0 reactions
Did some CoreMark testing on various x86 and Arm CPUs.
I started down the rabbit hole of Arm CPU cores and wanted to see how they all compare to each other on various devices I have in my house. I ended up comparing them to x86 CPUs on Windows as well since CoreMark can be compiled against both Windows, Linux, and Android. These are the results sorted by average speed:
|Device|CPU|Core|Average Speed (it/s)|Stdev|
|:-|:-|:-|:-|:-|
|Fire HD 8 8th Gen|MediaTek MT8163B|Cortex-A53 (CPU 0)|3511.98|2.03|
|Galaxy A51515U|Qualcomm Exynos 9611|Cortex-A53 (CPU 0)|5404.43|37.27|
|Galaxy A526U|Qualcomm Snapdragon 750G 5G|Kryo 570 Silver (Based on Cortex-A55) (CPU 0)|6457.92|91.14|
|Pixel 7 Pro|Google Tensor G2|Cortex-A55 (Core 0)|6557.86|75.39|
|Galaxy A51515U|Qualcomm Exynos 9611|Cortex-A73 (CPU 4)|11148.79|15.05|
|Dell T610|Intel Xeon E5649|Westmere-E (Core 0)|12712.49|67.22|
|Yoga 7 14IRL8|Intel Core i7-1355U|Gracemont E-Core (Core 6)|17314.18|282.95|
|Galaxy A526U|Qualcomm Snapdragon 750G 5G|Kryo 570 Gold (Based on Cortex-A77) (CPU 6)|17545.92|10.70|
|Raspberry Pi 5 4GB|Broadcom BCM2712|Cortex-A76 (Core 0)|17955.74|4.05|
|Pixel 7 Pro|Google Tensor G2|Cortex-A78 (Core 5)|20854.48|12.35|
|Yoga 7 14IRL8|Intel Core i7-1355U|Raptor Cove P-Core (Core 0)|24642.72|761.34|
|Pixel 7 Pro|Google Tensor G2|Cortex-X1 (Core 7)|26108.49|71.39|
|GP66 11UH-032|Intel Core i7-11800H|Cypress Cove (Core 0)|31447.33|72.81|
|Gaming Server|AMD Ryzen 5 7600X|Zen 4 (Core 0)|49792.97|147.94|
These are the results sorted by device:
|Device|CPU|Core|Average Speed (it/s)|Stdev|
|:-|:-|:-|:-|:-|
|Dell T610|Intel Xeon E5649|Westmere-E (Core 0)|12712.49|67.22|
|Fire HD 8 8th Gen|MediaTek MT8163B|Cortex-A53 (CPU 0)|3511.98|2.03|
|Galaxy A51515U|Qualcomm Exynos 9611|Cortex-A53 (CPU 0)|5404.43|37.27|
|Galaxy A51515U|Qualcomm Exynos 9611|Cortex-A73 (CPU 4)|11148.79|15.05|
|Galaxy A526U|Qualcomm Snapdragon 750G 5G|Kryo 570 Silver (Based on Cortex-A55) (CPU 0)|6457.92|91.14|
|Galaxy A526U|Qualcomm Snapdragon 750G 5G|Kryo 570 Gold (Based on Cortex-A77) (CPU 6)|17545.92|10.70|
|Gaming Server|AMD Ryzen 5 7600X|Zen 4 (Core 0)|49792.97|147.94|
|GP66 11UH-032|Intel Core i7-11800H|Cypress Cove (Core 0)|31447.33|72.81|
|Pixel 7 Pro|Google Tensor G2|Cortex-A55 (Core 0)|6557.86|75.39|
|Pixel 7 Pro|Google Tensor G2|Cortex-A78 (Core 5)|20854.48|12.35|
|Pixel 7 Pro|Google Tensor G2|Cortex-X1 (Core 7)|26108.49|71.39|
|Raspberry Pi 5 4GB|Broadcom BCM2712|Cortex-A76 (Core 0)|17955.74|4.05|
|Yoga 7 14IRL8|Intel Core i7-1355U|Gracemont E-Core (Core 6)|17314.18|282.95|
|Yoga 7 14IRL8|Intel Core i7-1355U|Raptor Cove P-Core (Core 0)|24642.72|761.34|
I compiled the benchmarking binary using the following command in Android and Linux: `make PORT_DIR=linux CC=gcc XCFLAGS="-O2 -march=armv8-a" NO_LIBRT=1 run`
For Windows, I ran the following command to build the executable: `make PORT_DIR=linux CC=gcc XCFLAGS="-O2 -march=x86-64" NO_LIBRT=1 run`
My rudimentary understanding is that CoreMark primarily stresses the ALUs of a CPU and that these results reflect the cores' ALU performance. Is this accurate?
I find it interesting that some Arm cores are shared in different devices, such as the Cortex-A53 in the Amazon Fire HD 8 8th Gen and the Samsung Galaxy A51515U. Interestingly, the Galaxy's Cortex-A53 is \~50% more performant than the Fire HD 8's Cortex-A53 even though the maximum core boost clock is only \~31% faster (1.7 GHz vs 1.3 GHz). Anyone have an idea as to why the ALU performance of the Cortex-A53 doesn't scale perfectly with clock speed?
