|Project Name||Stars||Downloads||Repos Using This||Packages Using This||Most Recent Commit||Total Releases||Latest Release||Open Issues||License||Language|
|Sbc Bench||486||a day ago||1||bsd-3-clause||Shell|
|Simple benchmark for single board computers|
|Mixbench||211||3 months ago||7||gpl-2.0||C++|
|A GPU benchmark tool for evaluating GPUs and CPUs on mixed operational intensity kernels (CUDA, OpenCL, HIP, SYCL, OpenMP)|
|Matrixmultiply||168||415||11||a month ago||24||November 20, 2021||7||apache-2.0||Rust|
|General matrix multiplication of f32 and f64 matrices in Rust. Supports matrices with general strides.|
|Freqbench||139||4 days ago||1||mit||Python|
|Comprehensive CPU frequency performance/power benchmark|
|Xup_vitis_network_example||97||a day ago||8||other||Jupyter Notebook|
|VNx: Vitis Network Examples|
|Krun||65||2 years ago||12||other||Python|
|High fidelity benchmark runner|
|Coremark Pro||57||2 years ago||3||other||C|
|Containing dozens of real-world and synthetic tests, CoreMark®-PRO (2015) is an industry-standard benchmark that measures the multi-processor performance of central processing units (CPU) and embedded microcrontrollers (MCU)|
|Sputnik||39||3 years ago||5||apache-2.0||C++|
|A library of GPU kernels for sparse matrix operations.|
|Shfllock||35||2 years ago||C|
|Numpy Benchmarks||30||2 years ago||1||bsd-3-clause||Python|
|A collection of scientific kernels using the numpy module for benchmarking purpose|
SBC is a shortcut for single-board computer and this whole repository is about performance considerations around those devices (with an initial focus on energy efficient server tasks).
This small set of different CPU performance tests focuses on 'headless' operation only (no GPU/display stuff, no floating point number crunching). Unlike many other 'kitchen-sink benchmarks' it tries to produce insights instead of fancy graphs.
It has eight entirely different usage modes:
-Tswitches (details/discussion, another example)
-gto measure efficiency of settings/devices
sbc-bench -G) or Phoronix (
sbc-bench -kshows kernel version info. Stuff like: still supported? BSP or mainline?
-R) are designed to help reviewers and participants of 'SBC debug parties' to quickly identify tunables and bottlenecks that need further attention: reports many performance relevant settings, switches them to max performance and lurks from then on in the background to monitor other benchmark executions and tests. By comparing scores made with defaults we are able to directly identify settings that need adjustments
The monitoring now also displays some hardware information when starting:
[email protected]:~$ sbc-bench -m Samsung Exynos EXYNOS5800 rev 1, Exynos 5422, Kernel: armv7l, Userland: armhf CPU sysfs topology (clusters, cpufreq members, clockspeeds) cpufreq min max CPU cluster policy speed speed core type 0 1 0 200 1400 Cortex-A7 / r0p3 1 1 0 200 1400 Cortex-A7 / r0p3 2 1 0 200 1400 Cortex-A7 / r0p3 3 1 0 200 1400 Cortex-A7 / r0p3 4 0 4 200 2000 Cortex-A15 / r2p3 5 0 4 200 2000 Cortex-A15 / r2p3 6 0 4 200 2000 Cortex-A15 / r2p3 7 0 4 200 2000 Cortex-A15 / r2p3 Thermal source: /sys/devices/virtual/thermal/thermal_zone0/ (cpu0-thermal) Time big.LITTLE load %cpu %sys %usr %nice %io %irq Temp 18:18:33: 800/ 500MHz 0.00 18% 0% 17% 0% 0% 0% 25.0°C 18:18:38: 800/ 600MHz 0.00 0% 0% 0% 0% 0% 0% 24.0°C 18:18:43: 700/ 500MHz 0.07 0% 0% 0% 0% 0% 0% 24.0°C 18:18:48: 800/ 600MHz 0.07 0% 0% 0% 0% 0% 0% 24.0°C ^C
The SoCs (system-on-chip) used on today's SBC are that performant that heat dissipation when running full load for some time becomes an issue. The strategies to deal with the problem differ by platform and kernel. We've seen CPU cores being shut down when overheating (Allwinner boards running with original Allwinner software), we know platforms where throttling works pretty well but by switching to a different kernel performance is trashed on exactly the same hardware. Sometimes it's pretty easy to spot what's going on, sometimes vendors cheat on us and it takes some efforts to get a clue what's really happening.
