8 posts / 0 new
Last post
Homebrew Your Own Supercomputer

I read about GreenArrays INC in the latest issue of QST Today, on page 97 in Vol 96 No 3
The company is offering for sale the GA144 processor with 144 cores! for the princly sum of $20.00
power consumption is speced at less then a watt, and they calim it can do 90 billion instructionss a second.
The minimum order is one 10-pack (ten chips)
~~~(the evaluation board is a bit steep, other then that I'm willing to chip in 1/5 of the funds to buy a pack.)~~~

Looks interesting, we used to

Looks interesting, we used to talk abou banks of 8051's in the old days.  I dont see anything like a library to act as a job/thread controller, I would think that they would have something to bridge an application to multiprocessor execution unless they expect all of the customers to write their own OS's. Maybe an integration between Matlab and codegeneration like we can do between Matlab and DSP haradwrae using Altera. :)

yes looks like an FPGA is called for to play with these toys!

My gut feeling is to slap a low dissipation FPGA on a board with some of these GA144 chips, and use the P2 VME pin out as a starting point.

By keeping the heat dissipation down, the fans can be left unused!!!!!
as to how to make the boards talk to one another, look no further then the 65K+ concepts...

The problem is not inter-CPU

The problem is not inter-CPU communication, but intra-CPU communication and coordinating the mini-programs of the 144 cores. They are using Forth, which is fun to program with, but you need a special mindset to like it and to use it successfully :-) I think they don't have an OS for it, it is intended for things like settop boxes.

I have started looking into

I have started looking into the people behind the GA144 chip, They are from a time where the mainframe ruled the day. that chip is designed to run code, and nothing else.
I got a chance to eyeball code for that thing running Ethernet as an interface. even their early drafts are very elegant.

Yeah the easiest way to

Yeah the easiest way to multiprocess is to divide up the work where each processor can do what it needs to do without interaction with the other "threads".  In the old days we would have killed for this capability for pixel/voxel processing, like modern GPU's do, or for array type processing like DSP's do.  SO these days we have GPU's, DSP's and dedicated encryption silicon so the question remains how to have a bunch of processors work together.
So either there is an underlying OS that tries to virtualize the processes, or you have compile time management or you just plain wing it with semaphores and process forking where a processor is engaged for the fork.
Back in the ole days the tasks were extremely repetative, think of oil companies modeling almost like raytracing discrete points and there was a gain from using vector processing which could be done with multiple processors I assume, however not much call for vector these days.
I used to refer to the Intel x86 architecture as a "shoebox" as the accumulator felt so limited that it was like running around with a shoebox in your hands and you could only pick up or put down one item at a time from your shoebox.as opposed to context via register arrays, etc.  However the Intel stuff getting faster and faster until it became one hell of a shoebox and made up for the inefficiencies and limitations of addressing and contextual modes by shear speed.  (including branch prediction, prefetching, etc)

one of the cool things about

one of the cool things about the GA144, is that when it's not needed a cell or all of them can just halt and resume as needed. a true statemachine.
They also talk about energy conservation, to the point the internals of this chip are so stingy they are measuring instructions in pico watts

Frankly, I'd like to use one

Frankly, I'd like to use one of those (or a couple or so) on my C128 as a "SUPER"-CPU. Real computing power to process information and free the 8502 to do what it does best.

Log in or register to post comments