Video : My first CPU benchmarks on the Rock5 in Linux

Hi all. I’ve gotten access to a Rock5b test board. I’ve done some quick benchmarks and made a video about it.
Interesting to see the great A76 performance.
Here it is, greetings. NicoD

2 Likes

OMG, you still spread this insane BS at 5:50

While I already explained this to you multiple times another attempt:

What you’re seeing with 7-zip not utilizing all cores at 100% has nothing to do with ‘uneven performance’ or whatever else sick explanation you pulled from somewhere but with the task being bottlenecked by memory access.

The four A55 cores tested by sbc-bench at 1830 MHz:

RAM size:   15723 MB,  # CPU hardware threads:   8
RAM usage:    882 MB,  # Benchmark threads:      4

                       Compressing  |                  Decompressing
Dict     Speed Usage    R/U Rating  |      Speed Usage    R/U Rating
         KiB/s     %   MIPS   MIPS  |      KiB/s     %   MIPS   MIPS

22:       4384   346   1233   4265  |      86337   396   1859   7366
23:       4254   353   1229   4335  |      85544   399   1854   7402
24:       4267   364   1262   4589  |      83051   397   1838   7291
25:       4156   369   1284   4746  |      81616   399   1822   7264
----------------------------------  | ------------------------------
Avr:             358   1252   4484  |              398   1843   7331
Tot:             378   1548   5907

RK3568, same cores, same kernel, same settings, same OS, same benchmark, but A55s clocked ~100 MHz higher:

RAM size:    1958 MB,  # CPU hardware threads:   4
RAM usage:    882 MB,  # Benchmark threads:      4

                       Compressing  |                  Decompressing
Dict     Speed Usage    R/U Rating  |      Speed Usage    R/U Rating
         KiB/s     %   MIPS   MIPS  |      KiB/s     %   MIPS   MIPS

22:       2897   354    797   2818  |      82936   394   1797   7076
23:       2853   370    786   2907  |      80935   394   1777   7003
24:       2728   376    780   2933  |      78674   393   1756   6906
25:       2627   381    788   3000  |      76137   393   1723   6776
----------------------------------  | ------------------------------
Avr:             370    788   2915  |              394   1763   6940
Tot:             382   1275   4928

What can we see from these numbers? Memory access matters!

While the A55 in little RK3568 are clocked higher, the 7-ZIP score is lower and also CPU utilization is lower when decompressing: 99.5% on RK3588 and only 98.5% on RK3568.

According to your theory about ‘uneven performance’ (or whatever you called it in the past – you know why YouTube tech videos are crap? Since no text you could quickly search through! One must go through all the annoying babbling all the time!) you would expect 100%, right?

Quick check of sbc-bench’s results list [1] reveals the following CPU utilization for different SoCs when executing 7z b on all cores:

SoC compression decompression
BCM2835 99 98
BCM2836/BCM2709 309 397
Allwinner A64 or https://tinyurl.com/yyf3d7fg 357 398
Allwinner H3/H2+ 333 398
Amlogic S905X2/S905Y2/S905D2/T962X2 331 394
SigmaStar SSD201/SSD202D 160 199
Allwinner H3/H2+ 332 395
2 x ThunderX CN8890 8692 8937
Amlogic S905 321 394
Allwinner H5 360 397
BCM2711B0 366 399
Amlogic S905X3 326 391
Allwinner A20 162 189
BCM2835 99 98
Amlogic S922X 567 508
Rockchip RK3399 545 530
Allwinner A20 164 189
BCM2711B0 349 397
Nvidia Jetson Nano 303 395
2 x ThunderX CN8890 7947 8663
2 x ThunderX CN8890 8344 8439
Allwinner A20 164 190
Amlogic S905X2/S905Y2/S905D2/T962X2 327 382
Rockchip RK3399 573 517
Amlogic Meson GXL (S905X) Revision 21:c (84:2) 301 394
Amlogic Meson SM1 (S905X3) Revision 2b:c (10:2) 370 392
Amlogic Meson GXBB (S905) Revision 1f:c (13:1) 299 377
Amlogic Meson G12B (S922X) Revision 29:c (40:2) 566 507
Amlogic Meson G12B (A311D) Revision 29:b (10:2) 568 508
Amlogic Meson8m2 (S812) RevA (1d - 0:74E) detected 330 395
Rockchip RK3568 (35681000) 367 394
Amlogic Meson8m2 (S812) RevA (1d - 0:74E) detected 339 387
Allwinner R40/V40 322 385
Amlogic Meson SM1 (S905X3) Revision 2b:c (10:2) 325 396
Rockchip RK3566 or RK3568 372 396
Nvidia Jetson Nano 301 367
Rockchip RK3568 (35682000) 365 395
BCM2711B0 370 395
Amlogic Meson SM1 (Unknown) Revision 2b:b (40:2) 359 394
Amlogic Meson G12B (A311D) Revision 29:b (10:2) 564 509
Amlogic A311D2 732 715
Rockchip RK3288 353 397
Rockchip RK3588 (35880000) 743 663
Rockchip RK3588 (35880000) 742 692
Nvidia AGX Xavier 509 589
Rockchip RK3328 367 398
Samsung/Nexell S5P6818 670 796
Rockchip RK3188 325 377
Allwinner D1 93 97
Rockchip RK3568 369 393
Amlogic Meson GXM (S912) Revision 22:a (82:2) 701 712
NXP i.MX6 Quad 320 388
Rockchip RK3588 (35880000) 742 679
Amlogic Meson8 (S802) RevC (19 - 0:27ED) 303 395
Kendryte K510 141 199
Phytium D2000 755 782
Phytium D2000 743 782

