Caffe con Troll: Shallow Ideas to Speed Up Deep Learning

By Firas Abuzaid, Stefan Hadjis, Ce Zhang, Christopher R'e. 2015

74 citations recorded by the source corpus.

We present Caffe con Troll (CcT), a fully compatible end-to-end version of the popular framework Caffe with rebuilt internals. We built CcT to examine the performance characteristics of training and deploying general-purpose convolutional neural networks across different hardware architectures. We find that, by employing standard batching optimizations for CPU training, we achieve a 6:3× throughput improvement over Caffe on popular networks like CaffeNet. Moreover, with these improvements, the end-to-end training time for CNNs is directly proportional to the FLOPS delivered by the CPU, which enables us to efficiently train hybrid CPU-GPU systems for CNNs.

Subfields: Artificial Intelligence, Computer Vision and Pattern Recognition, Electrical and Electronic Engineering

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Artificial Intelligence

1. Wujie Wen
2. Hai Helen Li
3. Yiran Chen
4. M. Franklin
5. Ping Tak Peter Tang
6. Shivaram Venkataraman
7. Sheng R. Li
8. Evan R. Sparks

In Computer Vision and Pattern Recognition

1. Wujie Wen
2. Hai Helen Li
3. Tushar Krishna
4. Yiran Chen
5. Hyoukjun Kwon
6. Atilim Gunes Baydin
7. Alexey Radul
8. Barak A. Pearlmutter

In Electrical and Electronic Engineering

1. Tushar Krishna
2. Hyoukjun Kwon
3. M. Franklin
4. Alexey Radul
5. Hai Helen Li
6. A. Samajdar
7. Yiran Chen
8. B. Recht

Citations in the corpus

• Automatic differentiation in machine learning: a survey (2015, background): “Deep learning systems are generally compute-bound and spend a considerable amount of computation time in highly-optimized numerical kernels for matrix operations (Hadjis et al., 2015; Chetlur et al., 2014).”
• SIGMA: A Sparse and Irregular GEMM Accelerator with Flexible Interconnects for DNN Training (2020, methodology): “For Convolutional Neural Networks (CNNs), GPUs remap the conv operation into a GEMM via the Im2Col operation [18] or other efficient ordering operations.”
• KeystoneML: Optimizing Pipelines for Large-Scale Advanced Analytics (2016, methodology): “eaturization we develop an extensible API in KeystoneML which allows for cost profiling of arbitrary nodes and uses these cost estimates to make node-level and whole-pipeline optimizations. Other work [60, 5] has looked at optimizing caching strategies and operator selection in the regime of feature ”
• Faster CNNs with Direct Sparse Convolutions and Guided Pruning (2016, background): “The lowering process has demonstrated overhead for dense convolution as in Hadjis et al. (2015); Chintala (2015), and is particularly problematic for sparse convolution with intensity already lower than its dense counter part.”
• Evaluating the Energy Efficiency of Deep Convolutional Neural Networks on CPUs and GPUs (2016, background): “The numbers on the bars are values of execution time per iteration.”