[ WWW.TECHLIFE.NET ] [ 077 ]
crunch data for high-resolution gaming at
incredible speeds. But through Nvidia’s CUDA
application programming interface (API) code,
software developers can tap those GPUs for
everything from mining Bitcoins to calculating
climate models (GPUs have also been known to
crack a password or two).
However, Microsoft ended up settling on
FPGA technology and has created what it calls
an ‘AI supercomputer’, linking its worldwide
data centres, with FPGA chips. That tech
now powers Bing searches and Microsoft’s
Azure cloud computing services
(tinyurl.com/goo4pm4).
TENSOR PROCESSING UNITS
Meanwhile, Google might be better known
for its Android operating system, but it, too, is
dabbling in accelerator chips, creating a custom
and still-mostly-secret Tensor Processing Unit
(TPU). A ‘tensor’ is a mathematical object that
can be described by ‘arrays’ or groups of
numbers. A row or column of a spreadsheet
can be a one-dimension array representing
a first-order tensor, a single-page spreadsheet
with rows and columns can represent a
two-dimensional array or second-order tensor.
Google isn’t handing out too much information
about how its TPU works just yet, other than it
being an ASIC rather than FPGA design. That
said, it has revealed that the custom ASIC
delivers roughly a ten-times or ‘order of
magnitude’ improvement in machine-learning
performance per watt compared with
traditional options, which is a massive gain
(tinyurl.com/henes56).THE CPU MARKET IS CHANGING
But this growing market demand for accelerator
chips is the sort of thing you’d expect to make
CPU makers nervous. More accelerator chips
will likely mean buying less CPUs, so if you’re a
CPU maker in this situation, you start thinking
about combining the two. As it turns out, chip
manufacturers have been dabbling in mixing
CPUs with FPGAs into a single chip since at
least 2010 — that year, Intel embedded an
Altera FPGA core into an Atom E600C chip,while FPGA maker Xilinx put a CPU from
phone chip designer ARM into its FPGA chips.
But with the FPGA market heating up, Intel
decided to put down a hefty US$16.7 billion at
the end of 2015 and buy FPGA maker Altera.
And it didn’t muck around — a few months
later it revealed a Broadwell-class Xeon server
processor with Altera’s Arria 10 FPGA in the
same chip package.THE FUTURE OF THE CPU
As cool as FPGA-powered consoles are, you
don’t spend nearly US$17 billion to make chips
for retro-gaming. The real goal for Google’s
TPU, as it is with Project Catapult and Intel’s
Altera purchase, is machine learning and data
centres. It’s a commonly held belief that we’re
producing data at an exponential rate —
according to IBM, 90% of the world’s data
was generated in just the last two years (tinyurl.
com/hrpkxsw), data that’s stored in data
warehouses and increasingly needs machine
learning to search and manage. Google says its
squadron of TPUs can find every piece of text
in its global Street View database in under five
days and just one TPU can process over 100
million photos each day in Google Photos.
Meanwhile, Microsoft says its ability to write
algorithms direct to its FPGA network can save
two years waiting for hardware upgrades. Now,
the hardware can be upgraded in code (tinyurl.
com/goo4pm4).
The rise of the accelerators will inevitably
lead to the question of whether or not the end
is in sight for the CPU. As more and more of
our data — and our data processing — ends up
in the cloud powered by all manner of ASICs
and FPGAs, there will likely be a corresponding
drop in the needs for CPUs, at least CPUs based
on currently available technology. With
continued research, particularly in Australia,
into quantum computing engines that promise
processing on a revolutionary scale, some might
believe current CPU tech is riding its last great
wave. But the reality is that traditional CPUs
will still be around for a long time to come, at
least until something else comes along that can
match them on a price-performance scale.
Now, where’s that NES console gone?FPGA-powered Analogue NT
Mini delivers full-compatibility
NES gaming.Microsoft is speeding up
Bing searches using
FPGA technology.Microsoft launched
Project Catapult in 2011 to
investigate post-CPU tech.FEATURERISE OF THE ACCELERATORS