# Int8 Quantization

Meta has released the next generation of Andromeda, its personalized ads retrieval system, designed jointly across model, system and GPU hardware. Retrieval previously could not afford ranking-stage model depth at its scale; the new architecture brings large-language-model-style mixture-of-experts into retrieval, approaching ranking-level performance at roughly 300x fewer inference FLOPs than the 2024 retrieval model. The combined single-GPU pipeline processes over one billion ads per second on one server, which Meta calls a new throughput benchmark for retrieval-stage models, and Meta reports a 2.4% conversion increase on Facebook since 2025.

Meta has released the next generation of Andromeda, its personalized ads retrieval system, designed jointly across model, system and GPU hardware. Retrieval previously could not afford ranking-stage model depth at its scale; the new architecture brings large-language-model-style mixture-of-experts into retrieval, approaching ranking-level performance at roughly 300x fewer inference FLOPs than the 2024 retrieval model. The combined single-GPU pipeline processes over one billion ads per second on one server, which Meta calls a new throughput benchmark for retrieval-stage models, and Meta reports a 2.4% conversion increase on Facebook since 2025.

This work presents a two-branch super-resolution deep-learning framework that simultaneously upscales low-resolution depth and intensity images from a consumer-grade VL53L9CX SPAD time-of-flight sensor, using a two-stage PixelShuffle reconstruction head and simplified sensor firmware to control activation memory, then exports the model to ONNX, quantizes it to INT8, generates C code, and deploys it on a bare-metal STM32H563ZI MCU with only 640 KB SRAM and 2 MB Flash for fully on-device super-resolution inference.

This work presents a two-branch super-resolution deep-learning framework that simultaneously upscales low-resolution depth and intensity images from a consumer-grade VL53L9CX SPAD time-of-flight sensor, using a two-stage PixelShuffle reconstruction head and simplified sensor firmware to control activation memory, then exports the model to ONNX, quantizes it to INT8, generates C code, and deploys it on a bare-metal STM32H563ZI MCU with only 640 KB SRAM and 2 MB Flash for fully on-device super-resolution inference.