This tool therefore focuses on a controlled environment and intensive monitoring running in the background and being added to results output. The tool returns with a brief performance overview (see screenshot above) but the real information will be uploaded to an online pasteboard service (Rock 5B example). Without checking this detailed output numbers are worthless (since we always need to check what really happened).
You need Debian Stretch/Buster/Bullseye or Ubuntu Bionic/Focal/Jammy. Older variants are not supported (due to distro packages being way too outdated). Then it's
wget https://raw.githubusercontent.com/ThomasKaiser/sbc-bench/master/sbc-bench.sh sudo /bin/bash ./sbc-bench.sh -c
Unfortunately to adjust the cpufreq governor and to collect monitoring data execution as root is needed. So do not run this on productive systems or if you don't understand what the script is doing.
I chose mhz, tinymembench, ramlat, cpuminer, stockfish, 7-zip and OpenSSL's AES benchmarks for the following reasons:
This tool is not a benchmark but instead measures real CPU clockspeeds. This is helpful on platforms where cpufreq support is not available yet or we can not rely on the clockspeed values returned by the kernel. This applies to platforms where vendors are cheating (RPi, Amlogic) or where actual clockspeeds are set via jumpers while the clockspeeds available to the kernel are derived from device-tree (DT) entries. On a Clearfog Pro routerboard it will look like this for example (DT defines 666/1332 MHz while I configured 800/1600 MHz via jumper):
Checking cpufreq OPP: Cpufreq OPP: 1332 Measured: 1599 (1598.621/1598.759/1598.324) (+20%) Cpufreq OPP: 666 Measured: 799 (799.502/798.295/799.115) (+20%)
mhz twice. At the begin of the benchmark with an idle and cold system walking through all cpufreq OPP and directly after the most demanding benchmark has finished with the device still under full load to see whether behaviour changes when SoC is overheated. This is on a Thundercomm Dragonboard 845c. Prior to benchmark execution it looked like this:
Checking cpufreq OPP for cpu4-cpu7 (Qualcomm Kryo 3XX Gold): Cpufreq OPP: 2803 Measured: 2704 (2705.057/2704.717/2704.717) (-3.5%) Cpufreq OPP: 2649 Measured: 2704 (2704.830/2704.717/2704.717) (+2.1%)
When running the multi-threaded 7zip benchmark, the SoC temperature exceeds 80°C and afterwards the 2803 MHz cpufreq OPP is gone while the reported 2649 MHz are in reality only ~1940:
Checking cpufreq OPP for cpu4-cpu7 (Qualcomm Kryo 3XX Gold): Cpufreq OPP: 2649 Measured: 1940 (1955.570/1943.795/1922.274) (-26.8%)
Unlike other 'RAM benchmarks' tinymembench checks for both memory bandwidth and latency in a lot of variations so it's even possible to get some insights about internal cache sizes. It also measures each mode at least two times and if sample standard deviation exceeds 0.1%, it is shown in brackets next to the result. So it's pretty easy to spot background activity ruining benchmark results.
On hybrid systems with different CPU cores (big.LITTLE, DynamicIQ, Alder/Raptor Lake) we pin execution one time to an efficiency/little and one time to a performance/big core to know the difference this makes. For the sake of simplicity we output memcpy and memset numbers at the end of the benchmark. On an overclocked RPi 3 B+ (arm_freq=1570, over_voltage=4, core_freq=500, sdram_freq=510, over_voltage_sdram=2) this will look like this
Memory performance: memcpy: 1316.0 MB/s (0.8%) memset: 1933.9 MB/s
On a NanoPC T4 (RK3399, 2xA72/4xA53 CPU cores) this will look like this with mainline kernel and conservative settings without any optimizations yet:
Memory performance: memcpy: 2054.9 MB/s memset: 8453.0 MB/s (0.2%) memcpy: 4238.8 MB/s (0.4%) memset: 9082.5 MB/s (0.9%)
(first two lines show execution on a little A53 core, the last ones when pinned to an A72 big core)
On ARM SoCs CPU and GPU/VPU usually share memory access so it's worth a try to experiment with disabling HDMI/GPU for headless use cases. Often memory bandwidth and therefore overall performance increases. Same when switching between kernel branches.