Even single core SoCs that feature a crappy memory controller are far away from reaching 100%, see Allwinner D1 for example. It has nothing to do with type of cores or ‘something uneven’ like you spread since years but with cores fighting over memory access!

So please stop spreading this BS! As well as wrong info about clockspeeds in this video when all you was reporting was cpufreq OPP and not clockspeeds (they need to be measured like sbc-bench is doing it).

BTW: if all you’re checking for is 100% CPU utilization then some lightweight joke like while true ; do yes >/dev/null; done on all cores is all that’s needed!

[1] just parsing the info everybody has at his hands since it happens in the open:

tk@mac-tk results % grep "SoC guess" *.txt | while read ; do
    SoCName="$(awk -F": " '{print $2}' <<<"${REPLY}")"
    ResultsFile=$(cut -f1 -d':' <<<"${REPLY}")
    echo "|  [${SoCName}](http://ix.io/$(basename ${ResultsFile} .txt)) | $(awk -F" " '/^Avr/ {print $2" | "$6}' "${ResultsFile}" | tail -n1) |"
done
1 Like

Utilization divided by core count and also older sbc-bench results considered that missed SoC guessing:

Device cores comp single decomp single comp multi decomp multi
v0.6.3 ODROID-N2 6 99 99 91 88
v0.6.2 ODROID-N2 6 99 99 93 90
v0.6.6 Realtek_Lion_Skin_1GB 4 - - - -
v0.6.6 SolidRun LX2160A COM express type 7 module 16 100 100 85 97
v0.6.7 jetson-nano 4 99 100 75 98
v0.6.7 Khadas VIM3 6 100 100 93 84
v0.6.7 RPi 4B Rev 1.1 4 99 100 89 98
v0.6.8 icosa 4 100 100 80 98
v0.6.9 Radxa ROCK Pi S 4 100 100 83 99
v0.4.3 Odroid XU4 8 99 99 64 78
v0.4.6 nexell soc 8 100 100 78 99
v0.4.6 Libre Technology CC 4 100 100 75 99
v0.4.6 Rockchip RK3288 Tinker Board 4 100 100 89 95
v0.4.6 ODROID-C2 4 100 100 77 96
v0.4.6 4 100 100 82 96
v0.5 Odroid XU4 8 100 100 64 78
v0.5 FriendlyElec NanoPC-T4 6 100 100 65 88
v0.5 Orange Pi Plus / Plus 2 4 100 100 82 97
v0.5 Pine64 Rock64 4 100 100 84 98
v0.5 Pine64 Rock64 4 100 100 84 98
v0.5 FriendlyElec NanoPC-T4 6 100 100 77 88
v0.5 Olimex A10-OLinuXino-LIME 1 93 94 - -
v0.5 Orange Pi PC Plus 4 100 100 81 98
v0.5.3 Helios4 2 100 100 79 99
v0.5.4 Pine H64 4 100 100 88 99
v0.5.3 nexell soc 8 100 100 78 99
v0.5.3 nexell soc 8 100 100 78 99
v0.5.4 Globalscale Marvell ESPRESSOBin Board 2 100 100 74 99
v0.6.1 Pine64 RockPro64 6 100 100 85 88
v0.6.1 Globalscale Marvell ESPRESSOBin Board 2 100 100 77 97
v0.5.4 FriendlyElec NanoPC-T4 6 100 100 88 86
v0.5.6 FriendlyElec NanoPi M4 6 100 100 70 86
v0.6.1 RPi Zero W Rev 1.1 1 98 99 - -
