Debunking the 100X GPU vs. CPU myth

Lee, Victor W.; Kim, Changkyu; Chhugani, Jatin; Deisher, Michael; Kim, Dae Hyun; Nguyen, Anthony D.; Satish, Nadathur; Smelyanskiy, Mikhail; Chennupaty, Srinivas; Hammarlund, Per; Singhal, Ronak; Dubey, Pradeep

doi:10.1145/1815961.1816021

Cited by 341 publications

(27 citation statements)

References 31 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…Moreover, XNN execution on a commodity 4-core Skylake chip (a standard desktop/laptop processor) was only 4-6x slower than this top-end GPU. These observations are consistent with the findings reported in [33], but here we support the claim on a modern CPU.…”

Section: A Gpu Is Not 100x Faster Than a Cpusupporting

confidence: 94%

A Multicore Path to Connectomics-on-Demand

Shavit

2016

Proceedings of the 28th ACM Symposium on Parallelism in Algorithms and Architectures

View full text Add to dashboard Cite

The current design trend in large scale machine learning is to use distributed clusters of CPUs and GPUs with MapReduce-style programming. Some have been led to believe that this type of horizontal scaling can reduce or even eliminate the need for traditional algorithm development, careful parallelization, and performance engineering. This paper is a case study showing the contrary: that the benefits of algorithms, parallelization, and performance engineering, can sometimes be so vast that it is possible to solve "clusterscale" problems on a single commodity multicore machine.Connectomics is an emerging area of neurobiology that uses cutting edge machine learning and image processing to extract brain connectivity graphs from electron microscopy images. It has long been assumed that the processing of connectomics data will require mass storage, farms of CPU/GPUs, and will take months (if not years) of processing time. We present a high-throughput connectomics-ondemand system that runs on a multicore machine with less than 100 cores and extracts connectomes at the terabyte per hour pace of modern electron microscopes.

show abstract

Section: A Gpu Is Not 100x Faster Than a Cpusupporting

confidence: 94%

A Multicore Path to Connectomics-on-Demand

Shavit

2016

Proceedings of the 28th ACM Symposium on Parallelism in Algorithms and Architectures

View full text Add to dashboard Cite

show abstract

“…As stated in [27], it is important to compare optimized code that is running on the GPU to optimized CPU code. For this reason, we compare the massively parallel GPU implementation approach to two non-distributed, non-D&R CPU implementations.…”

Section: Performance Evaluationmentioning

confidence: 99%

Fast Computation of Recurrences in Long Time Series

Rawald

Sips

Marwan

et al. 2014

Springer Proceedings in Mathematics &Amp; Statistics

View full text Add to dashboard Cite

We present an approach to recurrence quantification analysis (RQA) that allows to process very long time series fast. To do so, it utilizes the paradigm Divide and Recombine. We divide the underlying matrix of a recurrence plot (RP) into sub matrices. The processing of the sub matrices is distributed across multiple graphics processing unit (GPU) devices. GPU devices perform RQA computations very fast since they match the problem very well. The individual results of the sub matrices are recombined into a global RQA solution. To address the specific challenges of subdividing the recurrence matrix, we introduce means of synchronization as well as additional data structures. Outperforming existing implementations dramatically, our GPU implementation of RQA processes time series consisting of N ≈ 1, 000, 000 data points in about 5 minutes.

show abstract

“…This is an important feature; by allowing for the final code to execute in a hybrid manner on both the host CPU as well as the attached accelerator device(s), the present approach is not biased towards neither the former nor the latter of the two platforms (making for a fair comparison of the relative CPU vs. GPU throughput 70 ). If the implementation was instead formulated in terms of some low-level language such as CUDA, such a bias would be present.…”

Section: Implementationsmentioning

confidence: 99%

Efficient and portable acceleration of quantum chemical many-body methods in mixed floating point precision using OpenACC compiler directives

Eriksen

2016

Molecular Physics

View full text Add to dashboard Cite

It is demonstrated how the non-proprietary OpenACC standard of compiler directives may be used to compactly and efficiently accelerate the rate-determining steps of two of the most routinely applied many-body methods of electronic structure theory, namely the second-order Møller-Plesset (MP2) model in its resolution-of-the-identity (RI) approximated form and the (T) triples correction to the coupled cluster singles and doubles model (CCSD(T)). By means of compute directives as well as the use of optimized device math libraries, the operations involved in the energy kernels have been ported to graphics processing unit (GPU) accelerators, and the associated data transfers correspondingly optimized to such a degree that the final implementations (using either double and/or single precision arithmetics) are capable of scaling to as large systems as allowed for by the capacity of the host central processing unit (CPU) main memory. The performance of the hybrid CPU/GPU implementations is assessed through calculations on test systems of alanine amino acid chains using one-electron basis sets * To whom correspondence should be addressed 1 arXiv:1609.08094v2 [physics.chem-ph] 17 Jan 2017 of increasing size (ranging from double-to pentuple-ζ quality). For all but the smallest problem sizes of the present study, the optimized accelerated codes (using a single multi-core CPU host node in conjunction with six GPUs) are found to be capable of reducing the total timeto-solution by at least an order of magnitude over optimized, OpenMP-threaded CPU-only reference implementations.2

show abstract

Debunking the 100X GPU vs. CPU myth

Cited by 341 publications

References 31 publications

A Multicore Path to Connectomics-on-Demand

A Multicore Path to Connectomics-on-Demand

Fast Computation of Recurrences in Long Time Series

Efficient and portable acceleration of quantum chemical many-body methods in mixed floating point precision using OpenACC compiler directives

Contact Info

Product

Resources

About