2 Apr
2013
2 Apr
'13
6:15 p.m.
On Apr 2, 2013, at 9:03 PM, Jason McBrien wrote:
On Tue, Apr 2, 2013 at 6:34 PM, Brian Roth
<abacos_98 at yahoo.com> wrote:
They're about as difficult to program as any other supercomputer
processor. The whole point of the original Cray architecture
(and other Cray-designed CDC supercomputers, IIRC) was that there
were I/O processors performing the data transfers to provide the
CPU(s) enough data to chew on at full throttle. Anything less,
and you were wasting processor power. The Cell's ring data
structure wasn't all that different.
Which, I guess, is not to say that they weren't difficult to
program, but that supercomputer programmers probably should have
felt right at home.
- Dave
As time goes on the density of the processors goes up but its not just
the CPU's. Its what goes on behind the processors such as SGI's MIPS
supercomputers of a few years ago. They used a distributed shared memory
that required circuitry that connected processors, local memory, a network
interface, and the
I/O. These systems can be difficult to add onto brick wise and the only
way to increase speed is to add racks which adds power consumption.
The other problem with Roadrunner was it was a hybrid design - AMD Opterons
combined with cell processors, like the one in the Playstation 3. At the
time these were screaming fast, but difficult to program.