[Paper Review] CPU and/or GPU: Revisiting the GPU Vs. CPU Myth

Kishore Kothapalli, Dip Sankar Banerjee+14 more|arXiv (Cornell University)|Mar 9, 2013

Parallel Computing and Optimization Techniques35 references9 citations

TL;DR

This paper challenges the CPU vs. GPU dichotomy by proposing a hybrid computing model that leverages both CPU and GPU simultaneously across 13 diverse workloads. Using two hybrid platforms—Intel i7-980X + Tesla T10 and Intel E7400 + GT520—results show hybrid solutions achieve 90% average resource efficiency and outperform standalone CPU or GPU implementations, demonstrating that co-processing maximizes performance and efficiency in modern heterogeneous systems.

ABSTRACT

Parallel computing using accelerators has gained widespread research attention in the past few years. In particular, using GPUs for general purpose computing has brought forth several success stories with respect to time taken, cost, power, and other metrics. However, accelerator based computing has signifi- cantly relegated the role of CPUs in computation. As CPUs evolve and also offer matching computational resources, it is important to also include CPUs in the computation. We call this the hybrid computing model. Indeed, most computer systems of the present age offer a degree of heterogeneity and therefore such a model is quite natural. We reevaluate the claim of a recent paper by Lee et al.(ISCA 2010). We argue that the right question arising out of Lee et al. (ISCA 2010) should be how to use a CPU+GPU platform efficiently, instead of whether one should use a CPU or a GPU exclusively. To this end, we experiment with a set of 13 diverse workloads ranging from databases, image processing, sparse matrix kernels, and graphs. We experiment with two different hybrid platforms: one consisting of a 6-core Intel i7-980X CPU and an NVidia Tesla T10 GPU, and another consisting of an Intel E7400 dual core CPU with an NVidia GT520 GPU. On both these platforms, we show that hybrid solutions offer good advantage over CPU or GPU alone solutions. On both these platforms, we also show that our solutions are 90% resource efficient on average. Our work therefore suggests that hybrid computing can offer tremendous advantages at not only research-scale platforms but also the more realistic scale systems with significant performance gains and resource efficiency to the large scale user community.

Motivation & Objective

• To challenge the prevailing notion that GPU computing alone offers superior performance, arguing instead for integrated CPU+GPU utilization.
• To evaluate the feasibility and benefits of hybrid computing on realistic, commodity-scale platforms with diverse workloads.
• To demonstrate that both CPU and GPU can be actively utilized in parallel, avoiding idle resources common in accelerator-only models.
• To explore algorithmic design patterns—work sharing and task partitioning—that enable efficient hybrid execution across heterogeneous architectures.
• To provide empirical evidence that hybrid solutions outperform both CPU-only and GPU-only approaches in performance and resource efficiency.

Proposed method

• Selected 13 diverse workloads spanning databases, image processing, sparse matrix kernels, and graph algorithms to test hybrid computing across domains.
• Implemented hybrid solutions on two platforms: a high-end (6-core i7 + Tesla T10) and a low-end (dual-core E7400 + GT520) configuration.
• Employed two algorithmic strategies: work sharing (decomposing problems across CPU and GPU) and task partitioning (assigning different parts of an algorithm to different devices).
• Measured performance and resource efficiency by comparing hybrid solutions against standalone CPU and GPU implementations.
• Used performance metrics such as execution time, speedup, and resource utilization to evaluate efficiency and scalability.
• Analyzed workload characteristics (e.g., memory access patterns, SIMD parallelism) to guide optimal task mapping between CPU and GPU.

Experimental results

Research questions

• RQ1 Can hybrid CPU+GPU computing outperform standalone CPU or GPU solutions across a broad range of diverse workloads?
• RQ2 What are the performance and efficiency gains of using both CPU and GPU in parallel, rather than relying on one accelerator?
• RQ3 How does the choice of CPU and GPU combination affect the viability and performance of hybrid computing on commodity platforms?
• RQ4 What algorithmic strategies (e.g., work sharing, task partitioning) enable efficient and balanced utilization of CPU and GPU resources?
• RQ5 To what extent can hybrid computing achieve high resource efficiency (e.g., 90% utilization) in real-world systems with heterogeneous architectures?

Key findings

• Hybrid solutions consistently outperformed both CPU-only and GPU-only implementations across all 13 workloads tested on both platforms.
• On average, hybrid solutions achieved 90% resource efficiency, indicating near-optimal utilization of available computational resources.
• The high-end platform (i7-980X + Tesla T10) and low-end platform (E7400 + GT520) both demonstrated significant performance gains through hybrid execution.
• Workloads with irregular memory access or low SIMD parallelism—traditionally less suited for GPUs—benefited substantially from CPU involvement.
• The study confirms that hybrid computing is not limited to high-end systems but is viable and effective even on low-cost, commodity platforms.
• The results support the shift from a CPU-vs.-GPU mindset to a CPU-and-GPU paradigm, where co-processing unlocks greater performance and efficiency.