Permalink
Switch branches/tags
Nothing to show
Find file Copy path
d965f6a Nov 15, 2018
3 contributors

Users who have contributed to this file

@ThomasKaiser @igorpecovnik @g-provost
130 lines (115 sloc) 25.7 KB

Results

Below some results collected. Please keep in mind that these are NOT hardware performance numbers but depend on software/settings (see the differences kernel version makes for RockPro64 for example). The purpose of sbc-bench is to generate insights and not colorful graphs representing numbers without meaning. It's perfectly fine for the same hardware appearing multiple times with different numbers since those differ for a reason (software/settings).

Especially openssl numbers should be taken with a huge grain of salt since the benchmark numbers depend on kernel features and performance with other use cases (e.g. disk/filesystem encryption) might look differently.

So do not rely on collected numbers unless you carefully read through all the explanations and insights below and be prepared to conduct your own benchmarks if you really want to choose appropriate hardware for your use case.

Some numbers

Board Clockspeed Kernel Distro 7-zip AES-128 (16 byte) AES-256 (16 KB) memcpy memset kH/s URL
BPi R2 1300 MHz 4.4 Xenial armhf 2600 27550 25350 1500 3800 - http://ix.io/1iGV
Clearfog Pro 1600 MHz 4.14 Stretch armhf 2185 44500 43900 935 4940 - http://ix.io/1iFa
Helios4 1600 MHz 4.14 Stretch armhf 2210 44785 *1280 42500 *98560 910 4840 - http://ix.io/1jCy
Edge/Captain 2000/1500 MHz 4.4 Bionic arm64 6550 402150 1130400 2810 4860 10.50 http://ix.io/1rYm
EspressoBin 800 MHz 4.17 Stretch arm64 1138 54290 368330 1040 2490 1.23 http://ix.io/1kt2
EspressoBin 1200 MHz 4.18 Stretch arm64 1630 81900 555840 1000 2400 1.82 http://ix.io/1lCe
Le Potato 1410 MHz 4.18 Stretch arm64 3780 96680 657200 1810 5730 3.92 http://ix.io/1iSQ
Lime A10 910 MHz 4.14 Stretch armhf 550 25200 28250 440 1300 - http://ix.io/1j1L
NanoPC T3+ 1400 MHz 4.4 Xenial armhf 6400 143800 651000 1650 3700 - http://ix.io/1iyp
NanoPC T3+ 1400 MHz 4.14 Bionic arm64 7480 126000 652600 1440 4540 10.99 http://ix.io/1iRJ
NanoPC T4 1800/1400 MHz 4.17 Stretch arm64 6250 307200 1022500 4100 9000 8.24 http://ix.io/1iFz
NanoPC T4 1800/1400 MHz 4.17 Stretch arm64 6380 230280 1022600 4160 9000 9.36 http://ix.io/1iZq
NanoPC T4 1800/1400 MHz 4.17 Stretch arm64 6230 299600 1023600 4100 9060 10.30 http://ix.io/1iWU
NanoPC T4 2000/1500 MHz 4.4 Stretch arm64 5870 308370 1124040 2810 4890 8.70 http://ix.io/1lkG
NanoPi Fire3 1380 MHz 4.14 Bionic arm64 7440 126050 653000 1560 4600 10.96 http://ix.io/1jjm
NanoPi Fire3 1380 MHz 4.14 Stretch arm64 7420 95700 645400 1520 4570 8.53 http://ix.io/1jiU
NanoPi K1 Plus 1152 MHz 4.14 Stretch arm64 3030 78740 533380 1040 3070 3.32 http://ix.io/1m3x
NanoPi K2 1480 MHz 4.14 Stretch arm64 3850 43020 50370 1660 3870 4.61 http://ix.io/1iT1
NanoPi M4 2000/1500 MHz 4.19 Stretch arm64 6400 334650 1128330 4080 8270 8.86 http://ix.io/1lzP
NanoPi NEO4 2000/1500 MHz 4.4 Stretch arm64 6510 320600 1128860 2260 4770 8.71 http://ix.io/1oho
NanoPi NEO4 2000/1500 MHz 4.4 Stretch arm64 6030 342620 1121380 2230 4770 8.57 http://ix.io/1oib
NanoPi NEO4 2000/1500 MHz 4.4 Stretch arm64 6520 268720 1123190 2280 4770 8.83 http://ix.io/1oim
NanoPi NEO4 2000/1500 MHz 4.19 Stretch arm64 6750 278200 1139850 2370 6110 8.84 http://ix.io/1p3T
ODROID-C2 1750 MHz 3.14 Xenial arm64 4070 50500 48500 1750 3100 - http://ix.io/1ixI
ODROID-C2 1530 MHz 4.17 Stretch arm64 3870 43800 51280 1420 2600 4.63 http://ix.io/1iSh
ODROID-XU4 1900/1400 MHz 3.10 Jessie armhf 6750 74100 68200 2200 4800 - http://ix.io/1ixL
