Search

Barroso's seminal contributions in energy-proportional warehouse-scale computing launched an era where modern datacenters have become more energy efficient and cost effective than ever before. At the same time, modern AI applications have driven ever-increasing demands in computing, highlighting the importance of optimizing efficiency across the entire deep learning model development cycle. This paper characterizes the carbon impact of AI, including both operational carbon emissions from training and inference as well as embodied carbon emissions from datacenter construction and hardware manufacturing. We highlight key efficiency optimization opportunities for cutting-edge AI technologies, from deep learning recommendation models to multi-modal generative AI tasks. To scale AI sustainably, we must also go beyond efficiency and optimize across the life cycle of computing infrastructures, from hardware manufacturing to datacenter operations and end-of-life processing for the hardware.

A significant fraction (5-15%) of renewable energy generated goes into waste in the grids around the world today due to oversupply issues and transmission constraints. Being able to predict when and where renewable curtailment occurs would improve renewable utilization. The core of this work is to enable the machine learning community to help decarbonize electricity grids by unlocking the potential of renewable energy through curtailment prediction., Comment: The work was presented as a part of the Climate Change AI workshop at NeurIPS 2023

Training large-scale machine learning models poses distinct system challenges, given both the size and complexity of today's workloads. Recently, many organizations training state-of-the-art Generative AI models have reported cases of instability during training, often taking the form of loss spikes. Numeric deviation has emerged as a potential cause of this training instability, although quantifying this is especially challenging given the costly nature of training runs. In this work, we develop a principled approach to understanding the effects of numeric deviation, and construct proxies to put observations into context when downstream effects are difficult to quantify. As a case study, we apply this framework to analyze the widely-adopted Flash Attention optimization. We find that Flash Attention sees roughly an order of magnitude more numeric deviation as compared to Baseline Attention at BF16 when measured during an isolated forward pass. We then use a data-driven analysis based on the Wasserstein Distance to provide upper bounds on how this numeric deviation impacts model weights during training, finding that the numerical deviation present in Flash Attention is 2-5 times less significant than low-precision training.

We present LayerSkip, an end-to-end solution to speed-up inference of large language models (LLMs). First, during training we apply layer dropout, with low dropout rates for earlier layers and higher dropout rates for later layers, and an early exit loss where all transformer layers share the same exit. Second, during inference, we show that this training recipe increases the accuracy of early exit at earlier layers, without adding any auxiliary layers or modules to the model. Third, we present a novel self-speculative decoding solution where we exit at early layers and verify and correct with remaining layers of the model. Our proposed self-speculative decoding approach has less memory footprint than other speculative decoding approaches and benefits from shared compute and activations of the draft and verification stages. We run experiments on different Llama model sizes on different types of training: pretraining from scratch, continual pretraining, finetuning on specific data domain, and finetuning on specific task. We implement our inference solution and show speedups of up to 2.16x on summarization for CNN/DM documents, 1.82x on coding, and 2.0x on TOPv2 semantic parsing task., Comment: Code open sourcing is in progress

Large Language Models (LLMs) with hundreds of billions of parameters have transformed the field of machine learning. However, serving these models at inference time is both compute and memory intensive, where a single request can require multiple GPUs and tens of Gigabytes of memory. Multi-Head Attention is one of the key components of LLMs, which can account for over 50% of LLMs memory and compute requirement. We observe that there is a high amount of redundancy across heads on which tokens they pay attention to. Based on this insight, we propose Clustered Head Attention (CHAI). CHAI combines heads with a high amount of correlation for self-attention at runtime, thus reducing both memory and compute. In our experiments, we show that CHAI is able to reduce the memory requirements for storing K,V cache by up to 21.4% and inference time latency by up to 1.73x without any fine-tuning required. CHAI achieves this with a maximum 3.2% deviation in accuracy across 3 different models (i.e. OPT-66B, LLAMA-7B, LLAMA-33B) and 5 different evaluation datasets.

