http://wlog.flatlib.jp/item/1800 この表を見てて、そういえば、私もRaspberry Pi 3 が家に来たし似たような実験できるなと思ったので私も似たような実験をした。
binutilsのビルド時間です。消費電力はKILL A WATT目視。
raspberry pi 3 (idle 1.8W, load 4.7W, diff=2.9W)
Performance counter stats for 'sh -c ../configure ; make -j4':
797005.468900 task-clock (msec) # 2.953 CPUs utilized
121,732 context-switches # 0.153 K/sec
29,936 cpu-migrations # 0.038 K/sec
8,586,527 page-faults # 0.011 M/sec
954,251,209,471 cycles # 1.197 GHz
<not supported> stalled-cycles-frontend
<not supported> stalled-cycles-backend
422,118,328,351 instructions # 0.44 insns per cycle
53,087,297,010 branches # 66.608 M/sec
6,646,202,484 branch-misses # 12.52% of all branches
269.885547005 seconds time elapsed
raspberry pi 2 (idle 1.6W, load 2.9W, diff=1.3W)
Performance counter stats for 'sh -c ../configure ; make -j4':
1477301.886492 task-clock (msec) # 3.017 CPUs utilized
138,249 context-switches # 0.094 K/sec
31,201 cpu-migrations # 0.021 K/sec
8,587,257 page-faults # 0.006 M/sec
1,326,853,390,190 cycles # 0.898 GHz
<not supported> stalled-cycles-frontend
<not supported> stalled-cycles-backend
422,185,890,968 instructions # 0.32 insns per cycle
52,974,180,568 branches # 35.859 M/sec
13,626,377,155 branch-misses # 25.72% of all branches
489.620091792 seconds time elapsed
parallella (idle 2.8W, load 3.5W, diff=0.7W)
Performance counter stats for 'sh -c ../configure ; make -j2':
1477010.321602 task-clock (msec) # 1.674 CPUs utilized
138413 context-switches # 0.094 K/sec
12771 cpu-migrations # 0.009 K/sec
8209396 page-faults # 0.006 M/sec
982777694036 cycles # 0.665 GHz
83520783309 stalled-cycles-frontend # 8.50% frontend cycles idle
629510158808 stalled-cycles-backend # 64.05% backend cycles idle
553902409016 instructions # 0.56 insns per cycle
# 1.14 stalled cycles per insn
57338355895 branches # 38.821 M/sec
15218097669 branch-misses # 26.54% of all branches
882.186437213 seconds time elapsed
liva ecs (idle 3.5W, 7.2W, diff=3.7W) Celeron N2807 2core
Performance counter stats for 'sh -c ../configure ; make -j2':
399205.341262 task-clock (msec) # 1.790 CPUs utilized
180763 context-switches # 0.453 K/sec
24511 cpu-migrations # 0.061 K/sec
8537573 page-faults # 0.021 M/sec
832800007036 cycles # 2.086 GHz (51.63%)
<not supported> stalled-cycles-frontend
<not supported> stalled-cycles-backend
407625678041 instructions # 0.49 insns per cycle (76.75%)
88939447423 branches # 222.791 M/sec (76.15%)
7372201988 branch-misses # 8.29% of all branches (76.20%)
222.991643196 seconds time elapsed
| 時間[秒] | IPC | W(load) | W(idle) | W(load-idle) | ||
| rpi3 | 270 | 0.44 | 4.7 | 1.8 | 2.9 | Cortex A53 4core |
| rpi2 | 490 | 0.32 | 2.9 | 1.6 | 1.3 | Cortex A7 4core |
| parallella | 882 | 0.56 | 3.5 | 2.8 | 0.7 | Cortex A9 2core |
| liva ecs | 223 | 0.49 | 7.2 | 3.5 | 3.7 | Silvermont 2core |
わかることは、
- Cortex-A9 と Cortex-A53 だと、A53のほうがIPC良いと思ってたけど、A9のほうがいいんだな。まあA9一応OoOだからとは思うけど、Cortex-A9 → Cortex-A53って5年も時間経ってるしなんとかなってると思ってた
- 分岐ミスがA9,A7 は同じぐらいだけど、A53は半分くらいに減ってる。
まああんまり良い比較ではない、というのは
