CoChat
Sign Up
Cited by 569Parallel Computing and Optimization TechniquesDistributed and Parallel Computing SystemsEmbedded Systems Design Techniques

The DOI matches its Crossref record; 199 of the 199 references it cites were found in a registry or an index; 5 of the 49 sources its claims cite were checked for support.

A Survey of CPU-GPU Heterogeneous Computing Techniques

Open access
Mittal, Sparsh, Vetter, Jeffrey S.·2015·ACM Computing Surveys
DOI 10.1145/2788396
Open this paper in CoChat — free accountOpen PDF

Abstract

As both CPUs and GPUs become employed in a wide range of applications, it has been acknowledged that both of these Processing Units (PUs) have their unique features and strengths and hence, CPU-GPU collaboration is inevitable to achieve high-performance computing. This has motivated a significant amount of research on heterogeneous computing techniques, along with the design of CPU-GPU fused chips and petascale heterogeneous supercomputers. In this article, we survey Heterogeneous Computing Techniques (HCTs) such as workload partitioning that enable utilizing both CPUs and GPUs to improve performance and/or energy efficiency. We review heterogeneous computing approaches at runtime, algorithm, programming, compiler, and application levels. Further, we review both discrete and fused CPU-GPU systems and discuss benchmark suites designed for evaluating Heterogeneous Computing Systems (HCSs). We believe that this article will provide insights into the workings and scope of applications of HCTs to researchers and motivate them to further harness the computational powers of CPUs and GPUs to achieve the goal of exascale performance.

Overview

AI snapshot

Written from the paper’s full text.

TL;DR

This survey reviews CPU-GPU heterogeneous computing techniques (HCTs) that use both processing units to improve performance and/or energy efficiency. It classifies approaches across algorithm/runtime, programming/compiler, energy, fused-vs-discrete architecture, application-domain, and benchmark perspectives.

Key findings

  • HCTs are classified by scheduling nature (dynamic vs. static) and workload-partitioning basis, including relative PU performance/load balancing, subtask parallelism or characteristics, floating-point precision, and GPU memory-size limits; pipelining and MapReduce use are also identified.
  • Hardware trends show CPU core counts rising to 8–32 cores with >60 likely, CUDA cores increasing from 240 (GTX 280) to 3072 (GTX TITAN X), CPU LLC growing from 32MB (POWER7) to 96MB (POWER8), and GPU L2 caches from 768KB (Fermi) to 2048KB (Maxwell); NVLink promises 5–12× PCIe Gen3 bandwidth.
  • Fused CPU-GPU systems reduce data-transfer overhead: Daga et al. report an AMD E-series Zacate APU with only 80 GPU cores providing >3× improvement over a discrete Radeon HD 5870 with 1600 more powerful GPU cores; Spafford et al. find fused systems can have better energy efficiency due to low transfer overhead, but higher CPU-GPU contention and lower compute-intensive CPU performance than discrete designs.
  • Energy-saving approaches include intelligent workload partitioning, DVFS on CPU and/or GPU, and resource scaling/low-power modes; benchmark suites include SHOC, Parboil, Rodinia, and Valar. The June 2014 Green500 top 15 systems are all CPU-GPU heterogeneous, and Yang et al. achieve 0.8 PetaFLOP on Tianhe-1A for global atmospheric simulation.

Methods

  • Literature survey of heterogeneous computing techniques using both CPU and GPU, covering discrete and fused CPU-GPU architectures; scope excludes FPGA-based heterogeneous systems, GPU-only performance works, load balancing within a single GPU or across multiple GPUs only, and circuit/device/microarchitectural research.
  • Works are organized by abstraction level and theme: algorithm/runtime workload partitioning, programming languages/frameworks/compiler techniques, energy efficiency, fused CPU-GPU chips, application domains, and benchmark suites, with classification tables.
  • No primary experiments are conducted; the paper synthesizes cited research and reports their evaluation platforms and performance/energy results where available.

