Neural network quantization is a promising compression technique to reduce memory footprint and save energy consumption, potentially leading to real-time inference. However, there is a performance gap between quantized and fullprecision models. To reduce it, existing quantization approaches require highprecision INT32 or full-precision multiplication during inference for scaling or dequantization. This introduces a noticeable cost in terms of memory, speed, and required energy. To tackle these issues, we present F8Net, a novel quantization framework consisting of only fixed-point 8-bit multiplication. To derive our method, we first discuss the advantages of fixed-point multiplication with different formats of fixed-point numbers and study the statistical behavior of the associated fixedpoint numbers. Second, based on the statistical and algorithmic analysis, we apply different fixed-point formats for weights and activations of different layers. We introduce a novel algorithm to automatically determine the right format for each layer during training. Third, we analyze a previous quantization algorithmparameterized clipping activation (PACT)-and reformulate it using fixed-point arithmetic. Finally, we unify the recently proposed method for quantization finetuning and our fixed-point approach to show the potential of our method. We verify F8Net on ImageNet for MobileNet V1/V2 and ResNet18/50. Our approach achieves comparable and better performance, when compared not only to existing quantization techniques with INT32 multiplication or floating-point arithmetic, but also to the full-precision counterparts, achieving state-of-the-art performance.

Our deployment of cache-aware load balancing in the Google web search backend reduced cache misses by ∼0.5x, contributing to a double-digit percentage increase in the throughput of our serving clusters by relieving a bottleneck. This innovation has benefited all production workloads since 2015, serving billions of queries daily. A load balancer forwards each query to one of several identical serving replicas. The replica pulls each term's postings list into RAM from flash, either locally or over the network. Flash bandwidth is a critical bottleneck, motivating an application-directed RAM cache on each replica. Sending the same term reliably to the same replica would increase the chance it hits cache, and avoid polluting the other replicas' caches. However, most queries contain multiple terms and we have to send the whole query to one replica, so it is not possible to achieve a perfect partitioning of terms to replicas. We solve this via a voting scheme, whereby the load balancer conducts a weighted vote by the terms in each query, and sends the query to the winning replica. We develop a multi-stage scalable algorithm to learn these weights. We first construct a large-scale term-query graph from logs and apply a distributed balanced graph partitioning algorithm to cluster each term to a preferred replica. This yields a good but simplistic initial voting table, which we then iteratively refine via cache simulation to capture feedback effects.