Recommendation algorithms have been pivotal in handling the overwhelming volume of online content. However, these algorithms seldom consider direct user input, resulting in superficial interaction between them. Efforts have been made to include the user directly in the recommendation process through conversation, but these systems too have had limited interactivity. Recently, Large Language Models (LLMs) like ChatGPT have gained popularity due to their ease of use and their ability to adapt dynamically to various tasks while responding to feedback. In this paper, we investigate the effectiveness of ChatGPT as a top-n conversational recommendation system. We build a rigorous pipeline around ChatGPT to simulate how a user might realistically probe the model for recommendations: by first instructing and then reprompting with feedback to refine a set of recommendations. We further explore the effect of popularity bias in ChatGPT's recommendations, and compare its performance to baseline models. We find that reprompting ChatGPT with feedback is an effective strategy to improve recommendation relevancy, and that popularity bias can be mitigated through prompt engineering.

As the development of large-scale Generative AI models evolve beyond text (1D) generation to include image (2D) and video (3D) generation, processing spatial and temporal information presents unique challenges to quality, performance, and efficiency. We present the first work towards understanding this new system design space for multi-modal text-to-image (TTI) and text-to-video (TTV) generation models. Current model architecture designs are bifurcated into 2 categories: Diffusion- and Transformer-based models. Our systematic performance characterization on a suite of eight representative TTI/TTV models shows that after state-of-the-art optimization techniques such as Flash Attention are applied, Convolution accounts for up to 44% of execution time for Diffusion-based TTI models, while Linear layers consume up to 49% of execution time for Transformer-based models. We additionally observe that Diffusion-based TTI models resemble the Prefill stage of LLM inference, and benefit from 1.1-2.5x greater speedup from Flash Attention than Transformer-based TTI models that resemble the Decode phase. Since optimizations designed for LLMs do not map directly onto TTI/TTV models, we must conduct a thorough characterization of these workloads to gain insights for new optimization opportunities. In doing so, we define sequence length in the context of TTI/TTV models and observe sequence length can vary up to 4x in Diffusion model inference. We additionally observe temporal aspects of TTV workloads pose unique system bottlenecks, with Temporal Attention accounting for over 60% of total Attention time. Overall, our in-depth system performance characterization is a critical first step towards designing efficient and deployable systems for emerging TTI/TTV workloads., Comment: Published at 2024 IEEE International Symposium on Performance Analysis of Systems and Software (ISPASS)

As Machine Learning (ML) systems continue to grow, the demand for relevant and comprehensive datasets becomes imperative. There is limited study on the challenges of data acquisition due to ad-hoc processes and lack of consistent methodologies. We first present an investigation of current data marketplaces, revealing lack of platforms offering detailed information about datasets, transparent pricing, standardized data formats. With the objective of inciting participation from the data-centric AI community, we then introduce the DAM challenge, a benchmark to model the interaction between the data providers and acquirers. The benchmark was released as a part of DataPerf. Our evaluation of the submitted strategies underlines the need for effective data acquisition strategies in ML.

The increasing integration of renewable energy sources results in fluctuations in carbon intensity throughout the day. To mitigate their carbon footprint, datacenters can implement demand response (DR) by adjusting their load based on grid signals. However, this presents challenges for private datacenters with diverse workloads and services. One of the key challenges is efficiently and fairly allocating power curtailment across different workloads. In response to these challenges, we propose the Carbon Responder framework. The Carbon Responder framework aims to reduce the carbon footprint of heterogeneous workloads in datacenters by modulating their power usage. Unlike previous studies, Carbon Responder considers both online and batch workloads with different service level objectives and develops accurate performance models to achieve performance-aware power allocation. The framework supports three alternative policies: Efficient DR, Fair and Centralized DR, and Fair and Decentralized DR. We evaluate Carbon Responder polices using production workload traces from a private hyperscale datacenter. Our experimental results demonstrate that the efficient Carbon Responder policy reduces the carbon footprint by around 2x as much compared to baseline approaches adapted from existing methods. The fair Carbon Responder policies distribute the performance penalties and carbon reduction responsibility fairly among workloads.