Limitations

  • The paper states scope limitations: it includes only CPU-GPU systems/techniques, excludes FPGA-based heterogeneous systems, GPU-only works, single-GPU or multi-GPU-only load balancing, and circuit/device/microarchitectural works; it focuses on key research ideas while only mentioning CPU/GPU platforms used in cited experiments.
  • Evident from the text: as a survey, it provides no new experimental validation, and its classifications and conclusions depend on the selected literature and reported results, which may not generalize across all HCSs.
  • The paper also identifies open needs: fully automatic toolchains, lower-overhead porting of existing CPU/GPU codes, order-of-magnitude energy-efficiency improvement for mobile systems, and more extensive evaluation/development of fused HCSs.

Model-written summary. Check it against the paper before relying on it.

Integrity signals

Registry

Crossref · 10.1145/2788396

Cites

199 of 199 references verified0 citation errors · 0 not checkable

Statistics

Nothing to recompute.

Verification

Claims the paper makes

5 of 49 cited sources checked · 4 supported · 0 flagged · accuracy not yet measured

  • Interpretationcitation

    Different amount of work-divisions to CPU and GPU can lead to vastly different performance.

    “Due to this, different amount of work-divisions to CPU and GPU can lead to vastly different performance [Grewe and O'Boyle 2011].”
    Quote in the paperSubstantial claim1 of 1 sources checked
    • cites Grewe and O'Boyle 2011 source
  • Interpretationcitation

    A heterogeneous system can provide high performance for a much wider variety of applications and usage scenarios than using either CPU or GPU alone.

    “Thus, a heterogeneous system can provide high performance for a much wider variety of applications and usage scenarios than using either of CPU or GPU alone [Vetter and Mittal 2015].”
    Quote in the paperSubstantial claim1 of 1 sources checked
    • cites Vetter and Mittal 2015 source

      The Introduction says heterogeneous computing has increased computational rates in leading supercomputers. Section 2 describes CPU-like and GPU-like cores whose balance depends on workload. Neither passage establishes that a combined system delivers high performance across a much wider variety of applications and usage scenarios than either CPU or GPU alone. (LLM, conf=0.84)

      Experimental — accuracy not yet measured
  • Premisecitation

    OpenCL facilitates both task-parallelism and data-parallelism.

    “It facilitates both task-parallelism and data-parallelism.”
    Quote in the paperNot a substantial claim1 of 1 sources checked
    • cites Stone et al. 2010 source
  • Findingcitation

    A large fraction of TOP500 and Green500 supercomputers now use both CPUs and GPUs.

    “The importance of heterogeneous computing can be seen from the fact that an astonishingly large fraction of TOP500 and Green500 supercomputers now use both CPUs and GPUs [Green500 2014;Top500 2014].”
    Quote in the paperNot a substantial claim0 of 2 sources checked
    • cites Green500 2014

      Couldn’t be matched to a paper in Crossref, so it can’t be checked.

    • cites Top500 2014

      Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  • Findingcitation

    The June 2014 Green500 list shows all top 15 systems are CPU-GPU heterogeneous systems.

    “As a definite case in point, the June 2014 Green500 list of most energy efficient supercomputers shows that all the top 15 systems in the list are CPU-GPU heterogeneous systems [Green500 2014].”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites Green500 2014

      Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  • It may state two separate claims as one. 26
  • Probably not a substantial claim. 27

  • It may state two separate claims as one.

    Premisecitation

    OpenCL provides a higher abstraction programming framework and enables the programmer to write programs which are portable across a wide variety of processing units, although it may not achieve the highest possible performance on a PU.

    “It provides a higher abstraction programming framework and hence, it enables the programmer to write programs which are portable across wide variety of processing units, although it may not be able to achieve the highest possible performance on a PU.”
    Quote in the paperNot a substantial claim1 of 1 sources checked
    • cites Stone et al. 2010 source

  • Probably not a substantial claim. The paper’s conclusions may not depend on it. It may state two separate claims as one.

    Premisecitation

    OpenCL is an open standard and has been adopted by several vendors.

