ML Platform

• Performance & power optimization for TensorFlow models on Google Tensor chips

• ML Apps: Realtime Camera, Post Shot, and Speech use cases on Pixel phones

• Deep Learning: CNNs (MobileNet, Inception, ResNet) and RNNs (LSTMs) with quantization

• Integration of continuous regression testing with Google Silicon Device Cloud

• Automatic performance & power regression detection and culprit finding infrastructure

• ML Apps: FaceAuth, OCR, Speech Recognition, and Top Shot

• Fixed-point Fast Fourier Transform (FFT) written in C++ (Halide) on Pixel Visual Core for Pixel 2

• HDR+ burst photography for low-light imaging accelerated on Pixel Visual Core for Pixel 3

• Face single-shot detection performance & power optimization on Edge TPU for Pixel 4

• High-performance graph analytics engine for Big Data and recommender systems

• GPU-based ML (e.g., k-means, PageRank) and matrix ops (GEMM, GEMV) with "smart" CUDA kernels

• IoT-optimized graph algorithms > 1,000x more efficient than GraphLab and GraphX in perf / watt

• Large-scale PageRank on hyperlink graph with 128 billion edges < 26 seconds per iteration

• Freebase knowledge graph with 1 billion facts and 106 million entities compressed by 7.5x

• Worldwide GeoIP lookup < 16ns and realtime GeoNames search with 7.2M GPS locations from YAGO

• Online planning & scheduling of reconfigurable printers led to world's fastest duplex cut-sheet printer

• Parallel and external-memory graph search for domain-independent planning and model checking

• Model-based control and planner-assisted active diagnosis

• Anytime heuristic search for combinatorial optimization and sequence alignment