Benchmarking the VIM3

[Electr1]

Hi community,
Here are some benchmarks collected from the VIM3,

These benchmarks consist of may different types CPU based, GPU based, OpenCL and Graphics,
based benchmarks,

Forenote:
The Graphics benchmarks would be a bit disappointing due to the fact that some of the Required drivers are not present and and may underperform with even capable Hardware,

About the VIM3:
The VIM3 SBC is the latest addition to our popular Khadas VIM series. It is equipped with the powerful Amlogic A311D SoC, based on the big-little CPU architecture: x4 Cortex A73 performance-cores clocked at 2.2Ghz, and x2 Cortex A53 efficiency-cores clocked at 1.8Ghz, are merged into a hexa-core configuration, and fabricated with a 12nm process to maximise performance, thermal and electrical efficiency.

VIM3 is designed with the same form-factor as our popular VIM2 and VIM1. It is the size of a credit-card, with all I/O ports located on one-side. Although diminutive, VIM3 is competitively-priced and fully-featured; everything you need is already built-in. The Pro-model comes with onboard 4GB LPDDR4/4X RAM, 32GB of EMMC storage and 2T2R 802.11ac Wi-Fi with Bluetooth 5.0.

In addition, VIM3 comes with unique features designed for our customers. Digital-signage developers will love the onboard programmable MCU, MIPI-CSI for connecting to cameras, real-time clock and wake-on-lan functionality. We’ve also included a tri-axis digital accelerometer to detect signage orientation. The M2X Extension Board will provide additional developer options such as power-over-ethernet, dual-ethernet, 4G and GPIO.

The new VIM3 SBC also goes beyond what our original VIM models have to offer. It has an onboard 5.0 TOPS NPU for neural network applications, an M.2 slot for NVMe SSDs that can switch between PCIe and USB 2.0 via Khadas KBI commands, and a USB-C port with 5-20 volt USB power-delivery input. Connect with up to 2 independent displays via HDMI and MIPI-DSI, one of which may be touch-enabled using the built-in TP connector.

more info in this link,

Testing Environment:
These tests were conducted in Open air with a fan and no heatsink,
The Ambient temp was about 34 °C and low Humidity,

Firmware used:
Linux tests: Ubuntu server-4.9-Focal fossa-20.04 (except SBC-bench, it was ubuntu-18.04)
Android test: V200319

You would ask No heatsink !, what was the thermal performance,
The results are quite surprising, The CPU never throttled below 95% performance and the Max clock was 2.20 GHz with few dips to 2.10 GHz,
Throttletest was run in android and here are the Results for the Thermal performance,

So let’s get to some results shall we,

Test 1: GLmark 2

glmark2 is an OpenGL 2.0 and ES 2.0 benchmark.
glmark2 is developed by Alexandros Frantzis and Jesse Barker based on the
original glmark benchmark by Ben Smith.

Glmark2

GLmark Score:48
1080p full screen score:29

Test 2: 3D mark ( slingshot )
This is an Android Application used to test Graphics via OpenGL3.0 and The CPU using the physics test

3D mark ( Slingshot ) Score: 1663

Test 3: CLpeak
A synthetic benchmarking tool to measure peak capabilities of opencl devices. It only measures the peak metrics that can be achieved using vector operations and does not represent a real-world use case

Test 4: SBC-bench
This test was made by thomas kaiser, thanks to him
More about sbc-bench

sbc-bench v0.6.7

Installing needed tools. This may take some time… Done.
Checking cpufreq OPP… Done.
Executing tinymembench. This will take a long time… Done.
Executing OpenSSL benchmark. This will take 3 minutes… Done.
Executing 7-zip benchmark. This will take a long time… Done.
Executing cpuminer. This will take 5 minutes… Done.
Checking cpufreq OPP… Done.

Memory performance (big.LITTLE cores measured individually):
memcpy: 2270.3 MB/s
memset: 7465.5 MB/s
memcpy: 4983.3 MB/s
memset: 9304.4 MB/s (0.7%)

Cpuminer total scores (5 minutes execution): 13.12,13.11,13.10,13.09 kH/s

7-zip total scores (3 consecutive runs): 8605,8608,8474

OpenSSL results (big.LITTLE cores measured individually):
type 16 bytes 64 bytes 256 bytes 1024 bytes 8192 bytes 16384 bytes
aes-128-cbc 161383.40k 476855.04k 910941.70k 1215644.67k 1346942.29k 1355251.71k
aes-128-cbc 398365.98k 971936.34k 1465701.46k 1669396.48k 1752058.54k 1758106.97k
aes-192-cbc 153635.30k 420409.24k 742643.80k 936720.73k 1013822.81k 1016441.51k
aes-192-cbc 368655.52k 874884.01k 1235687.00k 1399264.94k 1461324.46k 1465789.10k
aes-256-cbc 148855.81k 388716.91k 642911.15k 783470.25k 836657.15k 839680.00k
aes-256-cbc 356255.06k 797466.37k 1099148.80k 1207970.82k 1253741.91k 1256909.48k

Full results uploaded to http://ix.io/1MFD. Please check the log for anomalies (e.g. swapping
or throttling happenend) and otherwise share this URL.