ODROID-XU4 2000/1400 MHz 4.9 Stretch armhf 6400 73350 72075 2230 4850 - http://ix.io/1iWL
ODROID-XU4 2000/1500 MHz 4.14 Bionic armhf 7100 74700 71500 2240 4880 - http://ix.io/1iLy
Orange Pi PC Plus 1300 MHz 4.14 Stretch armhf 2.880 20890 25270 900 3280 - http://ix.io/1j1d
Orange Pi Plus 2 1300 MHz 4.14 Stretch armhf 2.890 21480 25250 830 3240 - http://ix.io/1iX4
PineH64 900 (!) MHz 4.17 Stretch arm64 2550 62200 421000 1600 4840 2.84 http://ix.io/1iFT
PineH64 1800 MHz 4.18 Stretch arm64 4650 123400 836900 1380 5530 5.62 http://ix.io/1jEr
Renegade 1400 MHz 4.4 Stretch arm64 3710 95030 644200 1565 7435 3.92 http://ix.io/1iFx
Raspberry Pi 2 B+ 900 MHz 4.14 Debian Stretch 2070 14350 17450 615 1175 - http://ix.io/1iFf
Raspberry Pi 2 B+ 900 MHz 4.14 Raspbian Stretch 2130 14000 16300 1010 1170 - http://ix.io/1ivw
Raspberry Pi 3 B+ original 4.9 Raspbian Stretch 3600 35500 42700 1230 1640 - http://ix.io/1iI5
Raspberry Pi 3 B+ normal 4.14 Raspbian Stretch 3240 30500 36600 1130 1530 - http://ix.io/1ism
Raspberry Pi 3 B+ normal 4.14 Raspbian Stretch 3040 29500 36600 1050 1500 - http://ix.io/1iGM
Raspberry Pi 3 B+ UV/normal 4.14 Raspbian Stretch 2100 29500 36400 1040 1460 - http://ix.io/1iH0
Raspberry Pi 3 B+ OC/normal 4.14 Raspbian Stretch 3130 30500 36620 1230 1780 - http://ix.io/1iGz
Raspberry Pi 3 B+ with fan 4.14 Raspbian Stretch 3670 35800 42600 1120 1600 - http://ix.io/1isD
Raspberry Pi Zero 1000 MHz 4.14 Raspbian Stretch 450 13400 16820 400 1590 - http://ix.io/1niO
Rock64 1300 MHz 4.4 Bionic arm64 3410 89060 601200 1310 5680 4.46 http://ix.io/1iGW
Rock64 1300 MHz 4.18 Bionic arm64 3530 116100 605250 1340 5770 4.65 http://ix.io/1iH4
Rock64 1300 MHz 4.4 Stretch arm64 3430 88600 601000 1350 5680 3.64 http://ix.io/1iHo
Rock64 1300 MHz 4.18 Stretch arm64 3560 89070 603800 1340 5770 3.80 http://ix.io/1iHB
Rock64 1400 MHz 4.4 Stretch arm64 3610 95000 644250 1330 5700 3.85 http://ix.io/1iFm
Rock64 1400 MHz 4.4 Stretch arm64 3590 95000 643700 1320 5640 4.40 http://ix.io/1iZj
Rock64 1400 MHz 4.4 Stretch arm64 3580 94800 644380 1330 5680 4.63 http://ix.io/1iYK
Rock64 1400 MHz 4.4 Stretch armhf 3620 99400 624000 1430 3620 - http://ix.io/1iwz
Rock Pi 4B 1800/1400 MHz 4.4 Stretch armhf ~6250 261960 1007500 1900 4850 - http://ix.io/1pJi
Rock Pi 4B 2000/1500 MHz 4.4 Stretch armhf ~6450 301470 1113900 1870 4860 - http://ix.io/1rrO
RockPro64 1800/1400 MHz 4.4 Stretch arm64 6140 298800 1015600 2770 4850 8.14 http://ix.io/1lBC
RockPro64 1800/1400 MHz 4.4 Stretch armhf 6250 275000 1000150 2000 4835 - http://ix.io/1iFZ
RockPro64 1800/1400 MHz 4.18 Stretch arm64 6300 237700 1021500 3650 8450 8.20 http://ix.io/1iFp
Tinkerboard 1730 MHz 4.14 Stretch armhf 5350 63150 66600 1480 3900 - http://ix.io/1iSX
Vim2 1400/1000 MHz 4.9 Xenial arm64 4800 177600 659000 1690 5610 - http://ix.io/1ixi
Vim2 1400/1000 MHz 4.17 Bionic arm64 5450 126770 659600 1920 5920 8.59 http://ix.io/1iJ7
x5-Z8300 1420 MHz 4.9 Stretch amd64 3900 101580 178010 2380 2380 7.81 http://ix.io/1lgD
x5-Z8350 1920/1420 MHz 4.17 Manjaro amd64 4540 137900 237130 1970 1670 9.32 http://ix.io/1lBy
Celeron J3455 2300/1500 MHz 4.17 Stretch amd64 7000 316480 429660 4090 4050 17.26 http://ix.io/1m5p
Pentium N4200 2560/1100 MHz 4.14 Bionic amd64 7469 354328 468008 4682 4997 18.75 http://ix.io/1ngq
Pentium J4205 2560/1500 MHz 4.17 Stretch amd64 7570 355540 480640 5070 5170 18.82 http://ix.io/1m5t
Celeron J4105 2400/1500 MHz 4.15 Bionic amd64 9020 458670 697100 5500 7410 19.07 http://ix.io/1qal
Celeron J4105 2400/1500 MHz 4.15 Bionic amd64 8960 453860 697080 5620 7650 19.13 http://ix.io/1qb0

