Agarwal on Multi-Core computer chips
Wielding numerous analogies for his audience of MIT students, Anant Agarwal makes the
case that the next generation of computers, not to mention much of the technology in
everyday life, will be built with smaller, simpler parts “combined in a clever way.”
Agarwal starts with Puerto Rico’s enormous Arecibo radio telescope, 400 meters in diameter,
tuned to detect extraterrestrial life. Rather than being carved from a single gigantic material,
the dish consists of “a whole bunch of tiles” adjusted to create a spherical surface. In the
same way, CPU designers no longer make “one big honking processor,” says Agarwal, but
lots of little processing elements called tiles or cores. This engineering movement, which
MIT helped spark in the 1990s, has brought about multicore processors on chips, which
overcome not just the number-crunching limitations of single processors, but their power
drain as well.
Agarwal uses the example of eating ice cream: You really enjoy the first few spoonfuls, but by
the 30th or 40th taste, “you’re tapped out.” By illustrating the marginal value of eating one
more spoonful, Agarwal tries to get at the idea that once you’ve got a big processor, “making
it bigger doesn’t give you much return.” In fact, as he shows with some math and graphs,
having two or more processors works much better, including burning less power. He applies
Moore’s law and predicts that beyond the four or more cores on chips we now have (he’s
already developed a 64-core chip), we’ll be seeing 1000 tiles per chip in the next five years or
so -- assuming we can overcome three big “P” challenges. There’s the performance hurdle of
getting all those multicore chips to talk to each other and to the outside world without the
gridlock found on a busy city street; power efficiency, which will require rethinking CPU
architecture; and a very big programming obstacle, which may involve deploying an optical
broadcast medium. Crack these, and “multicore could replace all hardware in the future,”
claims Agarwal.
Principle of marginal utility.
I call it "reinventing the human brain from the computer science perspective".
The power consumed by a chip is proportional to the square of the voltage. It is also
proportional to the frequency at which we drive it. Therefore power is proportional to the
voltage cubed.
If we compare a single core processor to a dual core processor, we get the following
calculations:
Cores Freq Voltage Perf Power Power Efficiency
Single core 1 1 1 1 1 1
Bigger single core 1x 1.5x 1.5x 1.5x 3.3x 0.45x
Dual core 2x 0.75x 0.75x 1.5x 0.8x 1.88x
Therefore if we create multicore processors, we can run them at lower frequencies and
voltages and get more processing power per watt.
"We know that we are going toward hundreds of thousands of processors on a chip." Anant
Agarwal, MIT Professor.
http://mitworld.mit.edu/video/671
Wielding numerous analogies for his audience of MIT students, Anant Agarwal makes the
case that the next generation of computers, not to mention much of the technology in
everyday life, will be built with smaller, simpler parts “combined in a clever way.”
Agarwal starts with Puerto Rico’s enormous Arecibo radio telescope, 400 meters in diameter,
tuned to detect extraterrestrial life. Rather than being carved from a single gigantic material,
the dish consists of “a whole bunch of tiles” adjusted to create a spherical surface. In the
same way, CPU designers no longer make “one big honking processor,” says Agarwal, but
lots of little processing elements called tiles or cores. This engineering movement, which
MIT helped spark in the 1990s, has brought about multicore processors on chips, which
overcome not just the number-crunching limitations of single processors, but their power
drain as well.
Agarwal uses the example of eating ice cream: You really enjoy the first few spoonfuls, but by
the 30th or 40th taste, “you’re tapped out.” By illustrating the marginal value of eating one
more spoonful, Agarwal tries to get at the idea that once you’ve got a big processor, “making
it bigger doesn’t give you much return.” In fact, as he shows with some math and graphs,
having two or more processors works much better, including burning less power. He applies
Moore’s law and predicts that beyond the four or more cores on chips we now have (he’s
already developed a 64-core chip), we’ll be seeing 1000 tiles per chip in the next five years or
so -- assuming we can overcome three big “P” challenges. There’s the performance hurdle of
getting all those multicore chips to talk to each other and to the outside world without the
gridlock found on a busy city street; power efficiency, which will require rethinking CPU
architecture; and a very big programming obstacle, which may involve deploying an optical
broadcast medium. Crack these, and “multicore could replace all hardware in the future,”
claims Agarwal.
Principle of marginal utility.
I call it "reinventing the human brain from the computer science perspective".
The power consumed by a chip is proportional to the square of the voltage. It is also
proportional to the frequency at which we drive it. Therefore power is proportional to the
voltage cubed.
If we compare a single core processor to a dual core processor, we get the following
calculations:
Cores Freq Voltage Perf Power Power Efficiency
Single core 1 1 1 1 1 1
Bigger single core 1x 1.5x 1.5x 1.5x 3.3x 0.45x
Dual core 2x 0.75x 0.75x 1.5x 0.8x 1.88x
Therefore if we create multicore processors, we can run them at lower frequencies and
voltages and get more processing power per watt.
"We know that we are going toward hundreds of thousands of processors on a chip." Anant
Agarwal, MIT Professor.
http://mitworld.mit.edu/video/671
 | |  | |  | |  |