Thanks to some of the forum Members for helping me with some of the data collection,
Useful for your reference, to compare to other SBC’s etc.

have a good day !

[Spikerguy]

@Electr1

GLmark Score: 48

Looks good but as per the screenshoot, It looks like it is not in fullscreen mode.
Can you try glmark2 -s 1920x1080 and share the result.

Thank you for the benchmarking data.

[Electr1]

Sure will update it,

[Electr1]

for 1980x1080 the result is 29 points
but the problem is my TV which I am testing is more like 1600x900, really weird resolution. and I can’t see the entire screen, quite a problem.

[Vladimir.v.v]

no doubt, it will be better with active cooling, but the task is to prevent the fan from being heard, I have a quiet fan, otherwise I would go crazy

[Electr1]

Oh, it got active cooling all right,

the fan spits out Air at a speed of 20 CFM, (20 cubic feet of air per minute)

here is the link if you want to buy : https://www.alibaba.com/product-detail/Waterproof-Mini-Brushless-DC-Axial-Fan_60103740734.html?spm=a2700.icbuShop.41413.23.7b6029c7ilFLLE

I agree with the fan being silent is a much better situation, the last time I tried making a cooling fan with a coreless heli. motor and a 30 mm prop laying around, it screamed so loud that I swore that either I use a proper cooling fan or I would not use cooling at all, (my yet to learn more hobby took its toll on my eardrums)

[Vladimir.v.v]

yes, I know that you got a fan, from your other post

[Electr1]

Yep, it works like a charm, not the most beautiful beast in the room but gets the job done.
I was I a dilemma as to how to cool it, if I didn’t cool it when I was running one of high load applications on it it got super hot and I was scared I might cause permanent irreversible damage to the device all together,

Due to the situation prevailing I could not just buy a Khadas 3705 fan and heatsink as it would not arrive, because external country imports are not allowed right now, If I wanted to buy some thing from outside my country, I would have to wait for about 1 year for consumer imports to be allowed ! Yikes!

oh and plus the 30$ shipping charge for a country that that is nearly 0 KM apart

[Vladimir.v.v]

In any case, I think you did not lose

[Electr1]

yep totally worth the 300 ₹ I spent on it, ( 300 ₹ ≈ 4 USD )

[bizcocho85]

Amazing!! Thank you very much, @Electr1

I compared the Test 4 with the results from the github project and they match completely.

Also, it is amazing the clpeak results.

Looking forward to have OpenCL support for the CPU too, so, we can test clpeak, hashcat and another one to do load-balancing and compare the co-execution (I can propose to you). If you don’t mind, of course.

Great work!

Question 1: Why you need to use ubuntu-18.04 to execute SBC-bench?
Q2: why did you use the kernel 4.9 and not a 5.X kernel?

[Electr1]

@bizcocho85,
1. SBC-bench would fail to start due to reason that some of the testing software it uses is not present in this version (eg. tinymembench, or something like that and refused to work ), also it requires modification as it will not execute if the system load is above 0.1 % which earthly not possible as in being idle in headless mode it has Avg. load of minimum 2 % !, due to time constraint I actually referred to a previous test done by @numbqq

ref. here to the original post, original post

2. I have android in my eMMC, which means using Mainline kernel would give me the green screen problem, this might have been slightly fixed but for sake I have used it.

[Electr1]

Sure I am open to suggestions, I am a bit busy myself, but I will try to reply to you ASAP,
If it is openCL on CPU I will post that as well I just need to get POCL installed,

[Electr1]

Got of a bit of a Hashcat performance test done before I doozed off for the night ,
Take a look ,

Will upload the full test tomorrow morning,

[Electr1]

Here is the test I did this morning, the results are better,
but the process gets killed on the 5th test,

[bizcocho85]

Sure, thank you very much. Really interesting.

Did you kill the process or it got killed somehow?

[Electr1]

nope it got killed I am trying in the mainline kernel, will release results soon

[clort76]

I ran the Hardinfo 0.6-alpha benchmarks today.

I also note: While the Glmark2 score shown in a post above is not revealing usable game performance, it’s worth noting that the raspberri Pi4 WITH GPU only gets 150 - 195 in GLES under Xorg!!

For comparison the Nvidia Jetson Nano scores 2120 in OpenGL2 under Xorg!!

Excited to see the first Panfrost driver results on VIM3!

[Electr1]

Thank you for posting this data here @clort76

[clort76]

btw i get better glmark2

                              glmark2 Score: 108 

=======================================================