110 PART | III ITS business models
• The design of AI computation specific hardware that can further acceler-
ate data processing partially “at the edge” and partially “at the cloud” in a
transparent manner to the end-user. Cloud service providers such as Google
and Amazon hardware divisions are working on AI accelerator chips and
architectures such as the tensor processing unit (TPU). TPU is AI chip that
offers 15–30 times computations than GPU’s using 30–80 times less power.
The transparent processing of data, both at the edge and on the cloud is
expected to explode in the next few years, with investments on micro-chips
to reach 6.5 billion USD by 2021 and the respective investment on machine
learning-based knowledge inference going from zero to 1 billion USD each for
data centers and edge devices per year (Morgan, 2018). The rise is estimated to
be higher for edge devices than for the data centers, which will undertake all
the inference workload leaving training—which is, in essence, the preprocess-
ing—for the cloud backend.
In this same direction, Tesla has recently announced, at the Tesla Autonomy
Day event, its new full self-driving (FSD) chip (Pell, 2019). Manufactured by
Samsung in Austin, TX, the custom chip, says the company, was built with
autonomy and safety in mind and is currently shipping in its new models, in-
cluding S, X, and 3. The 260-mm two chip, features a pair of neural network
processing modules that can handle 36 trillion operations per second (TOPS)
each with a power consumption of 72 W. Two such chips will be installed on
each of the company’s FSD computer boards, delivering 144 TOPS for collect-
ing and processing data from radars, cameras, and ultrasonic sensors, using the
embedded deep neural network architecture. The company is claiming “FSD”
capability at the hardware level, for all the vehicles that are equipped with these
chips. For that, says Tesla founder and CEO Elon Musk, “All you need to do is
improve the software.”
According to the Vice President and General Manager of Automotive Nvidia
Rob Csongor, “It's not useful to compare the performance of Tesla's two-chip
FSD computer against Nvidia’s single-chip [Xavier] driver assistance system”
but the Nvidia DRIVE AGX Pegasus computer outperforms the 144 TOPs of
Tesla’s chip, running at 320 TOPS and offering AI perception, localization, and
path planning. Both companies agree that self-driving cars are the future of the
industry and with the embedded, AI-capable, chips, and algorithms they will be
able to provide safety, convenience, and efficiency at a better quality level, at the
expense of computational power.
9.7 Impact on growth and sustainability by compliance for
intelligent transport decision systems, Standards
The primary goal of ITS in this domain is to achieve cognitive decisions ac-
cording to human patterns of situation awareness, perception, and decision
making, based on machine perception from different kinds of sensors and data
sources and AI-based learning. The processes have to follow human-centered