Provides some insights about cache sizes/speed and memory latency/bandwidth. Stuff like this.
Prior to adding stockfish on most platforms this was the most demanding benchmark of the six and pretty efficient to check for appropriate heat dissipation and even instabilities under load. It makes heavy use of SIMD optimizations (NEON on ARM and SSE/AVX on x86) therefore generating more heat than unoptimized 'standard' code.
Heavy SIMD optimizations aren't really common, the generated scores depend a lot on compiler version and therefore this test is optional. Unless you execute
sbc-bench -c or with
MODE=extensive it will be skipped since results can be misleading. So consider this being a load generator to check whether your board will start to throttle or becomes unstable but take the benchmark numbers with a grain of salt unless you're a programmer and know what NEON, SSE and AVX really are and whether your application can make use of.
A typical result (Rock 5B with Ubuntu Focal) will look like this:
Cpuminer total scores (5 minutes execution): 25.32,25.31,25.30,25.29,25.28,25.12 kH/s
(result variation in this case is ok since all results are more or less the same. If the board would've started throttling or heavy background activitiy would've happened the later numbers would be much lower than the first ones)
Stockfish (open source chess engine) also makes heavy use of SIMD extensions but is heavy on memory access too putting even more load on devices than cpuminer which doesn't access RAM that much or at all since working set fits inside CPU caches.
As with cpuminer this test is optional (
sbc-bench -s or
MODE=extensive needed) since not representing any broader use case but being more of a stressor / load generator exposing thermal and stability issues. Consumption figures are higher compared to cpuminer since stockfish also stresses the DRAM interface and at least it's sufficient to expose a reliability issue with Rock 5B (most probably today RK3588 in general) since running this benchmark reliably freezes Rock 5B at 2112 MHz DRAM clock.
7-zip's internal benchmark mode is a pretty good representation of 'server workloads in general'. When running on all cores in parallel it doesn't utilize CPU cores fully (at least not on ARM, on x64 with Hyperthreading it's a different story), it depends somewhat on memory performance (low latency more important than high bandwidth) and amount of available memory. When running fully parallel on systems that have many cores but are low on memory we see just as in reality the kernel either killing processes due to 'out of memory' or starting to swap if configured.
On big.LITTLE systems we start with one run pinned to a little core followed by one pinned to a big core. Then follow 3 consecutive runs using all available cores. The results might look like this:
7-zip total scores (3 consecutive runs): 3313,3285,3050 7-zip total scores (3 consecutive runs): 3613,3598,3633 7-zip total scores (3 consecutive runs): 7382,7407,7426
(this is a RPi 3 B+ with latest firmware update applied destroying performance showing throttling symptoms followed by a Rock64 at 1.4GHz with Armbian standard settings passively cooled by small heatsink followed by an octa-core NanoPi Fire3 also at 1.4 GHz but with heatsink and fan this time)
How to interpret 7-zip MIPS scores: 7-zip ist all about integer CPU and memory performance. And by looking at the 'total score' (running on all CPU cores in parallel) you need to keep in mind that only a few use cases are really parallel and limited to 'integer performance'. That's why it's written 'server workloads in general' above since this applies here and overall performance scales well with count of CPU cores.
If your use case is different (desktop, rendering, video editing, number crunching and so on that either depends more on single-threaded performance and/or involves floating point arithmetic, vector extensions or GPGPU), 7-zip MIPS are rather irrelevant for you since they do not even remotely represent your use case!
With 'server workloads' in mind 7-zip MIPS give an estimate of what to expect. A system showing two times more 7-zip MIPS compared to another will be able to run more (maybe twice as much or even more) daemons/tasks as long as the stuff is only CPU bound. How an individual daemon/task performs is a totally different story and needs to be checked (single-core 7-zip MIPS are available via left column in results list).
