Google releases ML Drift GPU inference engine as open-source

in production, youtube shorts saw up to 40% lower average frame latency for real‑time visual effects, google photos unblur gained up to 2‑second speedup, adobe lightroom and photoshop reported up to 30% faster on‑device performance, and snap achieved 30% lower model latency. stable diffusion 1.4 peak gpu memory dropped from 4.31 gb to 387 mb, a 93% reduction. ml drift also achieved up to 12% lower memory overhead than llama.cpp and mlx lm on workstation hardware.

the engine is cross‑platform, supporting android, ios, windows, linux, and web via chrome’s built‑in ai apis. it uses a unified shader model built on ucl, co‑optimized with arm, intel, and qualcomm gpus, and includes an agentic coding guide for custom shaders. the release signals that developers can now use a production‑validated gpu inference layer without patent risk, while the legacy tflite gpu delegate will not receive further updates.