The EconomistDecember 7th 2019 Science & technology 812
1cause foreign aid accounts for two-thirds
of the money spent on malaria.
Another problem is patchy data about
local disease patterns. This makes it tricky
to work out the best mix of malaria-control
measures for a given area—and when to de-
ploy them. Still, it is better to use whatever
figures are available, because that will ini-
tiate a virtuous circle, says Dr Alonso. As
things stand, local health workers respon-
sible for collecting such data often do a
sloppy job because they do not see the data
being put to use.
Such things matter. The two countries
that stand out as successes in this year’s re-
port are India and Uganda. Both report dra-
matic falls in cases of malaria between 2017
and 2018. Not coincidentally, both have
been busy fine-tuning their regional ma-
laria-prevention strategies. If other coun-
tries followed suit, the world might get
back on track to beating the disease. 7Net gainsSource: WHO*Global Technical Strategy for Malaria 2016-30
†At peak rate, 2000-07Malaria, cases as a % of population at risk2000 05 10 15 20 25 30036FORECAST 9
Current estimatesExpected
trendWorst-case
scenario†Global target*S
ilicon chipshave lonely lives. They are
born together, often as tens of thou-
sands of identical siblings a few milli-
metres across, on a single wafer the size of
an old-fashioned vinyl record. They are
then broken from their natal wafers like
squares of chocolate from a bar, and pack-
aged individually in plastic and metal.
Only after this is a chip reconnected to oth-
ers of its kind, as the packages are wired up
to work together on circuit boards and in-
serted into products.
Many inventors over the years have not-
ed that if chips were instead wired together
from the beginning, on the wafer itself,much expense and trouble would be avoid-
ed. But efforts to implement such wafer-
scale integration have consistently failed,
either because the technology just did not
work or the resulting circuits could not
compete with new versions of conven-
tional designs.
Now Cerebras, a firm in Los Altos, Cali-
fornia, thinks the time is right to try again.
The heart of its new product, a supercom-
puter called the cs-1, could hardly be de-
scribed as a “chip”. It is a slab of silicon
measuring 21.5cm by 21.5cm that the firm
refers to as a wafer-scale engine. But what-
ever name you give it, it is a record-breaker.
A high-end modern computer chip might
have billions of transistors on its surface.
The wafer-scale engine has more than a
trillion of them.
Cerebras’s creation breaks many re-
cords besides the trillion-transistor barrier
(it actually has 1.2trn). Its transistors are or-
ganised into 400,000 individual process-
ing units, known in the trade as cores, and
it can shuttle nine petabytes (9,000trn
bytes) of data per second around inside it-
self. For comparison, Intel’s i9-9900k
chips, typical of those found in modern
pcs, have a mere eight cores and can shuttle
40 gigabytes per second.
The cs-1 has some notably small num-
bers, too. Admittedly, ibm’s Summit super-
computer, among the snazziest in the un-
classified world, offers some 2.4m cores.
However, Summit is constructed conven-
tionally, using package-laden circuit
boards. It weighs over 340 tonnes and oc-
cupies 520 square metres of floor space. A
cs-1 weighs around 250kg and is the size of
a domestic refrigerator. It also consumes a
mere 15-20kw of electricity. Summit re-
quires 1,000 times as much.
The purpose of all this computational
heft is to run linear algebra, the mathemat-
ics of data processing in general and mach-
ine learning in particular. Machine learn-
ing is at the heart of the trendy and
lucrative field of computing branded “arti-
ficial intelligence”.
The cs-1’s compiler—the software that
turns programs written by human beings
into binary code which a computer can un-
derstand—is tuned to keep the flow of data
from core to core as efficient as possible. It
does this by matching the structure of the
code generated to that of the hardware.
Also, as the cores are positioned within
fractions of a millimetre of the memory
they use, that flow of data is already much
faster from one part of a circuit board to an-
other than the long-distance trip which
would normally be required.
The wafer-scale engines themselves are
made by tsmc, a Taiwanese firm, using a
process claimed to be so accurate that each
has just 150-200 defects. These are easily
worked around, given the number of other
transistors available. Wafer-scale integra-tion has many other challenges, such as
keeping everything synchronised, pump-
ing in enough electric power, pumping out
the resultant heat, and efficiently moving
gigabytes of data to and from other parts of
a machine. But if the cs-1 survives contact
with the real world of commercial use, then
wafer-scale integration will at last have
proved itself, and the days of the lonely
chip may be numbered. 7How to make a small supercomputer
with a really big chipComputing recordsA trillion here, a
trillion there
C
entipedes do notgenerally get on well
together. Even members of the same
species may attack one another when they
meet. So it is a surprise to find mother cen-
tipedes sharing nests and a double surprise
to find that those co-residents are some-
times not even conspecifics. This, though,
is the conclusion of research published in
Biotropica by Farnon Ellwood and Josie
Phillips of the University of the West of
England, in Bristol.
Dr Ellwood studies the invertebrates of
the Danum Valley, an area of rainforest in
Sabah, a Malaysian state in north Borneo.
His past expeditions have found lots of
centipedes living in epiphytes called bird’s
nest ferns. These ferns tolerate the low illu-
mination beneath a forest’s light-absorb-
ing canopy and may weigh more than
200kg. They and their inhabitants are hard
to investigate because they grow on tree
trunks dozens of metres above the ground.
But when Dr Ellwood did bring a few down
to terra firmahe found that the largest of
them contained, besides the plethora of
herbivorous insects he was expecting, 126
centipedes. That led him to wonder wheth-
er, rather than migrating from the ground
as he had previously assumed was the ori-
gin of such myriapods in tree tops, the
creatures were actually being born there.
To investigate the matter he and Ms
Phillips collaborated with colleagues from
Sabah’s Forestry Department and the Natu-
ral History Museum, in London, to set up
climbing lines in local trees and use them
to collect bird’s nest ferns. Each specimen
was, as it broke loose from the tree, decant-
ed straight into a clear plastic bag to stop its
centipede inhabitants escaping. It was
then lowered to the ground using pulleys.
In total, the researchers nabbed 44 ferns in
this way—half from the highest part of the
canopy, above 40 metres, and the rest from
above 20 metres. Once a fern was safely
landed they dissected it and dropped everyEven the most aggressive animals will
co-operate if they have toCentipedesNesting instinct