An official website of the United States government
Deep neural network (DNN) accelerators as an example of domain-specific architecture have demonstrated great success in DNN inference. However, the architecture acceleration for equally important DNN training has not yet been fully studied. With data forward, error backward and gradient calculation, DNN training is a more complicated process with higher computation and communication intensity. Because the recent research demonstrates a diminishing specialization return, namely, “accelerator wall”, we believe that a promising approach is to explore coarse-grained parallelism among multiple performance-bounded accelerators to support DNN training. Distributing computations on multiple heterogeneous accelerators to achieve high throughput and balanced execution, however, remaining challenging. We present ACCPAR, a principled and systematic method of determining the tensor partition among heterogeneous accelerator arrays. Compared to prior empirical or unsystematic methods, ACCPAR considers the complete tensor partition space and can reveal previously unknown new parallelism configurations. ACCPAR optimizes the performance based on a cost model that takes into account both computation and communication costs of a heterogeneous execution environment. Hence, our method can avoid the drawbacks of existing approaches that use communication as a proxy of the performance. The enhanced flexibility of tensor partitioning in ACCPAR allows the flexible ratio of computations to be distributed among accelerators with different performances. The proposed search algorithm is also applicable to the emerging multi-path patterns in modern DNNs such as ResNet. We simulate ACCPAR on a heterogeneous accelerator array composed of both TPU-v2 and TPU-v3 accelerators for the training of large-scale DNN models such as Alexnet, Vgg series and Resnet series. The average performance improvements of the state-of-the-art “one weird trick” (OWT) and HYPAR, and ACCPAR, normalized to the baseline data parallelism scheme where each accelerator replicates the model and processes different input data in parallel, are 2.98×, 3.78×, and 6.30×, respectively.
more »
« less
This content will become publicly available on January 1, 2026
MTrain: Enable Efficient CNN Training on Heterogeneous FPGA-Based Edge Servers
FPGA-based edge servers are used in many applications in smart cities, hospitals, retail, etc. Equipped with heterogeneous FPGA-based accelerator cards, the servers can be implemented with multiple tasks including efficient video prepossessing, machine learning algorithm acceleration, etc. These servers are required to implement inference during the daytime while re-training the model during the night to adapt to new environments, domains, or new users. During the re-training, conventionally, the incoming data are transmitted to the cloud, and then the updated machine learning models will be transferred back to the edge server. Such a process is inefficient and cannot protect users’ privacy, so it is desirable for the models to be directly trained on the edge servers. Deploying convolutional neural network (CNN) training on heterogeneous resource-constrained FPGAs is challenging since it needs to consider both the complex data dependency of the training process and the communication bottleneck among different FPGAs. Previous multi-accelerator training algorithms select optimal scheduling strategies for data parallelism, tensor parallelism, and pipeline parallelism. However, pipeline parallelism cannot deal with batch normalization (BN) which is an essential CNN operator, while purely applying data parallelism and tensor parallelism suffers from resource under-utilization and intensive communication costs. In this work, we propose MTrain, a novel multi-accelerator training scheduling strategy that transfers the training process into a multi-branch workflow, thus independent sub-operations of different branches are executed on different training accelerators in parallelism for better utilization and reduced communication overhead. Experimental results show that we can achieve efficient CNN training on heterogeneous FPGA-based edge servers with 1.07x-2.21x speedup under 15 GB/s peer-to-peer bandwidth compared to the state-of-the-art work.