v0.6.1 FriendlyElec NanoPi M4 6 100 100 85 87
v0.6.1 FriendlyElec NanoPi NEO4 6 100 100 82 86
v0.6.1 FriendlyElec NanoPi M4 6 100 100 85 87
v0.6.1 FriendlyElec NanoPi NEO4 6 100 100 80 88
v0.6.1 ROCK PI 4B 6 100 100 86 90
v0.6.2 Khadas Captain 6 100 100 88 86
v0.6.2 ROCK PI 4B 6 100 100 87 86
v0.6.2 Olimex A64 Teres-I 4 100 100 83 97
v0.6.2 Cubietech Cubietruck 2 100 100 81 93
v0.6.2 4 100 100 93 99
v0.6.2 Khadas Captain 6 100 100 87 87
v0.6.2 Pine64 RockPro64 6 100 100 82 88
v0.6.9 Radxa ROCK Pi 4 6 100 100 89 86
v0.6.9 Pine H64 4 100 100 89 99
v0.6.9 Khadas VIM3L 4 100 100 88 98
v0.6.9 SolidRun i.MX8MQ HummingBoard Pulse 4 100 100 81 99
v0.7.4 RPi 400 Rev 1.0 4 100 100 92 98
v0.7.4 32 100 100 89 99
v0.7.5 1 100 100 - -
v0.7.5 1 100 100 - -
v0.7.5 Hugsun X99 TV BOX 6 99 100 90 89
v0.7.1 4 100 100 85 99
v0.7.2 Orange Pi Prime 4 100 100 88 98
v0.7.1 ODROID-C4 4 100 100 88 98
v0.7.2 Pine64 RockPro64 6 100 100 91 86
v0.7.2 Pine64 RockPro64 v2.1 6 100 100 94 86
v0.7.9 RPi Zero 2 Rev 1.0 4 100 100 81 99
v0.7.9 RPi Zero 2 Rev 1.0 4 100 100 75 99
v0.8.1 Nintendo Switch 4 97 98 76 97
v0.8.1 ODROID-N2Plus 6 99 100 93 84
v0.8.3 RPiB Rev 2 1 96 97 - -
v0.8.3 RPi Zero 2 Rev 1.0 4 100 100 76 99
v0.8.4 Olimex A20-OLinuXino-LIME2-eMMC 2 92 92 81 94
v0.8.4 RPi 4B Rev 1.1 4 100 100 84 99
v0.8.4 RPi 4B Rev 1.1 4 91 100 74 99
v0.8.4 RPi 4B Rev 1.1 4 100 100 88 99
v0.8.4 RPi 4B Rev 1.1 4 100 100 88 99
v0.8.5 RPi 4B Rev 1.1 4 100 100 88 99
v0.8.6 FriendlyARM NanoPi NEO4 6 100 100 92 86
v0.8.6 Odroid XU4 8 99 99 94 84
v0.8.8 RPi 4B Rev 1.1 4 100 100 86 99
v0.8.8 Radxa Zero 4 99 100 83 98
v0.9.0 RPi 4B Rev 1.4 4 100 100 91 97
v0.9.0 RPi 4B Rev 1.4 4 100 100 88 97
v0.9.1 ASUS Tinker Board 4 100 100 90 95
v0.8.7 RPi 4B Rev 1.4 4 100 100 90 98
v0.9.1 ODROID-N2Plus 6 99 100 92 84
v0.9.1 Generic RK322x TV Box board 4 100 100 85 96
v0.9.1 FriendlyElec NanoPi M4 Ver2.0 6 100 100 93 86
v0.9.1 BCM2835 1 99 98 - -
v0.9.1 BCM2836/BCM2709 4 100 100 77 99
v0.9.1 Allwinner A64 or https://tinyurl.com/yyf3d7fg 4 100 100 89 99
v0.9.1 Allwinner H3/H2+ 4 100 100 83 99
v0.9.1 Amlogic S905X2/S905Y2/S905D2/T962X2 4 99 100 82 98
v0.9.1 ODROID-HC4 4 99 100 84 98
v0.9.1 TRONFY MXQ S805 4 100 100 80 99
v0.9.1 ODROID-N2 6 99 100 90 91
v0.9.1 Khadas VIM 4 99 99 77 97
v0.9.2 SigmaStar SSD201/SSD202D 2 100 100 79 99
v0.9.2 Allwinner H3/H2+ 4 100 100 83 97
v0.9.2 2 x ThunderX CN8890 96 100 100 90 91
v0.9.2 Amlogic S905 4 99 100 80 98
v0.9.2 Allwinner H5 4 100 100 89 99
v0.9.2 BCM2711B0 4 100 100 91 99