* Number obtained with cryptodev (Marvell's CESA).

Explanations

  • 7-zip number is an averaged multi threaded score from 3 consecutive 7z b runs. Only relevant for server workloads where stuff happens in parallel. Check the links for single threaded results (on big.LITTLE SoCs individually) to get an idea how most typical (single threaded) workloads perform
  • AES-128 (16 byte) is a single threaded encryption score with very small chunks of data (useful to get an idea how initialization overhead influences crypto performance with small packets). On big.LITTLE SoCs numbers show big core performance
  • AES-256 (16 KB) is a single threaded encryption score with rather huge chunks of data. On big.LITTLE SoCs numbers show big core performance
  • memcpy and memset are tinymembench measurements for memory bandwidth. On big.LITTLE SoCs numbers show big core performance
  • kH/s is a multi threaded cpuminer score showing the board's performance when executing NEON optimized code. To get the performance difference between big and little cores check the links in the right column
  • Clearfog Pro and Helios4 use exactly same SoC (Armada 385), kernel and clockspeeds and the only reason why OpenSSL numbers differ is since Helios4 numbers were made using Marvell's CESA crypto accelerator via cryptodev which provides nice speed improvements with larger block sizes but also some initialization overhead with tiny block sizes. Also CPU utilization is way lower so the SoC is free for other stuff while performing better at the same time.
  • EspressoBin's boot BLOB claims to run at up to 1GHz while real clockspeeds are lower maxing out with this setting at 790MHz (obviously a kernel bug -- see details)
  • NanoPi K1 Plus numbers are preliminary. Currently in Armbian highest cpufreq OPP is 1152 MHz and throttling tresholds are way too low, once this is unlocked (SoC capable of almost 1.4GHz) numbers will improve further
  • NanoPi NEO4 numbers: 1st result is from my NEO4 N°1 running with a NanoPi M4 image. This NEO uses the vendor supplied thermal pad between SoC and heatsink. 2nd number from my 2nd NEO4 this time using NEO4 settings (rk3399-nanopi4-rev04.dtb loaded) with a copper shim between heatsink and SoC which as usual improves 'thermal performance' a lot. Since memory bandwidth and especially latency is too low another test needed with my NEO4 N°2, this time again with M4 settings (rk3399-nanopi4-rev01.dtb loaded) and an additional fan. Memory performance restored, slightly better performance due to colder SoC. 4th result made with 4.19.0-rc4. Please be aware that RK3399 memory performance numbers differ alot between 4.4 and mainline kernel for yet unknown reasons!
  • PineH64 numbers are both preliminary -- with mainline kernel no cpufreq scaling was working in the beginning. Comparison of 900 MHz and 1800 MHz numbers allows to estimate influence of CPU clockspeeds on DRAM bandwidth measurements and so on.
  • RPi 3 B+ performance shown as original was measured with an older ThreadX release (6e08617e7767b09ef97b3d6cee8b75eba6d7ee0b from Mar 13 2018). Back then the 3B+ was faster than the 3B. This changed with a newer ThreadX release (4800f08a139d6ca1c5ecbee345ea6682e2160881 from Jun 7 2018) since RPi Trading people decided to trash performance on every RPi 3 B+ to masquerade instability issues on a fraction of boards (details)
  • RPi 3 B+ performance numbers shown as normal were made with no or just a heatsink (in contrast to with fan)
  • RPi 3 B+ marked as 'UV/normal' means: normal settings and average Micro USB cable resulting in UV (undervoltage). Once the demanding 7-zip benchmark started voltage dropped below 4.63V and 'frequency capping' (downclocking to 600 MHz) happened destroying performance. See the detailed log: 1400 MHz are reported by the kernel while it's 600 MHz in reality. Is this just highly misleading or already cheating?
  • RPi 3 B+ marked as 'OC/normal' means: OC (overclocked) settings, stable voltage but no fan used. Since SoC temperature exceeds 60°C the 'firmware' starts to cheat and downclocks to 1200 MHz while the kernel reports running at 1570 MHz. At least memory overclocking is somewhat effective.
  • Rock Pi 4B numbers are preliminary. Board has been tested without heatsink first so throttling occured as expected. Second time with higher cpufreq OPPs just a fan was added (fan without heatsink == pretty inefficient). Memory performance seems rather low but that's due to testing with vendor's armhf Linaro images -- see other RK3399 devices running same software stack, e.g. RockPro64 numbers above with kernel 4.4, armhf and also being limited to 1.8/1.4GHz.
  • Vim2 is somewhat special: not a real big.LITTLE design but 2 A53 clusters controlled by a firmware BLOB that allows cluster 0 to clock up to 1414 MHz (reported falsely as 1512 MHz) and cluster 1 able to reach 1 GHz (details)
  • x86 numbers are meant as comparison. x5-Z8300 numbers were made with UP Board, x5-Z8350 with Alfawise X5 Mini, Celeron J3455 with an ASRock J3455-ITX mainboard, Pentium N4200 on UP2 Board, Pentium J4205 on an ASRock J4205-ITX and Celeron J4105 on two ODROID-H2 with different DDR4-PC19200 (2400MT/s) SO-DIMMs (remotely accessed via maze.odroid.com)