    “OpenCL [Stone et al. 2010] is an open standard and has been adopted by several vendors.”
    Quote in the paperNot a substantial claim1 of 1 sources checked
    • cites Stone et al. 2010 source

  • It may state two separate claims as one.

    Findingcitation

    Earlier GPUs only provided software-managed caches and not hardware-managed caches; as GPUs move towards general-purpose computing, they now feature increasingly larger-sized hardware-managed caches.

    “Earlier GPUs only provided softwaremanaged caches and not hardware-managed caches, however, as the GPUs move towards general-purpose computing, they are now featuring increasingly larger-sized hardware-managed caches.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites Mittal 2014a

      Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  • Probably not a substantial claim. The paper’s conclusions may not depend on it. It may state two separate claims as one.

    Findingcitation

    GT200 architecture GPU did not have an L2 cache; Fermi has 768KB LLC; Kepler has 1536KB LLC; Maxwell has 2048KB LLC.

    “For example, GT200 architecture GPU did not have an L2 cache, the Fermi GPU has 768KB LLC, the Kepler GPU has 1536KB LLC and the Maxwell GPU has 2048KB LLC [Mittal 2014a].”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites Mittal 2014a

      Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  • Probably not a substantial claim. The paper’s conclusions may not depend on it. It may state two separate claims as one.

    Findingcitation

    The 45nm POWER7 processor had 32MB eDRAM LLC, the 32nm POWER7+ had 80MB eDRAM LLC, and the 22nm POWER8 has 96MB eDRAM LLC.

    “The 45nm POWER7 processor had 32MB eDRAM (embedded DRAM) LLC [Wendel et al. 2010], the 32nm POWER7+ processor had an 80MB eDRAM LLC [Zyuban et al. 2013], while the 22nm POWER8 processor has a 96MB eDRAM LLC [Fluhr et al. 2014].”
    Quote in the paperNot a substantial claim0 of 3 sources checked
    • cites Wendel et al. 2010 source
    • cites Zyuban et al. 2013 source
    • cites Fluhr et al. 2014 source

  • Probably not a substantial claim. The paper’s conclusions may not depend on it. It may state two separate claims as one.

    Findingcitation

    Few years ago, CPUs had nearly 1B transistors; the recently announced Oracle SPARC M7 CPU will have more than 10B transistors on chip.

    “Few years ago, CPUs had nearly 1B transistors [Wendel et al. 2010] while the recently announced Oracle SPARC M7 CPU will have more than 10B transistors on chip [Morgan 2014].”
    Quote in the paperNot a substantial claim0 of 2 sources checked
    • cites Wendel et al. 2010 source
    • cites Morgan 2014

      Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  • It may state two separate claims as one.

    Findingreasoning

    For GPU programming, CUDA is most widely used, which is due to its similarity with C/C++ and ability to write optimized code.

    “From Table III, it is clear that for GPU programming, CUDA is most widely used, which is due to its similarity with C/C++ and ability to write optimized code.”
    Quote in the paperNot a substantial claim

  • It may state two separate claims as one.

    Findingreasoning

    Both CPUs and GPUs are increasingly moving towards use of 3D stacking; Intel's Knights Landing and NVIDIA's Pascal GPU will feature 3D stacked memory.

    “3D stacking facilitates high bandwidth and memory capacity [Poremba et al. 2015] and hence, both CPUs and GPUs are increasingly moving towards use of 3D stacking, for example, Intel's Knights Landing [Gardner 2014] and NVIDIA's Pascal GPU [Gupta 2014] will feature 3D stacked memory.”
    Quote in the paperNot a substantial claim0 of 2 sources checked
    • cites Gardner 2014

      Couldn’t be matched to a paper in Crossref, so it can’t be checked.

    • cites Gupta 2014

      Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  • Probably not a substantial claim. The paper’s conclusions may not depend on it.

    Findingcitation

    Many processors with CPU and GPU on the same chip have already been produced, such as AMD Llano, Intel Sandy Bridge, and Ivy Bridge.

