HP Innovation Journal Issue 12: Summer 2019 | Page 66
generates data or a compute resource no more than one
network hop from that device. Edge data must be analyzed
to be useful in driving improvements in how the devices
operate and deliver value to users. 1
In this article, we’ll explore three areas in which HP’s
innovation is focused on driving compute efficiency at the
edge: ultra-efficient processors, software 2.0, and virtual
machines/virtual factories.
ULTRA-EFFICIENT PROCESSORS
The rise of data-intensive workloads at the edge drives the
need for new energy-efficient compute architectures. A new
class of compute processors called Machine Learning Accel-
erators (MLAs) promises great advances in energy-efficient
performance in coming decades.
The data behind these workloads are largely machine-
generated, sensory and perceptual—from cameras,
microphones, and sensors proliferating the devices, such
as voice assistants, autonomous vehicles, and HP Web
Presses. The data can be processed with machine learning,
and thus can be accelerated by MLAs.
Recently, the rate of CPU power-performance efficiency
gains has been slowing. 2 For over 30 years, CPUs
exponentially increased in performance as each semicon-
ductor generation brought roughly two times the number
of transistors in design of new chips (Moore’s Law). It is
clear that we are getting closer to the limits of physics. 3
High-performance Graphics Processing Unit (GPU) archi-
tectures, which have thousands of processing cores, are
staying slightly ahead of the Moore’s Law slowdown pres-
ently. However, because of the data explosion at the edge
and increasing compute complexity, compute demand is
outpacing the power-performance growth of even a GPU.
IT industry analysts forecast that by 2025, the amount of
data against which Artificial Intelligence (AI)/Machine
Learning (ML) analysis will be conducted will increase by
100x, in turn driving compute requirements. Additionally,
compute complexity, by our estimate, will also grow by
a factor greater than 10x during this period, pushing the
compute requirements even higher. 4 It is not enough to
increase compute capacity to meet the demand–it must be
accomplished without increasing power. Hence the need for
Ultra-Efficient computing.
This steep challenge will not be
met just by packing on more
64
HP Innovation Journal Issue 12
transistors, but also through
architecture innovations that
can extract much higher energy
efficiencies with available transis-
tors—in other words, processors
that are ultra-efficient for the
emerging compute workloads.
A NEW WAY TO THINK ABOUT SOFTWARE
Hand in hand with the rise of new data-rich workloads
and new hardware architectures, a new paradigm in soft-
ware development has emerged: Software 2.0 (SW 2.0).
SW 2.0 enables faster development and greater flexibility
for solving traditional problems and tackling complex new
ones. This can lead to faster product-to-market, as well as
reduced development and maintenance costs.
In traditional or classic software development (SW
1.0), the goal is to write lines of code that are treated as
instructions to the computing environment. A typical
project for a software team may have many millions of
lines of code, some of which will come from outside devel-
opment teams, and some from the team itself.
SW 2.0 is a new way of thinking in which value creation
comes not from code writing, but from data curation. We
collect data from our devices, curate that data (select rele-
vant data sets, verify and label them), and use them to train
ML models. We then distribute the models and “run” them,
instead of creating algorithms and writing code.
Consider the data involved in running a large printing press,
and let’s imagine how data curation would be applied in the
digital and physical worlds. Curating data from a printing
press might include engineers looking at images of final press
output and marking some as defects (e.g., lines, splotches,
roller marks, etc.), then training an ML model to detect
defects. The model “runs” in real time at the press to mon-
itor the output. No engineer writes code to analyze images,
removing reliance on hand-built fragile algorithms. Develop-
ing algorithms and dealing with errors by hand is expensive.
Selecting data and then letting the ML training systems
run without regard to any specific algorithm is fast, from
a few hours to days for typical training. Really “big” prob-
lems like self-driving cars may require days of compute on
big cloud-based server farms. In such a complex case, the
equivalent algorithms are beyond our grasp. With SW 2.0,