those workloads from CPU-only systems to Nvidia's cuDF, a GPU-accelerated data processing library, running on Google Cloud's Kubernetes Engine, a managed platform for deploying software containers.
The migration, carried out in collaboration with Nvidia engineers, delivered four-times faster processing speeds using the same number of machines, according to data collected by Snap over a two-month period.
Engineers also reduced the number of concurrent GPUs required to 2,100, down from an initial estimate of around 5,500, by optimising pipelines on Google Cloud virtual machines powered by NVIDIA L4 graphics chips.
"Experimentation is at the core of our company," said Prudhvi Vatala, senior engineering manager at Snap. "Changing our data infrastructure from CPUs to GPUs allows us to efficiently scale this experimentation to more features, more metrics and more users over time."
Related reading
- OpenAI releases new faster and cheaper GPT-5.4 models
- Cambridge Aerospace in talks to raise $200 million at unicorn valuation
- Kering creates Kering jewellery unit, names Jean‑Marc Duplaix
Joshua Sambasivam, a backend engineer on the A/B testing team, said the early results exceeded expectations: "When I saw the results of the initial experiments, they were pretty crazy... we saw much higher cost savings than we had expected."
Snap said it plans to extend the GPU accelerator beyond its A/B testing team into wider production workloads and will present details of the migration at NVIDIA's GTC conference.
The recap
- Snap processes over 10 petabytes daily during A/B testing.
- NVIDIA cuDF delivered 4x runtime speedups with same machines.
- Snap reports 76% daily cost savings using NVIDIA GPUs.