BoldportsignsVellemanandPimoronilicences
WhenSaarDrimerannouncedtheBoldportClub
wasshuttingitsdoors,it wasdisappointing.Now,
though,it hasopenedupnewavenuesthanksto
licensingagreementsforthecompany’siconic
circuitdesigns.Drimerhasconfirmedlicensing
agreementswithVellemanandPimoroni,giving
thetwocompaniestherightstomanufacture
andsellBoldport-designedsolderingkits,
includingtheLigemdioLEDtester,new
CordwoodPuzzlevariantsandthebarebones
CuttleArduino-compatible.N EWSI NBRI EF
There’s nothing about offloading traditional
deep-learning workloads, such as TensorFlow
or Caffe, although Seeed indicates that its
engineers are looking into adding it.
If it arrives, however, it’s going to hit a
bottleneck. Using the GPIO header as its sole
connection to the Raspberry Pi sorely limits
the amount of data that could be shuffled
back and forth between the two otherwise
independent machines. The Studio Grove AI
HAT is by no means a cheap drop-in
replacement for an accelerator such as the
Google Coral USB Accelerator, which uses a
USB 3 link to provide a high-speed connection
to its well-supported Tensor Processing Unit
(TPU) co-processor.
Seeed positions the part as suitable for edge
AI projects, smart building monitoring, robotics
and even medical equipment. For the hobbyist,
the board’s low price – $28.90 US (around
£24 ex VAT and shipping) – could make it a
tempting toy even with its various limitations.There’s also a connector for a camera and an
LCD, both sold separately.
The hardware, then, is impressive. When
it comes to actually using the Grove AI
HAT, sadly, it starts to fall down. The first,
and biggest, issue is that there’s no way to
actually program the Grove AI HAT from
the Raspberry Pi itself. Seeed’s demo
program – a face-recognition application
based on a pre-trained data set, and which
requires the ‘optional’ camera and display
to be fitted to the Grove AI HAT with no
option to use an existing Raspberry Pi
Camera Module – does demonstrate an
interface between the Raspberry Pi and
the HAT. However, its purpose is simply
to count the number of faces detected.
The actual programming has to happen
on a traditional x86 desktop or laptop.
For a device that claims to be a HAT for the
Raspberry Pi, that’s a major oversight – but it’s
not wholly Seeed’s fault. The company is
leaning heavily on a toolchain developed by
Kendryte, the company behind the K210 SoC
powering the HAT, and at present, Kendryte
hasn’t released a toolchain that’s compatible
with ARM host systems.
The documentation is also poor. Seeed has
only released three example programs, and
only the facial recognition demo uses the AI
functionality or a Raspberry Pi; the remaining
two demonstrate simple Grove interfacing
with external hardware using the Arduino IDE.
The HAT, as its name implies, leans
heavily on Seeed’s Grove connectors
for external hardware interfacingUsed as a standalone development board,
the Grove AI HAT is surprisingly compact
The Sipeed Maix M1 module is built around a
Kendryte K210 dual-core RISC-V processor
with AI-accelerating ‘KPU’There’s even the potential to simply use it as
an unusual Arduino-compatible, or to play
around with the RISC-V ISA at a lower level.
For beginners, though, the documentation is
simply too sparse and the example projects
too limited. Coupled these shortcomings with
its incompatibility with the Raspberry Pi as a
programming environment – which will only
change if Kendryte decides to put in the time
and effort to make its toolchain compatible –
and the Grove AI HAT is a sadly mislabelled
and mistargeted device.
For professional developers, the Sipeed
M1 AI Module itself may be of more interest:
the company has aggressively priced it at
$9 US in individual unit quantities, dropping
to $6 in volume (around £7.50 and £5 exc
VAT and shipping).
More information on the Seeed Studio Grove
AI HAT is available from seeedstudio.com.
More information on the Maix M1 AI Module,
meanwhile, can be found on sipeed.com