Insights

  • Benchmarking the Raspberry Pi is useless when not taking into account that there always is a primary operating system running on the primary CPU (VideoCore) that fully controls the hardware. ARM cores are just guests here. That's why sbc-bench starting with v0.2 also logs ThreadX version and configuration (/boot/config.txt)
  • Looking at RPi 2 B+ numbers this is 2 times the same hardware, one time running latest Raspbian Stretch Lite and one time OMV/Armbian. Userland is both times Debian Stretch but Raspbian packages are built for ARMv6 while upstream Debian builds for ARMv7 (though with less effective compiler switches). Overall performance looks more or less the same except a very low memcopy bandwidth value with OMV. What's the reason since same ditro and kernel is used and same GCC to compile tinymembench? Is it firmware 'af8084725947aa2c7314172068f79dad9be1c8b4 from Apr 16 2018' vs. '47b05c853342eb6e4ea5b017d981e0ef247fb8be from Jul 3 2018'?
  • Looking at RPi 3 B+ numbers it's obvious that 'firmware' version is the most important factor. With original firmware (6e08617e7767b09ef97b3d6cee8b75eba6d7ee0b from Mar 13 2018) performance is ok just to get trashed after applying firmware 4800f08a139d6ca1c5ecbee345ea6682e2160881 from Jun 7 2018 which totally changes throttling behaviour. From then on you either need a fan for good performance or add a temp_soft_limit= entry to the firmware config file (we can't have a look what all those partially undocumented settings really do since RPi's main operating system is closed source)
  • tinymembench when executed on an A53 in an armhf userland compared to arm64 seems to generate lower memset numbers (78% on RK3399 -- see RockPro64 arm64 vs. RockPro64 armhf -- and 64% on RK3328 -- see Rock64 arm64 vs. Rock64 armhf). Status: needs further investigation and confirmation
  • Bionic vs. Stretch doesn't seem to make a difference with 7-zip scores. Applies to both armhf and arm64 too -- see Rock64 numbers above
  • 7-zip scores benefit slightly from memory performance. See RK3328 equipped Renegade at 1.4 GHz with 4.4 kernel and Rock64 with same setup
  • openssl numbers are not affected by memory performance and are the same with same CPU cores and same clockspeeds. At least with Cortex-A53 running at 1.4 GHz with a Debian Stretch arm64 binary: Le Potato, NanoPi Fire3, Renegade, Rock64 and RockPro64 with openssl pinned to an A53 core: ~96000k with AES-128/16bit and ~650000k with AES-256/16KB
  • It seems the combination arm64 Bionic with very recent kernel improves AES encryption results with small data chunks (less than 1KB -- see Rock64 with 4.18 at 1.3GHz and Vim2 with 4.17 at 1.4GHz vs. Rock64 with 4.4 at 1.3GHz). Status: Needs further investigations (most probably related to GCC version)
  • It seems running an armhf userland on 64-bit SoCs also improves AES encryption results with small data chunks (see armhf entries for NanoPC T3+, Rock64, RockPro64 and Vim2). Status: very interesting, needs further investigations
  • It seems running Xenial binaries even further improves AES/SSL performance when ARMv8 Crypto Extensions are available. Status: while interesting irrelevant, we should get rid of Xenial and Jessie numbers.
