In this white paper, I present my community effort to automatically co-design cheaper, faster and more energy-efficient software and hardware for AI, ML and other popular workloads with the help of the Collective Mind framework (CM), virtualized MLOps, MLPerf benchmarks and reproducible optimization tournaments. I developed CM to modularize, automate and virtualize the tedious process of building, running, profiling and optimizing complex applications across rapidly evolving open-source and proprietary AI/ML models, datasets, software and hardware. I achieved that with the help of portable, reusable and technology-agnostic automation recipes (ResearchOps) for MLOps and DevOps (CM4MLOps) discovered in close collaboration with academia and industry when reproducing more than 150 research papers and organizing the 1st mass-scale community benchmarking of ML and AI systems using CM and MLPerf. I donated CM and CM4MLOps to MLCommons to help connect academia and industry to learn how to build and run AI and other emerging workloads in the most efficient and cost-effective way using a common and technology-agnostic automation, virtualization and reproducibility framework while unifying knowledge exchange, protecting everyone's intellectual property, enabling portable skills, and accelerating transfer of the state-of-the-art research to production. My long-term vision is to make AI accessible to everyone by making it a commodity automatically produced from the most suitable open-source and proprietary components from different vendors based on user demand, requirements and constraints such as cost, latency, throughput, accuracy, energy, size and other important characteristics.

In this paper, we consider the problem of improving the inference latency of language model-based dense retrieval systems by introducing structural compression and model size asymmetry between the context and query encoders. First, we investigate the impact of pre and post-training compression on the MSMARCO, Natural Questions, TriviaQA, SQUAD, and SCIFACT, finding that asymmetry in the dual encoders in dense retrieval can lead to improved inference efficiency. Knowing this, we introduce Kullback Leibler Alignment of Embeddings (KALE), an efficient and accurate method for increasing the inference efficiency of dense retrieval methods by pruning and aligning the query encoder after training. Specifically, KALE extends traditional Knowledge Distillation after bi-encoder training, allowing for effective query encoder compression without full retraining or index generation. Using KALE and asymmetric training, we can generate models which exceed the performance of DistilBERT despite having 3x faster inference.

Vector-based retrieval systems have become a common staple for academic and industrial search applications because they provide a simple and scalable way of extending the search to leverage contextual representations for documents and queries. As these vector-based systems rely on contextual language models, their usage commonly requires GPUs, which can be expensive and difficult to manage. Given recent advances in introducing sparsity into language models for improved inference efficiency, in this paper, we study how sparse language models can be used for dense retrieval to improve inference efficiency. Using the popular retrieval library Tevatron and the MSMARCO, NQ, and TriviaQA datasets, we find that sparse language models can be used as direct replacements with little to no drop in accuracy and up to 4.3x improved inference speeds

Nvidia is touting another win on the latest set of MLPerf benchmarks released Wednesday. The GPU maker said it posted the fastest results on new MLPerf inference benchmarks, which measured the performance of AI inference workloads in data centers and at the edge. MLPerf's five inference benchmarks, applied across four inferencing scenarios, covered AI applications such as image classification, object detection and translation. Nvidia topped all five benchmarks for both data center-focused scenarios (server and offline), with its Turing GPUs. Meanwhile, the Xavier SoC turned in the highest performance among commercially available edge and mobile SoCs that submitted for MLPerf under edge-focused scenarios of single-stream and multi-stream.