more »
« less
- PAR ID:
- 10578865
- Publisher / Repository:
- IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems
- Date Published:
- Journal Name:
- IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems
- ISSN:
- 0278-0070
- Page Range / eLocation ID:
- 1 to 1
- Subject(s) / Keyword(s):
- edge server heterogeneous FPGAs CNN training
- Format(s):
- Medium: X
- Sponsoring Org:
- National Science Foundation
More Like this
-
-
In recent decades, due to the emerging requirements of computation acceleration, cloud FPGAs have become popular in public clouds. Major cloud service providers, e.g. AWS and Microsoft Azure have provided FPGA computing resources in their infrastructure and have enabled users to design and deploy their own accelerators on these FPGAs. Multi-tenancy FPGAs, where multiple users can share the same FPGA fabric with certain types of isolation to improve resource efficiency, have already been proved feasible. However, this also raises security concerns. Various types of side-channel attacks targeting multi-tenancy FPGAs have been proposed and validated. The awareness of security vulnerabilities in the cloud has motivated cloud providers to take action to enhance the security of their cloud environments. In FPGA security research papers, researchers always perform attacks under the assumption that attackers successfully co-locate with victims and are aware of the existence of victims on the same FPGA board. However, the way to reach this point, i.e., how attack- ers secretly obtain information regarding accelerators on the same fabric, is constantly ignored despite the fact that it is non-trivial and important for attackers. In this paper, we present a novel finger- printing attack to gain the types of co-located FPGA accelerators. We utilize a seemingly non-malicious benchmark accelerator to sniff the communication link and collect performance traces of the FPGA-host communication link. By analyzing these traces, we are able to achieve high classification accuracy for fingerprinting co-located accelerators, which proves that attackers can use our method to perform cloud FPGA accelerator fingerprinting with a high success rate. As far as we know, this is the first paper targeting multi-tenant FPGA accelerator fingerprinting with the communica- tion side-channel.more » « less
-
Genetic programming (GP) is a general, broadly effective procedure by which computable solutions are constructed from high-level objectives. As with other machine-learning endeavors, one continual trend for GP is to exploit ever-larger amounts of parallelism. In this paper, we explore the possibility of accelerating GP by way of modern field-programmable gate arrays (FPGAs), which is motivated by the fact that FPGAs can sometimes leverage larger amounts of both function and data parallelism—common characteristics of GP— when compared to CPUs and GPUs. As a first step towards more general acceleration, we present a preliminary accelerator for the evaluation phase of "tree-based GP"—the original, and still popular, flavor of GP—for which the FPGA dynamically compiles programs of varying shapes and sizes onto a reconfigurable function tree pipeline. Overall, when compared to a recent open-source GPU solution implemented on a modern 8nm process node, our accelerator implemented on an older 20nm FPGA achieves an average speedup of 9.7×. Although our accelerator is 7.9× slower than most examples of a state-of-the-art CPU solution implemented on a recent 7nm process node, we describe future extensions that can make FPGA acceleration provide attractive Pareto-optimal tradeoffs.more » « less
-
Various hardware accelerators have been developed for energy-efficient and real-time inference of neural networks on edge devices. However, most training is done on high-performance GPUs or servers, and the huge memory and computing costs prevent training neural networks on edge devices. This paper proposes a novel tensor-based training framework, which offers orders-of-magnitude memory reduction in the training process. We propose a novel rank-adaptive tensorized neural network model, and design a hardware-friendly low-precision algorithm to train this model. We present an FPGA accelerator to demonstrate the benefits of this training method on edge devices. Our preliminary FPGA implementation achieves 59× speedup and 123× energy reduction compared to embedded CPU, and 292× memory reduction over a standard full-size training.more » « less
-
With the rapid innovation of GPUs, heterogeneous GPU clusters in both public clouds and on-premise data centers have become increasingly commonplace. In this paper, we demonstrate how pipeline parallelism, a technique wellstudied for throughput-oriented deep learning model training, can be used effectively for serving latency-bound model inference, e.g., in video analytics systems, on heterogeneous GPU clusters. Our work exploits the synergy between diversity in model layers and diversity in GPU architectures, which results in comparable inference latency for many layers when running on low-class and high-class GPUs. We explore how such overlooked capability of low-class GPUs can be exploited using pipeline parallelism and present a novel inference serving system, PPipe, that employs pool-based pipeline parallelism via an MILP-based control plane and a data plane that performs resource reservation-based adaptive batching. Evaluation results on diverse workloads (18 CNN models) show that PPipe achieves 41.1%–65.5% higher utilization of low-class GPUs while maintaining high utilization of high-class GPUs, leading to 32.2%–75.1% higher serving throughput compared to various baselines.more » « less
