Meta-Style AI Recommendations Get Up to 49% Faster
Your online feeds could load faster and cost less to run.
Every time Facebook, Instagram, or a shopping app suggests something you might like, a recommendation AI has to quickly look up information about you and millions of products. Those lookups rely on massive tables of data, which take up huge amounts of memory. To save space, companies use a math trick called Tensor-Train compression, which squeezes those tables into a much smaller form.
But squeezing data makes it slower to search through. A new research paper from Seoul National University introduces FlowTT, a smarter way to run these compressed lookups on GPUs, the specialized chips that power most AI. Instead of repeating the same calculations over and over, FlowTT groups similar lookups together and reuses work they share. It also keeps intermediate results on the chip, rather than sending them back and forth to slower memory, which wastes time and energy.
On benchmarks built from Meta's real recommendation workloads, FlowTT made a big difference. Compared to an existing tool called EcoRec, it cut latency (the delay before a recommendation appears) by up to 42% during inference — when the model is actually serving users — and by up to 49% during training. It also used the least peak memory of any tested method. That means data centers could serve more users with the same hardware, or cut their electricity bills.
The catch: this is research, not a shipping product yet. FlowTT was tested on Meta-style simulated data, not live apps. Still, it points to practical ways to make the AI behind your feed faster, cheaper, and more energy-efficient — with no visible difference except quicker responses and lower costs for the companies behind them.
- FlowTT speeds up the lookup step in recommendation AI by reusing repeated calculations instead of doing them fresh each time.
- Compared to the existing EcoRec method, FlowTT reduced delays by up to 42% for serving recommendations and 49% for training.
- The technique also lowers peak memory use, which could cut server costs and energy consumption in large data centers.
Why It Matters
Faster, cheaper recommendation AI means snappier apps, lower energy bills, and better service for billions of users.