    “An even closer level of integration between CPU and GPU can be achieved by fabricating them on the same chip and many such processors have already been produced, such as AMD Llano [Branover et al. 2012], Intel Sandy Bridge [Yuffe et al. 2011], Ivy Bridge [Damaraju et al. 2012] etc.”
    Quote in the paperNot a substantial claim0 of 3 sources checked
    • cites Branover et al. 2012 source
    • cites Yuffe et al. 2011 source
    • cites Damaraju et al. 2012 source

  • Probably not a substantial claim. The paper’s conclusions may not depend on it. It may state two separate claims as one.

    Findingcitation

    Recent CPUs have 8 to 32 cores; CPUs with more than 60 cores are likely to be available in the near future.

    “Earlier CPUs had only one or two cores, while as recent CPUs have 8 to 32 cores [Fluhr et al. 2014;Morgan 2014] and CPUs with more than 60 cores are likely to be available in near future [Gardner 2014].”
    Quote in the paperNot a substantial claim0 of 3 sources checked
    • cites Fluhr et al. 2014 source
    • cites Morgan 2014

      Couldn’t be matched to a paper in Crossref, so it can’t be checked.

    • cites Gardner 2014

      Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  • Probably not a substantial claim. The paper’s conclusions may not depend on it.

    Findingreasoning

    For CPU programming, OpenMP is most widely used due to its portability and ease of programming by virtue of use of compiler directives.

    “For CPU programming, OpenMP is most widely used due to its portability and ease of programming by virtue of use of compiler directives.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim. The paper’s conclusions may not depend on it. It may state two separate claims as one.

    Findingcitation

    The GT200 GPU (2008) had 1.4B transistors; the recent Geforce GTX TITAN X GPU has 8B transistors.

    “Similarly, the GT200 GPU (2008) had 1.4B transistors, while the recent Geforce GTX TITAN X GPU has 8B transistors [Pirzada 2015].”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites Pirzada 2015

      Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  • Probably not a substantial claim. The paper’s conclusions may not depend on it. It may state two separate claims as one.

    Findingcitation

    The number of CUDA cores in GeForce GTX 280, GTX 480, GTX 680 and GTX TITAN X GPU is 240, 448, 1536 and 3072, respectively.

    “Along similar lines, the number of CUDA cores in GeForce GTX 280, GTX 480, GTX 680 and GTX TITAN X GPU is 240, 448, 1536 and3072, respectively [NVIDIA 2015].”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites NVIDIA 2015

      Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  • It may state two separate claims as one.

    Findingcitation

    GPUs use much larger number of cores, which are in-order cores that share their control unit; GPU cores use lower frequency and smaller sized caches.

    “In contrast, GPUs use much larger number of cores, which are in-order cores that share their control unit.Also, GPU cores use lower frequency, and smaller sized caches [Mittal 2014a].”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites Mittal 2014a

      Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  • It may state two separate claims as one.

    Interpretationreasoning

    While fused HCSs address several limitations of discrete HCSs, they also introduce new tradeoffs and require the design of suitable data-access strategies to take full advantage of their fast interconnection.

    “While fused HCSs address several limitations of discrete HCSs, they also introduce new tradeoffs and require the design of suitable data-access strategies to take full advantage of their fast interconnection.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim. It may be too vague to check against evidence.

    Interpretationreasoning

    CPU-GPU collaboration is inevitable to achieve high-performance computing.

    “As both CPU and GPU become employed in a wide range of applications, it has been acknowledged that both of these processing units (PUs) have their unique features and strengths and hence, CPU-GPU collaboration is inevitable to achieve high-performance computing.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim. The paper may give no reason why it holds. It may state two separate claims as one.

    Premisenone given

    Many algorithm-level techniques require the programmer to manually partition the workload between PUs, identify subtasks suitable for each PU, and/or profile the performance of each PU; this is a tedious process which does not scale well beyond small problems.

