🚀 go-pugleaf

In article <1992Dec14.221541.25270@dartvax.dartmouth.edu> pichet@coos.dartmouth.edu (Pichet Chintrakulchai) writes:
>Hello,
>
>I've been working on a project designing a processor with FPGA's.  One severe
>limitation I found is that their cells do not have tri-state outputs and thus
>forcing me to use MUXes on buses consuming a lot of the resources.
>
>Does anybody have any idea why they didn't make these cell outputs tri-state?
>
>Thanks.
>Pichet Chintrakulchai		(pichet@dartmouth.edu)

Maybe you are using the wrong FPGAs :-)

Enter advertising mode:

	One of the facilities of the Xilinx products that differentiates us
	from all the other lowly wanna-bees is that our XC3000 and XC4000
	products are full of tri-state buses. These are driven by what we
	call TBUFs, and there are 2 per logic block (i.e. just the right
	number). For each row of logic blocks, there are two rows of TBUFs
	with their associated horizontal long lines (the buses). On most of
	the larger chips, these long lines are split in the middle of the
	chip by splitters. For a given chip with N rows of logic blocks,
	you can have anywhere from N*2 bus lines (that span the width of
	the chip) to N*4 bus lines, in two groups of N*2, one for the left
	side of the chip, and one for the right. You can mix and match
	between half and full width within the one chip to get any number
	between N*2 and N*4.

	A sample of sizes follows:

	Part Numbers	Full width buses	Half width buses
	
	XC4005		28			56
	XC4006		32			64
	XC4008		36			72
	XC4010		40			80
	XC4013		48			96

	XC3090		40			80
	XC3190		40			80
	XC3195		44			88

Exit advertising mode.

With regard to building a processor, I have built (complete with front
panel with flashing lights and toggle switches) a 20MIPS 16 bit RISC
(native MIPS, not corelatable to anything else) with a XC4005 for the 
processor (used about 75% of the chip, the rest is available for 
peripheral intergration) and 6 XC3020 for the front pannel.

If you want to talk to me about your project, (408)879-5180, or
philip@xilinx.com



In article <1992Dec15.193837.8890@dartvax.dartmouth.edu> eht@northstar.dartmouth.edu (Edward H. Truex) writes:
>In article <1992Dec15.010020.9274@super.org> cfreese@super.org (Craig F. Reese) writes:
>>
>>You don't say which FPGA you are using but the Xilinx 4000 series does
>>contain tristate buffers for just the kind of purpose you describe.
>>The buffers are not on each CLB but adjacent.  There are some
>>limitations as to where the buffers can connect up.  I have found
>>designs where the functionallity did _not_ work for a given device (due
>>to the size of the FPGA, size of my design, and the tristate buffer/bus
>>resources) but a larger chip probably would have.  In many other
>>designs I have not used any tristate logic (so I'm glad the CLBs aren't
>>all tristate by default).
>
>
>I am working with Pichet (the originator of this thread) on a very 
>similar design, and I think his question can be rephrased.  What we
>are really trying to figure out is 
>
>Is there an electrical reason why the short pieces of interconnect 
>(those that are connected through Xilinx's 'PIP's) cannot be tristated?
>
>The long lines obviuosly don't present any electrical problems as they 
>are simply edge to edge pieces of wire.  Are the pass transistors in 
>the interconnect points the problem? 
>
>Here's the motivation for our inquiries...
>
>With the current Xilinx parts (I am using XC4010's) the number of
>busses that can be implemented is severely limited by the number
>of long lines.  In a design that requires many wide busses (I am
>currently working on a 32 bit microprocessor) the number of long
>lines available is not enough to implement a true bus structure.
>As a result, I am forced to use a large number of fairly wide
>multiplexers, and a correspondingly large number of CLBs and routing 
>resources.  The ability to tristate the output of each CLB while
>attaching it to a piece of "normal" routing would eliminate (or
>reduce) bus constraints on chip, in addition to (possibly) freeing 
>up some of the CLBs that were previously used for bus redirection
>to be used for "useful" logic.
>
>Thanks for any insights,
>Tad Truex 


OK!!!!!!!!!! so you are using the RIGHT FPGAs!! :-)

The reason that the local interconnect (short pieces of metal) dont have
tristate capability in current families (we may in a future family) is
for silicon real estate and speed reasons:

(for those of you just reading along, PIPs are Programable Interconect
Points, that link our chip routing resources together, either with
buffers, pass transistors, or other techniques)

1) tristate-able pips would be much bigger, needing a control line, an
input selector mux (for the tristate control line itself) and default
logic for the non-tristate case. versus 1 pass transistor.

2) pips between local to long lines include a buffer, so bidirectionality
would not work very well

3) tristateable pips are slower than direct pips  

4) we like to guarantee valid signal levels on all interconnect, so
we would have to include pull up resistors and 'weak keepers" everywhere
as well. this would grow the chip size a lot.

5) assorted other reasons.


-- 
Philip Freidin: Product Planning Manager, Xilinx, INC
(rest of clever .sig still under construction....
		coming to a terminal near you, Real Soon Now (tm))