Training and deploying large-scale machine learning models is time-consuming, requires significant distributed computing infrastructures, and incurs high operational costs. Our analysis, grounded in real-world large model training on datacenter-scale infrastructures, reveals that 14~32% of all GPU hours are spent on communication with no overlapping computation. To minimize this outstanding communication latency and other inherent at-scale inefficiencies, we introduce an agile performance modeling framework, MAD-Max. This framework is designed to optimize parallelization strategies and facilitate hardware-software co-design opportunities. Through the application of MAD-Max to a suite of real-world large-scale ML models on state-of-the-art GPU clusters, we showcase potential throughput enhancements of up to 2.24x for pre-training and up to 5.2x for inference scenarios, respectively., Comment: ISCA 2024

Deep learning recommendation systems serve personalized content under diverse tail-latency targets and input-query loads. In order to do so, state-of-the-art recommendation models rely on terabyte-scale embedding tables to learn user preferences over large bodies of contents. The reliance on a fixed embedding representation of embedding tables not only imposes significant memory capacity and bandwidth requirements but also limits the scope of compatible system solutions. This paper challenges the assumption of fixed embedding representations by showing how synergies between embedding representations and hardware platforms can lead to improvements in both algorithmic- and system performance. Based on our characterization of various embedding representations, we propose a hybrid embedding representation that achieves higher quality embeddings at the cost of increased memory and compute requirements. To address the system performance challenges of the hybrid representation, we propose MP-Rec -- a co-design technique that exploits heterogeneity and dynamic selection of embedding representations and underlying hardware platforms. On real system hardware, we demonstrate how matching custom accelerators, i.e., GPUs, TPUs, and IPUs, with compatible embedding representations can lead to 16.65x performance speedup. Additionally, in query-serving scenarios, MP-Rec achieves 2.49x and 3.76x higher correct prediction throughput and 0.19% and 0.22% better model quality on a CPU-GPU system for the Kaggle and Terabyte datasets, respectively.

Machine learning research has long focused on models rather than datasets, and prominent datasets are used for common ML tasks without regard to the breadth, difficulty, and faithfulness of the underlying problems. Neglecting the fundamental importance of data has given rise to inaccuracy, bias, and fragility in real-world applications, and research is hindered by saturation across existing dataset benchmarks. In response, we present DataPerf, a community-led benchmark suite for evaluating ML datasets and data-centric algorithms. We aim to foster innovation in data-centric AI through competition, comparability, and reproducibility. We enable the ML community to iterate on datasets, instead of just architectures, and we provide an open, online platform with multiple rounds of challenges to support this iterative development. The first iteration of DataPerf contains five benchmarks covering a wide spectrum of data-centric techniques, tasks, and modalities in vision, speech, acquisition, debugging, and diffusion prompting, and we support hosting new contributed benchmarks from the community. The benchmarks, online evaluation platform, and baseline implementations are open source, and the MLCommons Association will maintain DataPerf to ensure long-term benefits to academia and industry., Comment: NeurIPS 2023 Datasets and Benchmarks Track

We propose RecShard, a fine-grained embedding table (EMB) partitioning and placement technique for deep learning recommendation models (DLRMs). RecShard is designed based on two key observations. First, not all EMBs are equal, nor all rows within an EMB are equal in terms of access patterns. EMBs exhibit distinct memory characteristics, providing performance optimization opportunities for intelligent EMB partitioning and placement across a tiered memory hierarchy. Second, in modern DLRMs, EMBs function as hash tables. As a result, EMBs display interesting phenomena, such as the birthday paradox, leaving EMBs severely under-utilized. RecShard determines an optimal EMB sharding strategy for a set of EMBs based on training data distributions and model characteristics, along with the bandwidth characteristics of the underlying tiered memory hierarchy. In doing so, RecShard achieves over 6 times higher EMB training throughput on average for capacity constrained DLRMs. The throughput increase comes from improved EMB load balance by over 12 times and from the reduced access to the slower memory by over 87 times.

