Page 33 - EE Times Europe Magazine | April2019

EE|Times EUROPE 33

Top 10 Processors for AI Acceleration at the Endpoint

(MMA), which can achieve 8 TOPS. resources to save power and cost.
The SoC can handle a video stream from a front-mounted camera The KL520 is available now and can also be found on an accelerator
at up to 8 MP or a combination of four to six 3-MP cameras plus radar, card from manufacturer AAEON (the M2AI-2280-520).
LiDAR, and ultrasonic sensors. The MMA might be used to perform
sensor fusion on these inputs in an automated valet parking system, Gyrfalcon Lightspeeur 5801
for example. Designed for the consumer electronics market, Gyrfalcon’s Lightspeeur
The TDA4VM is designed for ADAS designs between 5 and 20 W. The 5801 offers 2.8 TOPS at 224-mW power consumption (the equivalent of
device is in pre-production, but development kits are available now. 12.6 TOPS/W) with 4-ms latency. Gyrfalcon uses a processor-in-memory
technique that is particularly power-efficient compared with other archi-
GPU tectures. Power consumption can actually be traded off with clock speed
Nvidia Corp. Jetson Nano by varying the clock speed between 50 and 200 MHz. Lightspeeur 5801
Nvidia’s well-known Jetson Nano is a small but powerful graphics contains 10 MB of memory, so entire models can fit on the chip.
processing unit (GPU) module for AI applications in endpoint devices. The part is the company’s fourth production chip and is already
Built on the same Maxwell architecture as larger members of the Jetson found in LG’s Q70 mid-range smartphone, where it handles inference
family (AGX Xavier for camera effects. A USB thumb drive development kit, the 5801 Plai
and TX2), the GPU Plug, is available now.
on the Nano module
ULTRA-LOW-POWER
Nvidia’s Jetson Eta Compute ECM3532
Nano module Eta Compute’s first production product, the ECM3532, is designed for
houses a powerful AI acceleration in battery-powered or energy-harvesting designs for
GPU with 128 cores IoT. Always-on applications in image processing and sensor fusion can
for AI at the edge. be achieved with a power budget as low as 100 µW.
(Image: Nvidia Corp.) The chip has two cores — an Arm Cortex-M3 microcontroller core
and an NXP CoolFlux DSP. The company uses a proprietary voltage and
has 128 cores and is capable of 0.5 TFLOPS, enough to run multiple frequency scaling technique, which adjusts every clock cycle, to wring
neural networks on several streams of data from high-resolution image every last drop of power out of both cores. Machine-learning workloads
sensors, according to the company. It consumes as little as 5 W when in can be processed by either core (some voice workloads, for example, are
use. The module also features a quad-core Arm Cortex-A57 CPU. better-suited to the DSP).
Like other parts in Nvidia’s range, the Jetson Nano uses CUDA X, Samples of the ECM3532 are available now, and mass production is
Nvidia’s collection of acceleration libraries for neural networks. Inex- expected to start in the second quarter.
pensive Jetson Nano development kits are widely available.
Syntiant Corp. NDP100
CONSUMER CO-PROCESSORS U.S. startup Syntiant’s NDP100
Kneron Inc. KL520 processor is designed for
The first offering from American-Taiwanese startup Kneron is the machine-learning inference on
KL520 neural network processor, designed for image processing and voice commands in applications
facial recognition in applications such as smart homes, security sys- in which power is tight. Its
tems, and mobile devices. It’s optimized to run convolutional neural processor-in-memory–based
networks (CNNs), the type commonly used in image processing today. silicon consumes less than 140 µW
The KL520 can run 0.3 TOPS and consumes 0.5 W (equivalent to of active power and can run models
0.6 TOPS/W), which the company said is sufficient for accurate facial for keyword spotting, wake word
recognition, given that the chip’s MAC efficiency is high (over 90%). detection, speaker identification, or
The chip architecture is reconfigurable and can be tailored to differ- event classification. Syntiant’s NDP100 is de-
ent CNN models. The company’s complementary compiler also uses The product will be used to enable signed for voice processing
compression techniques in order to run bigger models within the chip’s hands-free operation of consumer in ultra-low-power applica-
devices such as earbuds, hearing tions. (Image: Syntiant Corp.)
aids, smartwatches, and remote
controls, according to Syntiant.
Development kits are available now.
GreenWaves Technologies GAP9
GAP9, the first ultra-low-power application processor from French
startup GreenWaves, has a powerful compute cluster of nine RISC-V
cores whose instruction set has been heavily customized to optimize
the power consumed. It features bidirectional multi-channel audio
interfaces and 1.6 MB of internal RAM.
GAP9 can handle neural network workloads for images, sounds, and
vibration sensing in battery-powered IoT devices. GreenWaves’ figures
have GAP9 running MobileNet V1 on 160 × 160 images, with a channel
scaling of 0.25 in just 12 ms and with a power consumption of
Kneron’s KL520 uses a reconfigurable architecture and clever com- 806 µW/frame/second. ■
pression to run image processing in mobile and consumer devices.
(Image: Kneron Inc.) Sally Ward-Foxton is a staff correspondent at AspenCore.

www.eetimes.eu | APRIL 2020