v0.9.2 Amlogic S905X3 4 99 99 81 98
v0.9.2 Allwinner A20 2 92 92 81 94
v0.9.2 BCM2835 1 99 98 - -
v0.9.2 Amlogic S922X 6 99 100 94 84
v0.9.2 Rockchip RK3399 6 100 100 92 88
v0.9.2 Allwinner A20 2 100 100 81 94
v0.9.2 BCM2711B0 4 100 100 87 99
v0.9.3 Nvidia Jetson Nano 4 100 100 75 99
v0.9.2 2 x ThunderX CN8890 96 100 100 84 89
v0.9.2 2 x ThunderX CN8890 96 100 100 85 89
v0.9.3 Allwinner A20 2 92 92 81 95
v0.9.3 Amlogic S905X2/S905Y2/S905D2/T962X2 4 99 100 81 95
v0.9.3 Rockchip RK3399 6 100 100 93 86
v0.9.3 Amlogic Meson GXL (S905X) Revision 21:c (84:2) 4 99 100 75 98
v0.9.3 Amlogic Meson SM1 (S905X3) Revision 2b:c (10:2) 4 100 100 92 98
v0.9.3 Amlogic Meson GXBB (S905) Revision 1f:c (13:1) 4 96 96 74 94
v0.9.3 Amlogic Meson G12B (S922X) Revision 29:c (40:2) 6 99 100 94 84
v0.9.3 Amlogic Meson G12B (A311D) Revision 29:b (10:2) 6 99 99 94 84
v0.9.3 Amlogic Meson8m2 (S812) RevA (1d - 0:74E) detected 4 100 100 81 98
v0.9.3 Rockchip RK3568 (35681000) 4 100 100 91 98
v0.9.3 Amlogic Meson8m2 (S812) RevA (1d - 0:74E) detected 4 100 100 83 96
v0.9.3 Allwinner R40/V40 4 100 100 80 95
v0.9.3 Amlogic Meson SM1 (S905X3) Revision 2b:c (10:2) 4 99 100 81 98
v0.9.3 Rockchip RK3566 or RK3568 4 100 100 93 98
v0.9.3 Nvidia Jetson Nano 4 99 99 76 91
v0.9.3 Rockchip RK3568 (35682000) 4 100 100 91 98
v0.9.3 BCM2711B0 4 100 100 92 99
v0.9.3 Amlogic Meson SM1 (Unknown) Revision 2b:b (40:2) 4 99 100 89 98
v0.9.3 Amlogic Meson G12B (A311D) Revision 29:b (10:2) 6 99 99 93 84
v0.9.3 Khadas VIM4 8 99 99 93 90
v0.9.4 Amlogic A311D2 8 99 99 92 89
v0.9.4 Rockchip RK3288 4 100 100 87 99
v0.9.6 Rockchip RK3588 (35880000) 8 100 100 92 83
v0.9.6 Rockchip RK3588 (35880000) 8 100 100 92 86
v0.9.6 keeper.lan 16 100 100 92 98
v0.9.7 Nvidia AGX Xavier 6 98 99 84 98
v0.9.8 Rockchip RK3328 4 100 100 92 99
v0.7.7 Pine64 RK3566 Quartz64-A Board 4 99 99 83 99
v0.7.7 Radxa Zero 4 99 100 83 98
v0.9.8 Rockchip RK3188 4 95 97 81 94
v0.9.8 Allwinner D1 1 93 97 - -
v0.9.8 Rockchip RK3568 4 100 100 92 98
v0.9.8 Amlogic Meson GXM (S912) Revision 22:a (82:2) 8 100 100 87 88
v0.9.8 NXP i.MX6 Quad 4 97 98 79 97
v0.9.8 T-HEAD c910 ice 2 100 100 81 99
v0.9.8 / Celeron J1900 @ 1.99GHz 4 100 100 85 98
v0.9.8 Amlogic Meson8 (S802) RevC (19 - 0:27ED) 4 100 100 79 98
v0.9.8 Kendryte K510 2 100 100 71 99
v0.9.8 Apple MacBook Pro 10 100 100 90 77
v0.9.8 Phytium D2000 8 100 100 93 98
v0.9.8 Phytium D2000 8 100 100 93 98
v0.9.8 Silicom Minnowboard Turbot D0/D1 PLATFORM D0/D1 / Atom 2 99 99 85 98