With a system scoring 125% compared to another it's a different story and you need to examine individual results and your use case closely (time to switch from staring at numbers to Active Benchmarking).
A nice example is comparing two ARMv8 server designs: 32 Neoverse-N1 cores (Amazon m6g.8xlarge VM) vs. 96 ThunderX1 cores (dual CPU ThunderX CN8890 blade). Both systems share an identical multi-core score (~110000 7-zip MIPS) but any real server workload will perform better on the Neoverse-N1 design. Single-threaded performance there is at least twice as high, memory performance way better and this will make the difference with real-world stuff unless the use case is really all about 100% CPU utilisation on all cores all the time.
If those 7-zip MIPS apply only to a few selected use cases as performance indicator why are they used in sbc-bench?
sysbench cpubenchmark on the other hand 'performs' 10-15 times better on a 64-bit Raspbian which is not related to 64-bit vs. 32-bit but just due to ARMv8 ISA containing a
sysbench cpufor example runs completely inside CPU caches). With this benchmark it's easy to spot memory performance issues like this (after switching bootloaders DDR4 RAM got clocked with just 333 instead of the former 1056 MHz). It's one of the 'cheapest' tools for regression testing but unfortunately not widely used there
A good example for the latter is Odroid XU4, three times tested with different kernel and OS versions (Stretch, Bionic and Focal which all build packages with different GCC versions). Memory performance remained the same (for a way to quickly check this see included script snippets) but for whatever reasons only the multi-threaded performance fluctuated over time:
|Kernel / Compiler||7-zip single||7-zip multi||CPU utilisation compression||CPU utilisation decompression|
|Kernel 4.9 / GCC 6.3||1622||6370||64%||78%|
|Kernel 4.14 / GCC 7.3||1633||7100||64%||78%|
|Kernel 5.4 / GCC 9.3||1604||8980||94%||84%|
Smells like a scheduler problem with kernel 4.x. Only more detailed tests with more kernel/GCC combinations or switching to Active Benchmarking could really tell.
This test solely focuses on AES performance (VPN use case, full disk encryption). The test tries to quickly confirm whether an ARM SoC can make use of special crypto engines. Some SoC vendors don't care, some add proprietary engines to their SoCs (Marvell's CESA as an example), some vendors chose to license ARM's 'ARMv8 Crypto Extensions' (see here for some insights). So in case a board runs with an 64-bit ARM SoC this simple test shows the presence of crypto extensions or not.
Results might look like this on an overclocked Raspberry Pi 3 B+ at 1570 MHz lacking any crypto acceleration:
type 16 bytes 64 bytes 256 bytes 1024 bytes 8192 bytes aes-128-cbc 39393.73k 54173.16k 60220.67k 61720.92k 62518.61k aes-192-cbc 35676.65k 46311.68k 51358.21k 52840.11k 53157.89k aes-256-cbc 33339.62k 42962.13k 46476.37k 47619.07k 47925.93k
Vs. an Orange Pi Zero Plus based on Allwinner H5 heavily underclocked at just 816 MHz:
type 16 bytes 64 bytes 256 bytes 1024 bytes 8192 bytes aes-128-cbc 102568.41k 274205.76k 458456.23k 569923.58k 613422.42k aes-192-cbc 95781.66k 235775.72k 366295.72k 435745.79k 461294.25k aes-256-cbc 91725.44k 211677.08k 313433.77k 362907.31k 380482.90k
ARMv8 Crypto Extensions make the difference here. Even at almost half the CPU clockspeed with small data chunks at least 2.5 times faster and up to 9 times faster with larger chunks. Looking at different chunk sizes makes a lot of sense since some proprietary crypto engines suffer from high initialization overhead. See these numbers for a Banana Pi R2 based on a MediaTek MT7623 with proprietary crypto engine after compiling own kernel and OpenSSL (sources):
type 16 bytes 64 bytes 256 bytes 1024 bytes 8192 bytes aes-128-cbc 519.15k 1784.13k 6315.78k 25199.27k 124499.22k aes-192-cbc 512.39k 1794.01k 6375.59k 25382.23k 118693.89k aes-256-cbc 508.30k 1795.05k 6339.93k 25042.60k 112943.10k
Benchmarking a system that is otherwise busy will result in numbers without meaning. Therefore it's important to ensure the system is as idle as possible. That's the reason
sbc-bench will only start once '1 min average load' is reported as below 0.1 or CPU utilization less than 2.5% for 30 seconds:
Of course this is not sufficient since background tasks might become active later or cron jobs result in some peak activity in between. As much such services as possible should be stopped prior to benchmark execution or in best case a rather minimal image should be used for testing. On the other hand
sbc-bench can also easily be used to compare 'desktop' and 'minimal' images.
