General-purpose vs. gpu: Comparison of many-cores on irregular workloads

TitleGeneral-purpose vs. gpu: Comparison of many-cores on irregular workloads
Publication TypeConference Papers
Year of Publication2010
AuthorsCaragea GC, Keceli F, Tzannes A, Vishkin U
Conference NameProceedings of the Second Usenix Workshop on Hot Topics in Parallelism
Date Published2010///

XMT1 is a general-purpose many-core parallel architec-ture. The foremost design objective for XMT was to meet
the highest standards for ease of parallel programming.
GPUs, on the other hand, have acquired a strong reputa-
tion on performance, sometimes at the expense of ease-
of-programming. The current paper presents a perfor-
mance comparison on diverse workloads between XMT
and an NVIDIA CUDA-enabled GPU. Configured with
roughly the same amount of chip resources as the GPU,
XMT achieves an average speedup of 6.05x on irregu-
lar applications, while incurring an average slowdown of
2.07x on regular ones. Namely, XMT comes ahead for
significant applications without having to pay a (possibly
worthwhile) price for easier programming. This surpris-
ing result suggests a yet untapped opportunity: A high-
performance easy-to-program general-purpose 1000-core