efficient in other ways too, as Davies explains, “Conventional
computer architectures are great at invoking a lot of data
parallelism to solve lots of problems at once very fast. But as
you batch—like in traditional deep learning applications, for
example—the latency for getting a solution increases. You
have to wait until you accumulate all these data samples to
process them. What neuromorphic architectures are great for
is taking a single query, computing what needs to be
calculated, and producing an answer very quickly.”
BRAIN POWER
Those 770 Loihi processors give about 100 million neurones,
roughly the same as in the
brain of a mole-rat. The
human brain averages around
86 billion neurones, so there’s
a way to go yet before parity is
achieved. It’s more complex
than just loading more Loihi
boards into the case, however,
“Going beyond Pohoiki
Springs does raise some interesting issues,” says Davies. “If
you look in the brain, 80 percent of the volume is the white
matter—long distance communication. In a sense, Pohiki
Springs is at the level where it could still be called grey matter.
There’s a lot of local high-dimensional connectivity, but we can
get by with a plainer, two-dimensional fabric of connectivity in
Pohoiki Springs. To go beyond, we’ll probably have to start
investing more time and research into the right way to
introduce long-distance signalling in a brain-inspired way.”
Loihi doesn’t use long-distance signalling at the moment
because, despite being made of traditional silicon, it doesn’t
separate processing and memory the way our PCs do, “The
data flowing through the neuromorphic system doesn’t move
very far compared to what you would get in a conventional
architecture. You don’t have to bridge multiple caches, or even
beyond DRAM into flash memory. Everything is integrated
into one distributed fabric of compute and memory.”
Davies’ next step is in identifying what neuromorphic
computing is good at. Intel, in association with Western
Sydney University and the 2019 Telluride Neuromorphic
Cognition Engineering Workshop, has already produced a
game of table soccer that plays itself using input from
cameras, and a more serious application comes in the form of
the AMPRO3 prosthetic leg, from Caltech’s AMBER Lab and
Canada’s National Research
Council. The leg can track
objects and adapt to
unforeseen circumstances,
with processing handled by
Loihi chips. But this sort of
thing, while interesting, isn’t
where Davies’ heart lies,
“What we’re really interested
in is finding a new, relatively general-purpose architecture for
computing,” he says. “We’re not there yet.
“The most value will come from difficult computational
problems that correspond to problems our brains are solving
all the time. Our brains are constantly optimizing—if you think
about planning paths, or thinking about the trajectory of how
we might move our limbs, or planning our future, these are
cognitive, abstract actions and decisions. These can be
reduced to daily, well-defined computational problems. So we
think there is a class of problems that will run very efficiently
at scale in a system like Pohoiki Springs.”
Ian EvendenFAR LEFT: Ta b l e
soccer, or ‘foosball’,
that plays itself,
courtesy of Loihi and
Australian engineers.LEFT: “Energy
efficient three-
dimension
multi-contact
prosthetic walking” is
the aim of Caltech’s
prosthetic leg,
powered by Loihi.BIGGEST BRAINS And the heads that house them
3
BOTTLENOSE DOLPHIN
With a brain weighing up to 1.7kg,
dolphins demonstrate their intelligence
by having fun times in the water. Aww.2
ELEPHANT
Elephant brains weigh just over
5kg. They can also live up to 70 years,
and recognize themselves in mirrors.4
HUMAN
These apes with anxiety can only
muster a pathetic 1.5kg brain, although
the weight depends on body size.1
SPERM WHALE
The brain of a sperm whale weighs
about 8kg. These ocean giants can live
for 70 years and reach 16m in length.“THE INSPIRATION GOES
BACK TO THE EARLIEST
DAYS OF COMPUTING”Image (^) c
red
it: (^) I
nte
l
Ima
ge (^) c
red
it: (^) I
nte
l
Tech Report
HARDWARE