Look at those SoCs/CPUs that are built for server tasks with 16 or even 2 x 48 cores. They can keep up with memory access since they’re designed for the task unlike cheap SoCs from the Android e-waste category!

4 x the same cores


400% usage in decompression what I only use.
Small difference in core sizes. 2.25 vs 2.3Ghz for cluster 2 and 3.

You see small difference in usage. 397%
When difference is larger you see a lower percentage.

All 8 cores and only 676% used of the 800%.
So my point is that using all core 7z b isn’t a reliable source. Better to do clusters seperate, or only single cores. That’s just my point.

Nice example of ignorance. Care to understand that there’s tons of examples above with ‘4 x the same cores’ that do not get a 400% decompression utilization for the simple reason that these SoCs originate from the Android e-waste world and chip internals are massively bottlenecking fully parallel operation of all cores.

Check ‘v0.9.8 Rockchip RK3188’, ‘v0.9.1 ASUS Tinker Board’ or the Amlogic S905 and RPi 4B results.

As such it’s the obvious result of adding more cores that CPU utilization further decreases since that stuff is too demanding. And actually CPU utilization is information – see the notes about ODROID XU4.

And why do you use only the decompression value? Which real-world use case is represented by this task?

Why the hell? Since you’re obsessed by 100% CPU utilization or is there another reason why you throw away useful information?

SoC/device compression decompression
RK3188 325 377
RK3288 354 399
RK3288 332 387
RK3288 310 385
RK3288 284 386
RK3288 302 386
RK3288 353 397
RK3288 323 395
RK3288 335 396
RK3288 359 381
RK3288 358 381
RK3288 364 381
S905 320 397
S905 310 393
S905 310 393
S905 321 394
S905 312 377
RPi 4B 354 398
RPi 4B 351 396
RPi 4B 368 393
RPi 4B 353 392
RPi 4B 318 354
RPi 4B 339 387
RPi 4B 350 385
RPi 4B 315 375
RPi 4B 284 381
RPi 4B 364 395
RPi 4B 354 399
RPi 4B 362 388
RPi 4B 363 395
RPi 4B 357 397
RPi 4B 347 398
RPi 4B 365 393
RPi 4B 348 395
RPi 4B 357 397
RPi 4B 340 372

Don’t you agree that CPU utilization is actual information?

And keep in mind that these are all quad-core SoCs with same core types, no ‘different core types’ or ‘uneven performance’ BS.

When looking at RK3288 alone and checking also the kernel version (again, check the ODROID XU4 notes in sbc-bench documentation) then it should be obvious why CPU utilization is information:

SoC/device compression decompression kernel
RK3288 354 399 5.10
RK3288 332 387 5.15
RK3288 310 385 5.15
RK3288 284 386 5.15
RK3288 302 386 5.15
RK3288 353 397 5.10
RK3288 323 395 5.15
RK3288 335 396 5.15
RK3288 359 381 5.15
RK3288 358 381 5.15
RK3288 364 381 5.15

7-zip’s internal benchmark is such a cheap and effective way to do regression testing but people (like the Armbian folks) still just ignore it.

You are correct, memory bandwidth seems important here.
I don’t trow the info away, I often show it, but not this time.(not a hardware review video but CPU benchmark)

I prefer to use Blender for multi-core CPU benchmark, and 7zip for single core. But I still do the multi-core and keep the data.

My theory was that 7z b always gives equal tasks. So for 8-cores. 8 equal tasks for each core. When performing them, the big cores are finished before the small cores.
So I might be wrong and am not afraid to admit this.

No idea what that has to do with this discussion? It is a discussion between you and me, Armbian is not involved in me making this video or my benchmarks.

Thank you for sharing the information. Have a nice day.

Bandwidth? You know the difference between bandwidth and latency?

Why? Also why do you use only the 7-zip decompression score (already asked 5 days ago)?