Dark Silicon sounds like it should be the title of a
best-selling thriller novel, and in a way, it is a thriller when it comes to
the future of semiconductors. Will advanced nodes produce a huge mass of
transistors that will go "dark" because we can't afford to power them? Will all
our dreams about a more intelligent Internet come to a crashing halt because of
this unforeseen problem?
These provocative questions were raised by Mike Muller, ARM CTO, at the March 25 EE Times "Designing with ARM" virtual conference.
Co-sponsored by Cadence, the one-day event also included panels, chats, and
virtual exhibit booths.
Redefining the mobile
Internet
Muller's keynote didn't go to the "dark side" right away. He
first painted an optimistic picture of what the mobile Internet could be like
in another ten years. Muller envisioned a time in which clients will have
power, complexity, and performance that's equivalent to high-end servers today.
These clients could be embedded in any device - digital picture frames, smart
energy meters, printers, you name it. "The Internet will no longer be seen as
servers," he said. "It will become the software platform that defines what will
be run on the client."
Cloud computing will define the software platform, Muller
said. Software that is "programmed from the cloud, and communicated from the
cloud, is going to transform what the pervasive Internet is all about." As a
result, he said, a lot of today's "static" consumer products will gain Internet
capability in way that will transform the functionality they offer.
The dark lining
behind the clouds
But there's "trouble in paradise up in the clouds," Muller
said. And that trouble is what he calls "dark silicon."
It's a matter of simple math. From today's 45 nm node to the
11 nm technology of 2020, Muller said, we should be able to achieve a scaling
factor of 16. Frequency won't grow as much as it has in the past, but an 11 nm
design should run at 2.4X the speed of an equivalent 45 nm design. The power
consumed per transistor will fall to perhaps 60 percent.
So, given the same power budget used for the 45 nm design
today, if an 11 nm design has 16X the transistors running at 0.6X the power, "I
can actually use only 10 percent of them in my new design," Muller said. "The
rest is dark silicon. We need to find ways of lighting that silicon up."
Turning on the lights
So how to light things up? Muller started with three
suggestions:
- Push
forward on new silicon technologies such as silicon-on-insulator (SOI),
finding the optimal tradeoff between performance gains and power
reduction. (Coincidentally, on March 23, IBM, ARM and Cadence announced
a "Ready for SOI" program to develop an IP ecosystem for SOI).
- Use
energy-efficient, high-density memories to fill some of the "dark" space.
Use power gating to turn them off when not needed.
- Combine
the best process technologies to fulfill various functions with 3D ICs,
which will "become a critical part of how we deliver power-efficient
solutions."
Muller also identified several areas in which further
research and development is needed. These include stream programming for GPUs,
programming solutions for manycore SoCs, design of energy-efficient on-chip
interconnect, and "near and subthreshold circuits" that can use energy
harvesting from heat and light already present in the environment.
Finally, Muller talked about Razor,
a dynamic voltage scaling technique developed by ARM that can dynamically
detect and correct timing errors. Razor tunes the supply voltage by monitoring
the error rate during circuit operation, supposedly eliminating the need for
voltage margins. Muller said Razor makes it possible to build a fault-tolerant
processor that can recover from "fast moving and transient" timing errors.
It's the system,
stupid
Muller concluded by noting something he learned in college -
"it's the system, stupid." If you really want to save power, he said, you first
have to think about where all the power goes. Muller noted that 300 million
motors worldwide will represent 7 percent of global carbon emissions by 2020.
Smart control can reduce motor power consumption by 25 to 40 percent.
"The important problem that faces all of us," Muller said,
"is to optimize the real system, which is the world in which we live."
Conference co-chair John Donovan
summarizes other parts of the virtual conference in a blog at the Low Power Design web site.
Richard Goering