    “Many algorithm-level techniques require the programmer to manually partition the workload between PUs, identify the subtasks suitable for each PU and/or profile the performance of each PU.This is a tedious process which does not scale well beyond small problems.”
    Quote in the paperNot a substantial claim

  • It may state two separate claims as one.

    Interpretationreasoning

    Fused HCSs reduce data transfer time, which especially benefits applications requiring large communication between CPU and GPU.

    “This reduces their data transfer time which especially benefits the applications which require large communication between CPU and GPU.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim. The paper’s conclusions may not depend on it.

    Premisenone given

    In systems with GPUs, CPUs have been conventionally used as host for GPU to manage I/O and scheduling.

    “In systems with GPUs, CPUs have been conventionally used as host for GPU to manage I/O and scheduling, however, as continuing innovations improve CPU performance even further (refer Section 2.2), using their computation capabilities also has become more attractive.”
    Quote in the paperNot a substantial claim

  • It may state two separate claims as one.

    Premisereasoning

    Existing fused HCSs are not strictly superior to discrete GPUs on all parameters, and extensive evaluation and development on them has been lacking.

    “Moreover, existing fused HCSs are not strictly superior to discrete GPUs on all parameters and hence, despite their potential, extensive evaluation and development on them has been lacking.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim. The paper may give no reason why it holds. It may state two separate claims as one.

    Premisenone given

    Both CPU and GPU have unique features and strengths.

    “As both CPU and GPU become employed in a wide range of applications, it has been acknowledged that both of these processing units (PUs) have their unique features and strengths and hence, CPU-GPU collaboration is inevitable to achieve high-performance computing.”
    Quote in the paperNot a substantial claim

  • It may state two separate claims as one.

    Premisecitation

    CPU-only or GPU-only systems are usually over-provisioned to meet worst-case performance requirements, but their average utilization remains low.

    “To meet the worse-case performance requirements, the CPU-only or GPU-only systems are usually over-provisioned, however, their average utilization remains low [Mittal 2012;Mittal and Vetter 2015].”
    Quote in the paperSubstantial claim0 of 2 sources checked
    • cites Mittal 2012 source
    • cites Mittal and Vetter 2015

      Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  • Probably not a substantial claim.

    Premisereasoning

    HCSs require special benchmarks to gain insights into their functioning.

    “Due to their unique characteristics, HCSs require special benchmarks to gain insights into their functioning.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim. It may state two separate claims as one.

    Premisecitation

    Several initial works report GPUs provide up to 100× to 1000× speedup; other researchers claim that on applying careful optimizations on both CPUs and GPUs, CPUs may equal or even outperform GPUs.

    “Further, while several initial works report that GPUs provide up to 100× to 1000× speedup, other researchers claim that on applying careful optimizations on both CPUs and GPUs; CPUs may equal or even outperform the performance of GPUs [Gummaraju et al. 2010;Lee et al. 2010].”
    Quote in the paperSubstantial claim0 of 2 sources checked
    • cites Gummaraju et al. 2010 source
    • cites Lee et al. 2010 source

  • Probably not a substantial claim.

    Interpretationnone given

    The vastly different architectures and programming models of CPU and GPU present several challenges in achieving collaborative computing.

    “The vastly different architectures and programming models of CPU and GPU, however, also present several challenges in achieving such collaborative computing.”
    Quote in the paperSubstantial claim

  • It may state two separate claims as one.

    Premisecitation

    Mobile computing systems generally provide multimedia services under real-time performance constraints and hence are attractive targets for acceleration through heterogeneous computing.

    “Mobile computing systems, which now number nearly equal to the population of the earth [International Telecommunication Union 2012], generally provide multimedia services under real-time performance constraints and hence they are attractive targets for acceleration through the use of heterogeneous computing.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites International Telecommunication Union 2012

      Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  • It may state two separate claims as one.

    Interpretationnone given

    For applications where data transfers dominate execution time, or branch divergence does not allow uninterrupted execution on all GPU cores, CPUs can provide better performance than GPUs.

    “For applications where data transfers dominate execution time, or branch divergence does not allow for the uninterrupted execution on all GPU cores, CPUs can provide better performance than GPUs.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim. The paper may give no reason why it holds.

