In this talk, we introduce Merlin HugeCTR. Merlin HugeCTR is an open source, GPU-accelerated integration framework for click-through rate estimation. It optimizes both training and inference, whilst enabling model training at scale with model-parallel embeddings and data-parallel neural networks. In particular, Merlin HugeCTR combines a high-performance GPU embedding cache with an hierarchical storage architecture, to realize low-latency retrieval of embeddings for online model inference tasks. In the MLPerf v1.0 DLRM model training benchmark, Merlin HugeCTR achieves a speedup of up to 24.6x on a single DGX A100 (8x A100) over PyTorch on 4x4-socket CPU nodes (4x4x28 cores). Merlin HugeCTR can also take advantage of multi-node environments to accelerate training even further. Since late 2021, Merlin HugeCTR additionally features a hierarchical parameter server (HPS) and supports deployment via the NVIDIA Triton server framework, to leverage the computational capabilities of GPUs for high-speed recommendation model inference. Using this HPS, Merlin HugeCTR users can achieve a 5~62x speedup (batch size dependent) for popular recommendation models over CPU baseline implementations, and dramatically reduce their end-to-end inference latency.

Differential phase-contrast computed tomography (DPC-CT) is a powerful analysis tool for soft-tissue and low-atomic-number samples. Limited by the implementation conditions, DPC-CT with incomplete projections happens quite often. Conventional reconstruction algorithms face difficulty when given incomplete data. They usually involve complicated parameter selection operations, which are also sensitive to noise and are time-consuming. In this paper, we report a new deep learning reconstruction framework for incomplete data DPC-CT. It involves the tight coupling of the deep learning neural network and DPC-CT reconstruction algorithm in the domain of DPC projection sinograms. The estimated result is not an artifact caused by the incomplete data, but a complete phase-contrast projection sinogram. After training, this framework is determined and can be used to reconstruct the final DPC-CT images for a given incomplete projection sinogram. Taking the sparse-view, limited-view and missing-view DPC-CT as examples, this framework is validated and demonstrated with synthetic and experimental data sets. Compared with other methods, our framework can achieve the best imaging quality at a faster speed and with fewer parameters. This work supports the application of the state-of-the-art deep learning theory in the field of DPC-CT.