FPGA VAX update

Jim Battle frustum at pacbell.net
Sat Nov 5 02:39:41 CST 2005


Sean 'Captain Napalm' Conner wrote:

> It was thus said that the Great Dwight K. Elvey once stated:
> 
>> My thought is to make a variable word length processor using
>>a single bit ALU. Of course, I've been thinking in terms of a relay
>>based machine.
> 
> 
>   The CM-1 (Connection Machine) was just such a beast, although with that
> you had 65,536 1-bit CPUs, each with 4096 bits of memory.
> 
>   -spc (Interesting machine, and certainly on topic ... )

that's baby stuff!  :-)

UNC worked for many years on a chip architecture called pixel planes.  they 
developed a few generations of custom VLSI chips for it.

anyway, everything inside was bit serial and was controlled by SIMD 
instructions.  one chip controlled a 128x128 group of ALUs, and each ALU had 
something like 2 kbits of DRAM.  sure, that is only 16K processors -- but that 
was just one chip.  they had 256 or so of these chips on a board, and had 16 or 
so boards in a rack, and their system had something like 6 racks.  the SIMD'ness 
was only local to a chip -- they didn't have all the ALUs in the system driven 
by one instruction stream!

because of the scale of the thing, they had to worry about clock skew, as all of 
these boards were communicating over a token ring bus (not IBM's token ring -- 
their own).  their solution was interesting.  their backplane boards had a clock 
plane that was driven from one side, and the other side was unterminated.  they 
picked their geometry and frequency such that they would set up a standing wave 
and the antinodes of the wave were coincident with where each board was plugged in.

although it was primarily used for graphics, there were some attempts at using 
it for more general computation.

anyway, I saw it (pixel planes 5) in operation at siggraph in 93 I think.  they 
were giving away buttons saying "I saw it!", referring to the fact it was doing 
2M phong shaded triangles in real time.  It was amazing for the time, but now 
single GPU chips crank out 100x that many triangles that have much more complex 
shading.




More information about the cctalk mailing list