But comparisons only make some sense if execution of the benchmark can be observed. That's what
sbc-bench's background monitoring is for that will be appended to detailed result list. See this example for Rock64. We can there look for the following problems:
The 7-zip benchmark when running on all cores can result in the system starting to swap when running low on memory. A good example for an affected board is the inexpensive NanoPi Fire3 with 8 A53 cores but just 1 GB DRAM. When we search in the detailed result output for Swap we'll find 2 occurences. One check prior to the benchmarks and one afterwards. With a Fire3 this might look like:
Swap: 495M 0B 495M Swap: 495M 34M 460M
So we know swapping has happened which negatively affected performance to some degree based on how swap is implemented. If swapping to SD card is configured performance will be severely impacted but in this case since it's a recent Armbian image the effects are negligible since Armbian implements zram based swap in the meantime (that's why kind of swap is also recorded in detailed result list).
While executing the multi-core 7-zip benchmark monitoring looked like this:
System health while running 7-zip multi core benchmark: Time big.LITTLE load %cpu %sys %usr %nice %io %irq Temp 10:50:25: 1400/1400MHz 6.23 9% 0% 8% 0% 0% 0% 44.0°C 10:50:58: 1400/1400MHz 5.16 50% 0% 50% 0% 0% 0% 54.0°C 10:51:29: 1400/1400MHz 5.63 74% 0% 73% 0% 0% 0% 58.0°C 10:52:00: 1400/1400MHz 6.23 80% 0% 79% 0% 0% 0% 59.0°C 10:52:31: 1400/1400MHz 6.39 72% 0% 71% 0% 0% 0% 56.0°C
Always 0% in the
%io column reported so not a big deal. With swap on SD card especially when using cards with low random IO performance we would've seen high occurences of %iowait activity and way lower performance numbers.
We have 3 benchmark executions that run completely single threaded: tinymembench, the first 7-zip run limited to a single CPU core and the openssl test. In all these cases the overall
%cpu percentage has to match count of CPU cores (the first two lines can be ignored). So on an octa-core board like NanoPi Fire3 it has to show exactly 12% and nothing more:
Time big.LITTLE load %cpu %sys %usr %nice %io %irq Temp 10:40:05: 1400/1400MHz 0.18 2% 0% 0% 0% 1% 0% 40.0°C 10:41:05: 1400/1400MHz 0.63 10% 0% 10% 0% 0% 0% 44.0°C 10:42:05: 1400/1400MHz 0.94 12% 0% 12% 0% 0% 0% 44.0°C 10:43:05: 1400/1400MHz 0.98 12% 0% 12% 0% 0% 0% 40.0°C 10:44:05: 1400/1400MHz 0.99 12% 0% 12% 0% 0% 0% 40.0°C 10:45:05: 1400/1400MHz 1.00 12% 0% 12% 0% 0% 0% 40.0°C 10:46:06: 1400/1400MHz 1.04 12% 0% 12% 0% 0% 0% 40.0°C
On a dual-core board we're talking about 50% max, on hexa-cores it's 16% and on a quad-core board it must not exceed 25% (100 / 4):
Time CPU load %cpu %sys %usr %nice %io %irq Temp 10:18:10: 1392MHz 1.05 17% 2% 15% 0% 0% 0% 59.5°C 10:19:10: 1392MHz 0.95 21% 0% 21% 0% 0% 0% 62.5°C 10:20:10: 1392MHz 1.02 25% 0% 25% 0% 0% 0% 61.7°C 10:21:10: 1392MHz 1.13 27% 1% 26% 0% 0% 0% 59.5°C 10:22:10: 1392MHz 1.05 25% 0% 25% 0% 0% 0% 60.0°C 10:23:10: 1392MHz 1.09 25% 0% 25% 0% 0% 0% 61.2°C 10:24:10: 1392MHz 1.03 25% 0% 25% 0% 0% 0% 61.7°C
In this case we were able to spot some background activity in this line:
10:21:10: 1392MHz 1.13 27% 1% 26% 0% 0% 0% 59.5°C
$something happened in parallel which will slightly lower the generated benchmark score. While 2% CPU utilisation for other stuff won't hurt that much at least we need to have an eye on this since when there are higher utilisation numbers reported when running the single threaded stuff the system shows way too much background activity to report reasonable benchmark scores. Then we simply generated numbers without meaning.