  • It makes a huge difference whether ARMv8 Crypto Extensions can be used or not. See the many 64-bit SBC results above and compare with 32-bit SoCs or RPi 3B+, ODROID-C2 and NanoPi K2 (the latter 3 basing on 64-bit ARMv8 SoCs without crypto engine licensed/available)
  • Bionic vs. Stretch makes a big difference with cpuminer. Libs and GCC versions obviously matter (GCC 7.3 on Bionic vs. 6.3 on Stretch -- some benchmarks heavily depend on compiler versions). Stretch with GCC 7.3 provides a 15% performance increase with cpuminer on RK3328 and RK3399 (see Rock64 and NanoPC T4 numbers above and there the logs to compare performance of big and little cores). With GCC 8.2 and Stretch it's 20% with RK3328 and even 25% with RK3399 (the A72 performance increasing more compared to the A53 cores -- check individual kH/s numbers in the logs)
  • (more to come soon)

The bigger picture

  • To compare different hardware exactly the same software environment (apps, libs, compiler, kernel) is needed. Ignoring this will produce numbers without meaning.
  • ARM's big cores (A15, A17, A72) perform a lot better than the little cores (A7, A53). Everything that needs high single threaded performance will hugely benefit from running on such a core. This puts SoCs like RK3288 (Tinkerboard), Exynos 5244 (ODROID XU4) or RK3399 in a better position. For the big.LITTLE designs a working HMP scheduler is mandatory since otherwise performance hungry tasks end up on a slow core. This is even true for pseudo big.LITTLE like on the Vim2/S912
  • 7-zip's benchmark still looks like a nice indicator for a 'server workloads' performance index (multi threaded tasks that do not rely on floating point arithmetics but partially on memory performance). Though these scores are totally irrelevant when it's about SBC use cases that focus on something different (e.g. a 'Desktop Linux' needing high single threaded CPU performance, HW accelerated GPU and VPU and also fast random IO on the rootfs)
  • We see a huge variation in tinymembench numbers with some boards outperforming others by magnitudes while the effect in reality for CPU bound workloads is rather minimal though high memory bandwidth is a requirement for certain other tasks (e.g. playing 4K video). At least numbers are there to generate further insights.
  • Identical SoCs perform more or less identical if 'environmental conditions' (clockspeeds) are the same -- see Renegade vs. Rock64 numbers or NanoPC T4 vs. RockPro64 or ODROID-C2 vs. NanoPi K2.
  • Same could be said for different Cortex-A cores. One A53 performs like the other as long as both run at the same clockspeed (with some exceptions most probably due to internal cache sizes -- see cpuminer numbers for Amlogic S905 vs. S905X/RK3328). With same count of cores you get similar performance (if the task(s) in question benefits from parallel execution)
  • Cortex-A53 running at the same clockspeed as A7 shows almost ~30% better performance (~3500 7-zip MIPS vs. ~2700). This is even true when running ARMv7 code (see RPi 3 B+ numbers). In general it seems irrelevant whether the A53 cores run an armhf or arm64 userland, some numbers are even higher when running armhf code. This is very interesting since there are scenarios where running an armhf userland results in needing way less physical memory for the same task while performing identical. Please note: it's about the userland (32-bit vs. 64-bit) and not kernel (64-bit of course)

TODO