I'm starting to get really interested in this stuff. What other benchmarks can I use to test other functional units of a CPU? I'd like some benchmarks that are:
* Cross platform between Windows, Android, Linux, and MacOS
* FOSS
* Usable on >=armv6 and >=x86-32
#hardware #technology
source
I started down the rabbit hole of Arm CPU cores and wanted to see how they all compare to each other on various devices I have in my house. I ended up comparing them to x86 CPUs on Windows as well since CoreMark can be compiled against both Windows, Linux, and Android. These are the results sorted by average speed:
|Device|CPU|Core|Average Speed (it/s)|Stdev|
|:-|:-|:-|:-|:-|
|Fire HD 8 8th Gen|MediaTek MT8163B|Cortex-A53 (CPU 0)|3511.98|2.03|
|Galaxy A51515U|Qualcomm Exynos 9611|Cortex-A53 (CPU 0)|5404.43|37.27|
|Galaxy A526U|Qualcomm Snapdragon 750G 5G|Kryo 570 Silver (Based on Cortex-A55) (CPU 0)|6457.92|91.14|
|Pixel 7 Pro|Google Tensor G2|Cortex-A55 (Core 0)|6557.86|75.39|
|Galaxy A51515U|Qualcomm Exynos 9611|Cortex-A73 (CPU 4)|11148.79|15.05|
|Dell T610|Intel Xeon E5649|Westmere-E (Core 0)|12712.49|67.22|
|Yoga 7 14IRL8|Intel Core i7-1355U|Gracemont E-Core (Core 6)|17314.18|282.95|
|Galaxy A526U|Qualcomm Snapdragon 750G 5G|Kryo 570 Gold (Based on Cortex-A77) (CPU 6)|17545.92|10.70|
|Raspberry Pi 5 4GB|Broadcom BCM2712|Cortex-A76 (Core 0)|17955.74|4.05|
|Pixel 7 Pro|Google Tensor G2|Cortex-A78 (Core 5)|20854.48|12.35|
|Yoga 7 14IRL8|Intel Core i7-1355U|Raptor Cove P-Core (Core 0)|24642.72|761.34|
|Pixel 7 Pro|Google Tensor G2|Cortex-X1 (Core 7)|26108.49|71.39|
|GP66 11UH-032|Intel Core i7-11800H|Cypress Cove (Core 0)|31447.33|72.81|
|Gaming Server|AMD Ryzen 5 7600X|Zen 4 (Core 0)|49792.97|147.94|
These are the results sorted by device:
|Device|CPU|Core|Average Speed (it/s)|Stdev|
|:-|:-|:-|:-|:-|
|Dell T610|Intel Xeon E5649|Westmere-E (Core 0)|12712.49|67.22|
|Fire HD 8 8th Gen|MediaTek MT8163B|Cortex-A53 (CPU 0)|3511.98|2.03|
|Galaxy A51515U|Qualcomm Exynos 9611|Cortex-A53 (CPU 0)|5404.43|37.27|
|Galaxy A51515U|Qualcomm Exynos 9611|Cortex-A73 (CPU 4)|11148.79|15.05|
|Galaxy A526U|Qualcomm Snapdragon 750G 5G|Kryo 570 Silver (Based on Cortex-A55) (CPU 0)|6457.92|91.14|
|Galaxy A526U|Qualcomm Snapdragon 750G 5G|Kryo 570 Gold (Based on Cortex-A77) (CPU 6)|17545.92|10.70|
|Gaming Server|AMD Ryzen 5 7600X|Zen 4 (Core 0)|49792.97|147.94|
|GP66 11UH-032|Intel Core i7-11800H|Cypress Cove (Core 0)|31447.33|72.81|
|Pixel 7 Pro|Google Tensor G2|Cortex-A55 (Core 0)|6557.86|75.39|
|Pixel 7 Pro|Google Tensor G2|Cortex-A78 (Core 5)|20854.48|12.35|
|Pixel 7 Pro|Google Tensor G2|Cortex-X1 (Core 7)|26108.49|71.39|
|Raspberry Pi 5 4GB|Broadcom BCM2712|Cortex-A76 (Core 0)|17955.74|4.05|
|Yoga 7 14IRL8|Intel Core i7-1355U|Gracemont E-Core (Core 6)|17314.18|282.95|
|Yoga 7 14IRL8|Intel Core i7-1355U|Raptor Cove P-Core (Core 0)|24642.72|761.34|
I compiled the benchmarking binary using the following command in Android and Linux: `make PORT_DIR=linux CC=gcc XCFLAGS="-O2 -march=armv8-a" NO_LIBRT=1 run`
For Windows, I ran the following command to build the executable: `make PORT_DIR=linux CC=gcc XCFLAGS="-O2 -march=x86-64" NO_LIBRT=1 run`
My rudimentary understanding is that CoreMark primarily stresses the ALUs of a CPU and that these results reflect the cores' ALU performance. Is this accurate?
I find it interesting that some Arm cores are shared in different devices, such as the Cortex-A53 in the Amazon Fire HD 8 8th Gen and the Samsung Galaxy A51515U. Interestingly, the Galaxy's Cortex-A53 is \~50% more performant than the Fire HD 8's Cortex-A53 even though the maximum core boost clock is only \~31% faster (1.7 GHz vs 1.3 GHz). Anyone have an idea as to why the ALU performance of the Cortex-A53 doesn't scale perfectly with clock speed?
I'm starting to get really interested in this stuff. What other benchmarks can I use to test other functional units of a CPU? I'd like some benchmarks that are:
* Cross platform between Windows, Android, Linux, and MacOS
* FOSS
* Usable on >=armv6 and >=x86-32
#hardware #technology
source
| 2,000 mlx | |
| 2 | |
| 0 |