28 May 2019 in IEEE Transactions on Intelligent Vehicles
We benchmark several widely-used deep learning frameworks and investigate the FPGA deployment for perform- ing traffic sign classification and detection. We evaluate the training speed and inference accuracy of these frameworks on the GPU by training FPGA-deployment-suitable models with various input sizes on GTSRB, a traffic sign classification dataset. Then, selected trained classification models and various object detection models that we train on GTSRB's detection counterpart (i.e., GTSDB) are evaluated with inference speed, accuracy, and FPGA power efficiency by varying different parameters such as floating-point precisions, batch sizes, etc. We discover that Neon and MXNet deliver the best training speed and classification accuracy on the GPU in general for all test cases, while TensorFlow is always among the frameworks with the highest inference accuracies. We observe that with the cur- rent OpenVINO release, the performance of lightweight models (e.g., MobileNet-v1-SSD, etc) usually exceeds the requirement of real-time detection without losing much accuracy, while other models (e.g., VGG-SSD, ResNet-50-SSD) generally fail to do so. We also demonstrate that we can adjust the precision of bitstreams and the batch sizes to balance inference speed and accuracy of the applications deployed on the FPGA. Finally, we show that for all test cases, the FPGA always achieves higher power efficiency than the GPU.
28 November 2024 in IEEE Transactions on Parallel and Distributed Systems
Characterizing and predicting the training performance of modern machine learning (ML) workloads on compute systems with compute and communication spread between CPUs, GPUs, and network devices is not only the key to optimization and planning but also a complex goal to achieve. The primary challenges include the complexity of synchronization and load balancing between CPUs and GPUs, the variance in input data distribution, and the use of different communication devices and topologies (e.g., NVLink, PCIe, network cards) that connect multiple compute devices, coupled with the desire for flexible training configurations. Built on top of our prior work for single-GPU platforms, we address these challenges and enable multi-GPU performance modeling1 by incorporating (1) data-distribution-aware performance models for embedding table lookup, and (2) data movement prediction of communication collectives, into our upgraded performance modeling pipeline equipped with inter-and intra-rank synchronization for ML workloads trained on multi-GPU platforms. Beyond accurately predicting the per-iteration training time of deep learning recommendation models (DLRM) models with random configurations with a geomean error of 5.21% on two multi-GPU platforms, our prediction pipeline generalizes well to other types of ML workloads, such as Transformer-based natural language processing (NLP) models with a geomean error of 3.00%. Moreover, even without actually running ML workloads like DLRMs on the hardware, it is capable of generating insights such as quickly selecting the fastest embedding table sharding configuration (with a success rate of 85%).