Page 43 - EE Times Europe Magazine | April2019

EE|Times EUROPE — Boards & Solutions Insert 43

Chip Startups for AI in Edge and Endpoint Applications

might include everything from security cam- Mythic Journey 2 offers 4 TOPS at 2 W for
eras to consumer electronics and appliances. Mythic was founded in 2012 at the Univer- L3/L4 autonomous driving and, combined
Of course, there is some gray area, given that sity of Michigan. The company, now based with Horizon’s own algorithms for percep-
some devices can be used as either gateways in Austin, Texas, has raised US$86 million to tion, achieves 90% core utilization.
or endpoints (consider autonomous vehicles develop its analog compute chip, which uses
and smartphones). a processor-in-memory technique based on AI IN THE IOT (TINYML)
flash transistors for power, performance, In this category, we consider chips with
Hailo Technologies and cost advantages over CPUs and GPUs. microcontroller levels of compute that oper-
Founded in 2017 in Tel-Aviv, Israel, by former Processor-in-memory is not new, but ate in resource-constrained environments
members of the Israel Defense Forces’ elite Mythic says it has figured out the tricky at ultra-low power. AI inference on the end-
intelligence unit, Hailo has about 60 employ- compensation and calibration techniques point device in these circumstances is very
ees and has raised US$21 million to date. that cancel out noise and allow reliable attractive because it increases latency, saves
8-bit computation. The company plans to bandwidth, helps privacy, and saves power
sell standalone chips as well as multichip associated with RF transmission of data to
processing cards. the cloud.
Because the device can handle image
processing on HD video at 30 frames/sec- GreenWaves
ond, one of Mythic’s key target markets is A spin-out from the University of Bologna,
security cameras and on-premises aggrega- GreenWaves (Grenoble, France) uses multiple
tors for security camera systems. RISC-V cores in an ultra-low-power ML
Mythic CEO Mike Henry said in an inter- application processor for battery-powered
view at CES that its chip would be sampling sensing devices. The company relies on its
“shortly.” custom instruction set extensions to facil-
itate DSP operations and AI acceleration at
Horizon Robotics minimal power consumption.
Hailo’s AI co-processor, the Hailo-8, can Founded in 2015 in Beijing, Horizon Robot- GreenWaves’ second-generation product,
handle 26 TOPS with a power efficiency of ics had raised about US$600 million by the GAP9, uses 10 cores. One is used as a fabric
2.8 TOPS/W. (Image: Hailo) end of 2019, for a valuation of US$3 billion. controller, and nine make up the compute
Today, the company has more than 500 cluster, with the controller and cluster in
Hailo’s AI co-processor, the Hailo-8, can employees and holds more than 600 patents. separate voltage and frequency domains, so
handle 26 TOPS with a power efficiency of Horizon Robotics’ Brain Processing they each consume power only when nec-
2.8 TOPS/W. It targets advanced driver- Unit (BPU) was originally designed for essary. The GAP9 also takes advantage of a
assistance systems (ADAS) and autonomous computer-vision applications. The second state-of-the-art fully depleted silicon-on-in-
driving applications. Its architecture mixes generation of the BPU is a 64-bit multi-in- sulator (FD-SOI) process technology to
memory, control, and compute blocks, and struction, multi-data (MIMD) core that can minimize power consumption.
adjacent blocks are allocated to compute each handle all types of neural networks (not just GreenWaves’ figures have GAP9 running
layer of a neural network by software. Mini- convolutional networks). It uses the compa- MobileNet V1 on 160 × 160 images with a
mizing data transmission on and off the chip ny’s sparse neural networks to predict the channel scaling of 0.25 in just 12 ms, with
helps save power. movement of objects and for scene parsing. power consumption of 806 μW/frame/second.
Mass production of the Hailo-8 is due to A third generation will add acceleration for GAP9 samples are set to arrive in this
begin in the first half. decision-making algorithms and for other year’s first half.
parts of AI outside of deep learning.
Kneron Horizon Robotics has two chip product Eta Compute
With 150 employees in San Diego and Taiwan lines: Journey, for automotive, and Sunrise, Eta Compute’s design for AI processing in
and US$73 million in funding, Kneron was for AIoT. The first generation of Journey and ultra-low-power IoT devices uses two cores:
one of the first startups to get silicon on Sunrise chips launched in December 2017, an Arm Cortex-M3 microcontroller core plus
the market, achieving the goal in May 2019. with a second generation based on BPU 2.0 a DSP. Both leverage the company’s clever
The company has several customers already arriving in autumn 2019. dynamic voltage and frequency-scaling
announced for its first-generation KL520 techniques to run at the lowest possible
chip and says it made “millions of dollars” in power levels, which is achieved without
revenue in 2019. phase-locked loops (PLLs). The AI workload,
The KL520 is optimized for convolutional allocated via software, can run on either core
neural networks (CNNs) and can run 0.3 TOPS or both. Using this technique, always-on
at 0.5 W (equivalent to 0.6 TOPS/W). That is image processing and sensor fusion can be
enough for facial recognition in IP security achieved with a power budget of 100 μW.
cameras, for example, but the chip can also be The company also optimizes neural net-
found in smart door locks and doorbells. works for ultra-low-power applications that
The company started out making neural will run on its ECM3532 device.
networks for facial recognition and now offers Founded in 2015, Eta Compute employs
those alongside intellectual property for its 35 people in the U.S. and India and has raised
neural processing unit (NPU). A second-gen- US$19 million in funding to date. ■
eration chip, due this summer, will be able to
accelerate both CNNs and recurrent neural Kneron’s KL520 will be succeeded by the Sally Ward-Foxton is a staff correspondent
networks (RNNs), the company said. KL720 this summer. (Image: Kneron) at AspenCore.

www.eetimes.eu | APRIL 2020