Custom PC - UK (2020-05)

Not that a lack of networking is an issue.

The Orange Pi 4 B includes integrated dual-

band Wi-Fi with a UFL connector for an

external antenna, plus true Gigabit Ethernet.

The latter easily matches the network

performance of a Raspberry Pi 4 and the

former exceeds it, although it’s at the odd cost

of a significantly reduced maximum range.

Elsewhere, the board includes a Rockchip

RK3399 system-on-chip, which packs two

high-performance ARM Cortex-A72 cores

and four ARM Cortex-A54 cores into a pair

of clusters, both running at ‘up to’ 2GHz.

There’s an ARM Mali-T864 GPU, 4GB of

dual-channel LPDDR4 memory and 16GB of

eMMC storage, which is preloaded with an

Android Open Source Project (AOSP) build.

As well as the HDMI 2 and DisplayPort 1.2

ports, both supporting 60Hz 4K displays,

there are two LCD connectors and two

MIPI CSI camera connectors. However, one

of the latter pulls double duty, meaning

it’s possible to connect two LCDs and one

camera, or two cameras and one LCD, at

a time. Storage expansion is handled via

micro SD, and there’s a 5V 3A DC barrel-jack

power input.

In general testing,

running Xunlong’s

customised Ubuntu

18.04 Linux distribution,

the Orange Pi 4 B is no

slouch. It exceeds the

Raspberry Pi 4 in some

benchmarks, thanks

to higher memory

bandwidth and those

two extra CPU cores,

and trails behind in

others – likely due to

four of the six cores

being weaker ARM

Cortex-A54 units.

97

Firing up the NPU unveils the Orange Pi 4

B’s true potential. Running a simple image

inference workload, the NPU was able to

accurately categorise images at a rate of

between 15 and 25 frames per second, while

increasing the power draw by just 1W over idle.

Sadly, the software side of the package

lets it all down. To use the NPU with pre-

configured software and pre-installed

demos, you need to use an outdated build of

Ubuntu, which fails to resize its root partition

on boot. A newer, bug-fixed build offers the

3D acceleration missing from the earlier

version, but doesn’t include the software for

the NPU. In both cases, stability was an issue

and crashes occurred frequently.

The other software is simply outdated, rather

than buggy. The bundled AOSP build is based

on Android 8.1 with a November 2018 security

patch level, and there’s no sign from Xunlong to

suggest a newer build is in the works.

At $69.90 US (ex VAT) from orangepi.org,

the Orange Pi 4 B isn’t the cheapest SBC on

the market, but given the high cost of buying

an NPU accelerator separately, it’s priced

pretty aggressively. For anyone investigating

edge and embedded neural network

projects, it could be worth a punt, as long as

you’re willing to spend time polishing the

software side of the setup yourself.

Thermal imaging reveals a fairly hot-running

SoC which fails to bleed its heat to the board

Running an image inference task on the

NPU reveals impressive performance

Notthata lackofnetworkingis anissue.
TheOrangePi4 B includesintegrateddual-
bandWi-Fiwitha UFLconnectorforan
externalantenna,plustrueGigabitEthernet.
Thelattereasilymatchesthenetwork
performanceofa RaspberryPi4 andthe
formerexceedsit, althoughit’sattheoddcost
ofa significantlyreducedmaximumrange.
Elsewhere,theboardincludesa Rockchip
RK3399system-on-chip,whichpackstwo
high-performanceARMCortex-A72cores
andfourARMCortex-A54 coresintoa pair
ofclusters,bothrunningat‘upto’2GHz.

There’sanARMMali-T864GPU,4GBof

dual-channelLPDDR4memoryand16GBof

eMMCstorage,whichispreloadedwithan

AndroidOpenSourceProject(AOSP)build.

AswellastheHDMI2 andDisplayPort1.2

ports,bothsupporting60Hz4Kdisplays,

therearetwoLCDconnectorsandtwo

MIPICSIcameraconnectors.However,one

ofthelatterpullsdoubleduty,meaning

it’spossibletoconnecttwoLCDsandone

camera,ortwocamerasandoneLCD,at

a time.Storageexpansionishandledvia

microSD,andthere’sa 5V3ADCbarrel-jack

powerinput.

Ingeneraltesting,

runningXunlong’s

customisedUbuntu

18.04Linuxdistribution,

theOrangePi4 B isno

slouch.It exceedsthe

RaspberryPi4 insome

benchmarks,thanks

tohighermemory

bandwidthandthose

twoextraCPUcores,

andtrailsbehindin

others– likelydueto

fourofthesixcores

beingweakerARM

Cortex-A54units.

97

FiringuptheNPUunveilstheOrangePi 4

B’struepotential.Runninga simpleimage

inferenceworkload,theNPUwasableto

accuratelycategoriseimagesata rateof

between 15 and 25 framespersecond,while

increasingthepowerdrawbyjust1Woveridle.

Sadly,thesoftwaresideofthepackage

letsit alldown.TousetheNPUwithpre-

configuredsoftwareandpre-installed

demos,youneedtouseanoutdatedbuildof

Ubuntu,whichfailstoresizeitsrootpartition

onboot.A newer,bug-fixedbuildoffersthe

3Daccelerationmissingfromtheearlier

version,butdoesn’tincludethesoftwarefor

theNPU.Inbothcases,stabilitywasanissue

andcrashesoccurredfrequently.

Theothersoftwareis simplyoutdated,rather

thanbuggy.ThebundledAOSPbuildis based

onAndroid8.1witha November 2018 security

patchlevel,andthere’snosignfromXunlongto

suggesta newerbuildis intheworks.

At$69.90US(exVAT)fromorangepi.org,

theOrangePi4 B isn’tthecheapestSBCon

themarket,butgiventhehighcostofbuying

anNPUacceleratorseparately,it’spriced

prettyaggressively.Foranyoneinvestigating

edgeandembeddedneuralnetwork

projects,it couldbewortha punt,aslongas

you’rewillingtospendtimepolishingthe

softwaresideofthesetupyourself.

Thermal imaging reveals a fairly hot-running

SoC which fails to bleed its heat to the board

Running an image inference task on the

NPU reveals impressive performance