ShowDaily Day3 - Flipbook - Page 4
sponsored by
Dickerson
continued from page 3
enable future scaling of power, performance, cost, and density (PPAC).
(see related story)
Yieh said magnetic RAM has
the potential to replace SRAM
and DRAM. While noting that the
multiple layers in MRAM are “not
easy to integrate,” she said MRAM
is non-volatile and has a “rich
roadmap” that could reduce active
power consumption by 10X compared with today’s memory types.
Cliff Young, a Google research
manager, said Google is open to
new memory types, and also sees
long-term promise in analog computation, which could lead to compute-in-memory. Over the past eight
years, since the AlexNET benchmark
came into use among AI developers,
energy efficiency for neural network
training has improved by 50 times,
Young said.
The Google researcher compared today’s AI training methods to the early
18th century steam engines. Those
engines were “clunky” but improved
quickly enough to power the Industrial Age. The AI community is in
a similar mode, looking at a variety
of application-specific processors
and new semiconductor materials to
sharply improve the energy efficiency
of the training function.
Samantha Alt, an Intel researcher,
noted that renewable energies, such as
wind and solar, could be used to operate data centers, bringing the carbon
footprint down sharply, but only if
the data centers could accommodate
the intermittent nature of solar- and
wind-generated energy. Turning data
centers on and off would present
challenges, but Alt said Google and
others already are studying such
“time shifting” approaches to data
center operation.
Alt also said the industry’s efficiency can be enhanced by Open
Source projects. Yieh concurred,
pointing out that PPAC scaling can
only proceed if various parts of the
industry engage in “co-optimization
within the ecosystem.”
Chip Manufacturing
continued from page 1
power the expansion. The wafer fab
equipment segment – which includes
wafer processing, fab facilities,
and mask/reticle equipment – is
expected to rise five percent in 2020
followed by 13 percent growth in
2021 driven by a memory spending
recovery and investments in leading-edge and China. Foundry and
logic spending, accounting for about
half of total wafer fab equipment
sales, will see single-digit increases
in 2020 and 2021. Both DRAM
and NAND spending in 2020 will
surpass 2019 levels and are projected to grow over 20 percent,
respectively, in 2021.
The assembly and packaging
equipment segment is forecast to
grow 10 percent to $3.2 billion
in 2020 and 8 percent to $3.4
billion in 2021 driven by advanced
packaging capacity buildup. The
experts tackled the subject of AI
and energy efficiencies, moderated
by Eric Masonet of the University
of California at Santa Barbara and
organized by Applied Materials.
Bob Aitken, an ARM Ltd. Fellow,
said he believes that energy efficiencies in high-performance computing will come partly by using the
third dimension, stacking memory
and logic to lower the capacitance
(which results in higher power)
and delays. “The 3D aspect has not
begun in earnest yet,” Aitken said,
“but there are several 3D solutions
out there.” In the longer term, logic
and memory transistors could be
fabricated monolithically, stacking
layers of logic and memory.
Elie Yieh, vice president for Advanced Product Technology Development at Applied Materials, pointed
to the just-announced selective
tungsten deposition technology,
which she said would serve to reduce
the resistance at the contacts and
SEMICON West exposition.
Growth across a number of semiconductor segments is expected to
Results are shown in
terms of market size in
billions of U.S. dollars.
Source: SEMI July 2020,
Equipment Market
Data Subscription. New
equipment includes
wafer fab, test, and
assembly and packaging.
Total Equipment
does not include
wafer manufacturing
equipment. Totals may
not add due to rounding.
4 | Wednesday, July 22www.semiconductordigest.com