Depending on settings (kernel or some 'firmware' controlling the hardware) the clockspeeds might be dynamically reduced when the SoC starts to overheat. When clockspeeds are reduced then this obviously slows down operation.
sbc-bench continually monitors the clockspeeds but since we can only query every few seconds we might not catch short clockspeed decreases. That's why we check whether the kernel's cpufreq driver supports statistics. If true we record contents of
stats/time_in_state prior to and after benchmark execution and compare afterwards. This way we are able to detect even short amounts of downclocking which will result in a warning like this: ATTENTION: Throttling occured. Check the log for details.
The detailed log then will contain information how much time (in milliseconds) has been spent on which clockspeed while executing the benchmarks. Might look like this on a NanoPC T4 without fan (only vendor's heatsink) after running the full set (NEON test included which resulted in the big cluster clocking down to even 408 MHz):
Throttling statistics (time spent on each cpufreq OPP) for CPUs 4-5: 1800 MHz: 1344.39 sec 1608 MHz: 372.95 sec 1416 MHz: 117.69 sec 1200 MHz: 48.28 sec 1008 MHz: 41.58 sec 816 MHz: 55.24 sec 600 MHz: 127.08 sec 408 MHz: 352.72 sec
Important: to get throttling notifications running a kernel with
CONFIG_CPU_FREQ_STAT=y is needed since otherwise cpufreq statistics are not available. And this will not work on Raspberries since there cpufreq driver has not the slightest idea what's going on.
And all of this doesn't work reliably on
x86_64. Here you need to check
stockfish scores. If they got lower during execution your device ran into thermal or powercapping issues.
sbc-bench should benchmark in an automated fashion then exporting
MODE=unattended prior to execution will prevent warning dialogs but of course
sbc-bench will still check whether average load or CPU utilization is too high and refuse to start since benchmarking a busy system is useless.
Everything sent to
stdout can be ignored (but parsing for 'check the log' is highly recommended since hinting at too much background activity and/or swapping resulting in numbers without meaning instead of benchmark scores). Full benchmark results are available at
/var/log/sbc-bench.log with the last line containing a performance summary. So something like this could be used for regression testing and similar stuff:
MODE=unattended sbc-bench.sh -c | grep -q 'check the log' || tail -n1 /var/log/sbc-bench.log
MODE=extensive (not compatible with
MODE=unattended so use either/or) then
sbc-bench conducts additional tests:
opensslbenchmarks will also be executed in parallel on all CPU cores (takes an additional minute)
cpuminertest will be fired up (5 more minutes)
stockfishstress tester will be fired up 3 times to check further for throttling and stability issues
7-ziptests per cluster are done (no duration estimate possible since depends on SoC architecture)
This operation mode will be extended further over time to get insights into SoC internals.
$MaxKHz is exported prior to benchmark execution (e.g. by
MODE=extensive MaxKHz=1416000 sbc-bench.sh) then cpufreq OPP higher than this value are skipped. On many platforms this allows CPU core comparisons at same clockspeeds (e.g. limiting all cores to 1.8 GHz on RK3588 or 1.4 GHz on RK3399). For a list of available values check
$CPUINFOFILE is exported prior to benchmark execution then SoC guessing and similar stuff happens not based on
/proc/cpuinfo but on the supplied file that obviously needs to have a compatible format.