Evaluation of Successive CPUs/APUs/GPUs Based on an OpenCL Finite Difference Stencil

Abstract

International audienceThe AMD APU (Accelerated Processing Unit) architecture, which combines CPU and GPU cores on the same die, is promising for GPU applications which performance is bottlenecked by the low PCI Express communication rate. However the first APU generations still have different CPU and GPU memory partitions. Currently, the APU integrated GPUs are also less powerful than discrete GPUs. In this paper we therefore investigate the interest of APUs for scientific computing by evaluating and comparing the performance of two successive AMD APUs (family codename Llano and Trinity), two successive discrete GPUs (chip codename Cayman and Tahiti) and one hexa-core AMD CPU. For this purpose, we rely on a 3D finite difference stencil, that is optimized and tuned in OpenCL. We detail the most interesting optimizations for each architecture and show very good performance in OpenCL: up to 500 Gflops on Tahiti. Finally, our results show that APU integrated GPUs outperform CPUs, and that integrated GPUs of upcoming APUs may match discrete GPUs for problems with high communication requirements

Electrical-Level Attacks on CPUs, FPGAs, and GPUs: Survey and Implications in the Heterogeneous Era

“…Access to the main memory is often shared among components; on heterogeneous CPU-GPU platforms, both components can access the main memory, and the GPU can use both coherent and noncoherent communication [32].…”

Towards Seismic Wave Modeling on Heterogeneous Many-Core Architectures Using Task-Based Runtime System

“…In [24], the authors implement a seismic model using task-based programming but they don't use heterogeneous cores for the simulations. Calandra et al [25] evaluate execution of a finite-differences stencil on CPUs, APUs and GPUs but they don't combine heterogeneous cores, neither parallel tasks programming. Figure 7: Impact of the granularity on the efficiency of seismic wave modelling on GPUs+CPUs.…”

Auxílio a Portabilidade De Código Em Aplicações De Alto Desempenho

“…Com isso é possível fazer operações do tipo DMA (direct memory access), habilitando a sobreposição das cópias de memória com execuções de kernels (em CUDA esta sobreposição é feita com as chamadas assíncronas, que só funcionam se a memória alocada na CPU for pinned [68]). O uso desse tipo de memória para transferências CPU ⇔ dispositivo é recomendado pois essas transferências ocorrem no barramento PCI Express, que é o gargalo [69] de muitas arquiteturas heterogêneas.…”