On Wed, 15 Mar 2000, Spike Jones wrote about the FUZION 150:
> Holy schlaMOLY here it comes:
Hmmm... well someone there has trouble counting since one story says
80+ million. But IBM is going for 170 million!
See: http://www.zdnet.com/zdnn/stories/news/0,4586,2466014,00.html
> PixelFusion's FUZION 150 is a 0.25 micron, single-chip,
Gads, its obsolete before it hits the street. SotA is 0.18 microns now
while IBM is going to 0.13 microns Q3 next year. If the fabless
companies get into a position of being 2 generations behind, you
have to wonder if they will be able to compete.
> 24 megabits of on-chip embedded DRAM.
Great! A *real* PIM (processor in memory) chip that you can buy.
> This ultra high-performance chip delivers more than 1.5 trillion
> operations or 3 billion floating-point operations per second,
Well the 1.5 TeraOps are integer operations, but if I can do "synapse"
arithmetic with them, with ~700 of them you get a peta-op which is about
in the middle of estimates for "brain-equivalence". However since it
consumes about 30W, that works out to ~20,000 Watts. I doubt I'm
going to run it at home. Gotta do something about that.
Of course you get interesting architectures, here is my 1 W transmeta
chip doing general purpose computing connected to my 30 W video
chip to display all the fancy graphics...
> along with 600 gigabytes per second of on-chip memory bandwidth.
This is what PIM *really* buys you which is great. It isn't the
arithmetic the brain does that gives it the power, its the
internal communciations bandwidth. So with this much on chip
bandwidth a few thousand of them (with optical interconnects) gets
in the ballpark of the bandwidth in the brain.
Now you did miss the *interesting* thing about the chip -- it comes
in a 1060 pin HPBGA (high performance ball grid array?) package
and measures 4.5 cm square (big!).
So, we assume in ~10 years or less we can knock an order of magnitude
off of the power budget and # of processors. So, looks like by 2010
we will get me a desktop machine with a reasonable power budget and
something close to brain capacity. So we only have 10 years to figure
out how to program the thing.
Robert
This archive was generated by hypermail 2b29 : Thu Jul 27 2000 - 14:05:23 MDT