Technology companies have been leading the way to a renewable energy transformation, by investing in renewable energy sources to reduce the carbon footprint of their datacenters. In addition to helping build new solar and wind farms, companies make power purchase agreements or purchase carbon offsets, rather than relying on renewable energy every hour of the day, every day of the week (24/7). Relying on renewable energy 24/7 is challenging due to the intermittent nature of wind and solar energy. Inherent variations in solar and wind energy production causes excess or lack of supply at different times. To cope with the fluctuations of renewable energy generation, multiple solutions must be applied. These include: capacity sizing with a mix of solar and wind power, energy storage options, and carbon aware workload scheduling. However, depending on the region and datacenter workload characteristics, the carbon-optimal solution varies. Existing work in this space does not give a holistic view of the trade-offs of each solution and often ignore the embodied carbon cost of the solutions. In this work, we provide a framework, Carbon Explorer, to analyze the multi-dimensional solution space by taking into account operational and embodided footprint of the solutions to help make datacenters operate on renewable energy 24/7. The solutions we analyze include capacity sizing with a mix of solar and wind power, battery storage, and carbon aware workload scheduling, which entails shifting the workloads from times when there is lack of renewable supply to times with abundant supply., Comment: Published at ASPLOS'23: Proceedings of the 28th ACM International Conference on Architectural Support for Programming Languages and Operating Systems, Volume 2

This paper explores the environmental impact of the super-linear growth trends for AI from a holistic perspective, spanning Data, Algorithms, and System Hardware. We characterize the carbon footprint of AI computing by examining the model development cycle across industry-scale machine learning use cases and, at the same time, considering the life cycle of system hardware. Taking a step further, we capture the operational and manufacturing carbon footprint of AI computing and present an end-to-end analysis for what and how hardware-software design and at-scale optimization can help reduce the overall carbon footprint of AI. Based on the industry experience and lessons learned, we share the key challenges and chart out important development directions across the many dimensions of AI. We hope the key messages and insights presented in this paper can inspire the community to advance the field of AI in an environmentally-responsible manner.

The memory capacity of embedding tables in deep learning recommendation models (DLRMs) is increasing dramatically from tens of GBs to TBs across the industry. Given the fast growth in DLRMs, novel solutions are urgently needed, in order to enable fast and efficient DLRM innovations. At the same time, this must be done without having to exponentially increase infrastructure capacity demands. In this paper, we demonstrate the promising potential of Tensor Train decomposition for DLRMs (TT-Rec), an important yet under-investigated context. We design and implement optimized kernels (TT-EmbeddingBag) to evaluate the proposed TT-Rec design. TT-EmbeddingBag is 3 times faster than the SOTA TT implementation. The performance of TT-Rec is further optimized with the batched matrix multiplication and caching strategies for embedding vector lookup operations. In addition, we present mathematically and empirically the effect of weight initialization distribution on DLRM accuracy and propose to initialize the tensor cores of TT-Rec following the sampled Gaussian distribution. We evaluate TT-Rec across three important design space dimensions -- memory capacity, accuracy, and timing performance -- by training MLPerf-DLRM with Criteo's Kaggle and Terabyte data sets. TT-Rec achieves 117 times and 112 times model size compression, for Kaggle and Terabyte, respectively. This impressive model size reduction can come with no accuracy nor training time overhead as compared to the uncompressed baseline., Comment: To appear in Conference on Machine Learning and Systems (MlSys 2021, https://mlsys.org/)

The use of GPUs has proliferated for machine learning workflows and is now considered mainstream for many deep learning models. Meanwhile, when training state-of-the-art personal recommendation models, which consume the highest number of compute cycles at our large-scale datacenters, the use of GPUs came with various challenges due to having both compute-intensive and memory-intensive components. GPU performance and efficiency of these recommendation models are largely affected by model architecture configurations such as dense and sparse features, MLP dimensions. Furthermore, these models often contain large embedding tables that do not fit into limited GPU memory. The goal of this paper is to explain the intricacies of using GPUs for training recommendation models, factors affecting hardware efficiency at scale, and learnings from a new scale-up GPU server design, Zion., Comment: To appear in IEEE International Symposium on High-Performance Computer Architecture (HPCA 2021)