    Interpretationnone given

    The vastly different architecture, programming model and performance of CPU and GPU present unique challenges in heterogeneous computing.

    “The vastly different architecture, programming model and performance (for a given program) of CPU and GPU present unique challenges in heterogeneous computing.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim. The paper’s conclusions may not depend on it. It may state two separate claims as one.

    Premisecitation

    For a query processing application, CPUs are more efficient for queries involving short lists while GPUs are more efficient for those involving long lists.

    “For example, Ding et al. [2009] note that for a query processing application, CPUs are more efficient for queries involving short lists while GPUs are more efficient for those involving long lists.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites Ding et al. 2009 source

  • Probably not a substantial claim. It may be too vague to check against evidence. The paper may give no reason why it holds.

    Premisenone given

    Significant amount of work needs to be done before recently introduced HCS programming frameworks can reach the level of maturity, robustness and popularity as enjoyed by conventional CPU programming languages.

    “Significant amount of work needs to be done before recently introduced HCS programming frameworks can reach the level of maturity, robustness and popularity as enjoyed by the conventional CPU programming languages.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim. The paper’s conclusions may not depend on it.

    Premisereasoning

    Heterogeneous computing has been actively researched and utilized in a variety of applications ranging from natural sciences to engineering.

    “As a result, heterogeneous computing has been actively researched and utilized in variety of applications ranging from natural sciences to engineering etc.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim. The paper may give no reason why it holds. It may state two separate claims as one.

    Premisenone given

    Several existing large-scale codes are written in either CPU or GPU languages, and porting them to HCSs will incur large overhead and is error-prone.

    “Several existing large-scale codes are written in either CPU or GPU languages.Porting them to HCSs will incur large overhead and is also error-prone.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim. The paper’s conclusions may not depend on it.

    Premisecitation

    CPU-GPU workload division based techniques on heterogeneous systems use CPU also to get performance gain, not just as a host.

    “However, given that CPU-GPU workload division based techniques on heterogeneous systems use CPU also to get performance gain (and not just as a host), sometimes CPU is also referred to as an accelerator [Sun et al. 2012].”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites Sun et al. 2012 source

  • It may state two separate claims as one.

    Interpretationcitation

    NVLink promises to offer 5 to 12× bandwidth compared to PCIe Gen3 interconnect and is likely to offset the bandwidth limitation of conventional interconnects.

    “However, a recently proposed interconnect, called NVLink, promises to offer 5 to 12× bandwidth compared to PCIe Gen3 interconnect [Gupta 2014] and this is likely to offset the bandwidth limitation of conventional interconnects.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites Gupta 2014

      Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  • Probably not a substantial claim. The paper’s conclusions may not depend on it. It may state two separate claims as one.

    Premisereasoning

    The domains listed in Table VII are not exhaustive; heterogeneous computing can be applied in several other domains which are computationally intensive and/or provide opportunity for parallelization.

    “These domains are, by no means, exhaustive as heterogeneous computing can be applied in several other domains also which are computationally intensive and/or provide opportunity for parallelization.”
    Quote in the paperNot a substantial claim

  • It may state two separate claims as one.

    Premisereasoning

    Data transfer between CPU and GPU in discrete systems happens over a connection bus (e.g., PCIe) and incurs large overhead.

    “Discrete GPUs have separate memory spaces from the CPU and hence, data transfer happens over a connection bus (e.g.PCIe bus), which incurs large overhead.”
    Quote in the paperSubstantial claim
    • Rests on an assumption: Discrete GPUs have separate memory spaces from the CPU.

  • Probably not a substantial claim. The paper’s conclusions may not depend on it. The paper may give no reason why it holds.

    Premisenone given

    Many other techniques which aim to improve performance may also save energy.

    “Note that while many other techniques which aim to improve performance may also save energy, in this table we only show those techniques which have been evaluated on the basis of energy efficiency.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim. The paper’s conclusions may not depend on it.

