Page 42 - EE Times Europe Magazine | April2019

42 EE|Times EUROPE — Boards & Solutions Insert

SPECIAL REPORT: AI AT THE EDGE
Chip Startups for AI in Edge and

Endpoint Applications

By Sally Ward-Foxton

s the industry grapples with the best company at more than US$1 billion, making it
way to accelerate artificial- the first Western semiconductor startup to be
intelligence performance to keep up designated a unicorn.
A with requirements from cutting-edge The company’s Intelligence Processing Unit
neural networks, there are many startup com- (IPU) chip has a massively parallel architec-
panies springing up around the world with new ture with more than 1,200 specialized cores,
ideas about how this is best achieved. They are which can each run six program threads.
attracting a lot of venture capital funding, and There is also substantial on-chip memory
the result is a sector rich in not just cash but in — hundreds of megabytes of RAM — plus,
novel ideas for computing architectures. importantly, 45 terabytes of memory band- The Cerebras chip takes up an entire
width. That allows entire machine-learning wafer. (Image: Cerebras)
A data center can count as models to be stored on the chip. and packs 400,000 cores and 18 GB of memory
Graphcore’s IPU chip is available in a Dell
the edge, depending on where server for edge compute applications. onto 84 processor tiles. If the stats seem
it is. The key concept of edge Groq staggering, remember that one Cerebras chip
is designed to replace thousands of GPUs.
Cerebras says it has solved problems that
computing is that the data is Founded in Silicon Valley by a team from previously plagued wafer-scale designs,
Google, Groq employs 70 people and has
processed in or near the same raised US$67 million in funding to date. It such as yield (it routes around defects), and
geographical location as the officially unveiled its enormous chip, which is that it has invented packaging that count-
capable of 1,000 TOPS (1 peta-OPS), late last
ers thermal effects. The company has raised
data is generated or gathered. year at SC19. more than US$200 million and says its rack
The company’s software-first approach system is running in a handful of customer
means its compiler handles many of the data centers.
Here at EE Times Europe, we are currently control functions that would normally happen
tracking about 60 AI chip startups in the in hardware, such as execution planning. Soft- Cambricon Technologies
United States, Europe, and Asia, ranging from ware orchestrates all the dataflow and timing One of China’s first AI chip companies, but
companies reinventing programmable logic required to make sure calculations happen by no means its last, Cambricon was founded
and multicore designs to developers of entirely without stalls, making latency, performance, in 2016 by two researchers from the Chinese
new architectures and others using futuristic and power consumption entirely predictable Academy of Sciences who are brothers.
technologies such as neuromorphic (brain-in- at compile time. Citing the lack of agility in CPU and
spired) architectures and optical computing. Groq is targeting data-center applications general-purpose GPU (GPGPU) instruction
Here is a snapshot of 10 that we think show and autonomous vehicles with its tensor sets for the acceleration of neural networks,
promise or at least have some interesting streaming processor (TSP) chip. The part is they developed their own instruction set
ideas. We’ve categorized them by where in sampling now on a PCIe board. architecture (ISA), a load-store architecture
the network their products are targeted: data that integrates scalar, vector, matrix, logical,
centers, endpoints, or devices for the artificial Cerebras Systems data transfer, and control instructions.
intelligence of things (AIoT). California-based Cerebras is famous for res- Cambricon’s first product, Cambricon-1A,
urrecting the wafer-scale chip idea, which was is used in tens of millions of smartphones
AI IN THE DATA CENTER abandoned in the 1980s. and other endpoint devices such as drones
Yes, a data center can count as the edge — The company’s mammoth, 46,225-mm die and wearables. Today, second-generation
2
depending on where it is. The key concept of takes up an entire wafer. It consumes 15 kW chips include two parts for the cloud plus an
edge computing is that the data is processed edge compute chip, the Siyuan 220, that was
in (or near) the same geographical location designed to fill a gap in the company’s portfo-
as the data is generated or gathered. This lio. The chip offers 8-TOPS performance and
includes gateway or hub devices but also consumes 10 W.
on-premises servers that accelerate compa- Cambricon (along with Horizon Robotics;
nies’ individual AI applications. Think servers see below) is currently one of the world’s
that accelerate image classification for X-rays most valuable chip startups: The company
or CT scans in a hospital or medical research has raised US$200 million thus far, giving it a
facility, or gateways that receive status data market valuation of about US$2.5 billion.
from the factory floor and process it on-site.
AI IN THE ENDPOINT
Graphcore “Endpoint” refers to devices at the end of the
Based in Bristol, U.K., Graphcore hit the news Graphcore’s IPU chip has more than 1,200 network, wherein the data is processed inside
when an early funding round valued the cores. (Image: Graphcore) the same device that collected the data. This

APRIL 2020 | www.eetimes.eu