    Premisenone given

    Some benchmark suites provide implementations in different languages which allow comparison using those programming models.

    “Notice that some of these benchmarks provide implementations in different languages which allow comparison using those programming models.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim.

    Interpretationreasoning

    Using HCSs in mobile systems calls for an order-of-magnitude improvement in their energy efficiency.

    “While the large cost and power consumption of HCSs may be acceptable in high-end computing systems, using them in mobile systems calls for an order-of-magnitude improvement in their energy efficiency.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim. The paper may give no reason why it holds.

    Premisenone given

    In recent years, CPUs and GPUs are increasingly being seen as indispensable coprocessors, instead of substitutes for each other.

    “In recent years, CPUs and GPUs are increasingly being seen as indispensable coprocessors, instead of substitutes for each other.”
    Quote in the paperNot a substantial claim
40 of 40

  • The paper relies on this without support.

    Modern multicore CPUs use up to a few tens of cores, which are typically out-of-order, multi-instruction issue cores; CPU cores run at high frequency and use large sized caches to minimize the latency of a single-thread.

    “Modern multicore CPUs use upto a few tens of cores, which are typically out-of-order, multi-instruction issue cores.Also, CPU cores run at high-frequency and use large sized caches to minimize the latency of a single-thread.”

  • The paper relies on this without support.

    After allocating a task to GPU (starting the kernel), CPU stays idle, leading to wastage of energy.

    “Further, after allocating the task to GPU (i.e. starting the kernel), CPU stays idle which leads to wastage of energy.”

  • The authors state this as an assumption.

    The different subtasks are of similar nature and can run on any of the PU.

    “For example, assuming that the different subtasks are of similar nature and can run on any of the PU, the decision about where a subtask should be mapped depends on which PU provides higher performance and which mapping helps in achieving load-balancing.”

  • The paper relies on this without support.

    Different languages offer different tradeoffs between ease of programming, ability to write optimized code, ability to target multiple PUs and products from different vendors etc.

    “Different languages offer different tradeoffs between ease of programming, ability to write optimized code, ability to target multiple PUs and products from different vendors etc.”

  • The paper relies on this without support.

    OpenCL can be used for both CPU and GPU, and hence, applications written in OpenCL can be easily scheduled on any PU.

    “For example, OpenCL can be used for both CPU and GPU, and hence, applications written in OpenCL can be easily scheduled on any PU.”

  • The paper relies on this without support.

    CUDA works only on NVIDIA GPUs.

    “By comparison, CUDA works only on NVIDIA GPUs”

  • The paper relies on this without support.

    ACML and Brook+ can be used to program AMD GPUs only.

    “ACML (AMD Core Math Library) and Brook+ can be used to program AMD GPUs only.”

  • The paper relies on this without support.

    OpenMP, Pthread, TBB and MKL are used to write parallel programs on CPUs.

    “Similarly OpenMP, Pthread, TBB (thread building block) and MKL (math kernel library) are used to write parallel programs on CPUs.”

  • The paper relies on this without support.

    MKL is internally parallelized by OpenMP threading.

    “Note that MKL is internally parallelized by OpenMP threading; for sake of clarity, we mention it separately in Table III.”

  • The paper relies on this without support.

    Small task sizes provide better load-balancing, while large task sizes provide better resource utilization and an optimal value of task sizes achieves a trade-off between these two factors.

    “Small task sizes provide better load-balancing, while large task sizes provide better resource utilization and an optimal value of task sizes achieves a trade-off between these two factors.”

  • The paper relies on this without support.

    HMMER algorithm has large memory requirement and thus, for large sized sequences, shared memory of GPU becomes unsuitable while global memory incurs large performance penalty.

    “HMMER algorithm has large memory requirement and thus, for large sized sequences, shared memory of GPU becomes unsuitable while global memory incurs large performance penalty.”

  • The paper relies on this without support.

    Discrete GPUs have separate memory spaces from the CPU.

    “Discrete GPUs have separate memory spaces from the CPU and hence, data transfer happens over a connection bus (e.g.PCIe bus), which incurs large overhead.”
12 of 12

Self-check

Percentages, sums and reported differences re-computed from the paper’s own numbers.

Each contradiction between two passages listed here was confirmed by an independent check.

  • Dithering step 3 is assigned to both GPU and CPU

    Minor

    The handover description says step 3 runs on the GPU, but the subsequent description says the CPU performs steps 1 and 3. The account gives conflicting assignments for the final step.

1 of 1

Citations

Cites

Flagged first, then by citations

Research works printed but not linked: 12

  1. Chris Gregg, Jeff Brantley, and Kim Hazelwood. 2010. Contention-aware scheduling of parallel code for heterogeneous systems. In Proceedings of the USENIX Workshop on Hot Topics in Parallelism (HotPar’10).
  2. 10.5555/2015039.2015535 · DOI
  3. 10.5555/1987237.1987259 · DOI
  4. 10.5555/1885276.1885302 · DOI
  5. 10.5555/1964238.1964251 · DOI
  6. 10.5555/1761728.1761755 · DOI
  7. Usman Pirzada. 2015. Nvidia Geforce GTX TITAN X Unveiled - GM200 ‘Big Daddy Maxwell’, 12GB VRam and 8 Billion Transistors. Retrieved from wccftech.com/nvidia-gtx-titan-x-revealed-gdc-2015/.
  8. 10.5555/1855610.1855621 · DOI
  9. 10.5555/1885276.1885305 · DOI
  10. 10.5555/2001252.2001258 · DOI
  11. 10.5555/1863086.1863099 · DOI
  12. Dieter Wendel, Ronald Kalla, Joshua Friedrich, James Kahle, Jens Leenstra, Cedric Lichtenau, Balaram Sinharoy, William Starke, and Victor Zyuban. 2010. The POWER7 processor SoC. In Proceedings of the International Conference on IC Design and Technology (ICICDT’10). 71--73.

Cited by

Sorted by citations

Related

Metadata

Authors
Mittal, S., & Vetter, J. S.
Affiliations
Oak Ridge National Laboratory (US)
ORCID
Venue
ACM Computing Surveys
Publisher
Association for Computing Machinery (ACM)
Published
2015-07-21
Volume
47
Issue
4
Pages
1-35
ISSN
  • 0360-0300 (ISSN-L)
  • 1557-7341
DOI
10.1145/2788396
OpenAlex
W1864199185
Semantic Scholar
a14bb15298961545b83e7c7cefff0e7af79828f7
Cited by
568
References
197
Citations by year
2026: 43 · 2025: 51 · 2024: 57 · 2023: 46 · 2022: 45 · 2021: 58 · 2020: 61 · 2019: 67 · 2018: 68 · 2017: 35 · 2016: 35 · 2015: 2
FWCI
71.01
Type
journal article
Language
en
Open access
Licence
Locations
  • ACM Computing Surveys (publishedVersion, public-domain)
  • OSTI OAI (U.S. Department of Energy Office of Scientific and Technical Information) (submittedVersion)
Funders
  • U.S. Department of Energy
  • Office of Science
  • Advanced Scientific Computing Research
Topics
  • Physical Sciences › Computer Science › Hardware and Architecture › Parallel Computing and Optimization Techniques
  • Physical Sciences › Computer Science › Computer Networks and Communications › Distributed and Parallel Computing Systems
  • Physical Sciences › Computer Science › Hardware and Architecture › Embedded Systems Design Techniques
Keywords
CPU-GPU heterogeneous computing, workload partitioning, heterogeneous supercomputers, energy efficiency, benchmark suite, exascale computing

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

A language model drew the claims from the paper on 2026-11-08, under version 3 of the claim definition (model deepseek-v4.1-flash, prompt extract-v3).

A second model labelled each claim on 2026-11-08 (model jev-1.13-20260917, definition revision gold-guideline-v0.2).

Cited sources are checked with a language model’s help (gpt-6-sol).

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review