Newsgroups: comp.sys.transputer,comp.arch.fpga
Subject: Re: Emulating a transputer on FPGA
Date: Tue, 10 Aug 1999 10:19:54 -0700

Ram Meenakshisundaram wrote in message <37B04CFB.DA77A8F0@olf.com>...
>What would it take to emulate say a T225 or T425 transputer on a FPGA
> (without any external links).  Since I am new at this, what would be an
>ideal FPGA to do this.  Can I accomplish this on a XC4003E??  How many
>gates would I need on the FPGA to do this.  Thanks.

I have thought on this before.  (If you configure a big array of FPGAs as
CPUs, you might as well have them talk to each other.)

The answer: it depends.  By "emulate" do you mean drop-in-replace?  If so, I
can't say, because I never studied the transputer instruction set
architecture, and because it depends upon whether you are willing to handle
divide, floating point, etc., in software.

If you mean "build a new machine in the transputer *style*", with high
integration of on-chip components and instruction set support for message
sends and fast task switching, etc., then here's a few back-of-the-envelope
estimates for you.

I recall the inmos T414, 1985, which had a 32-bit processor, 2KB of on-chip
RAM, 4 10? Mb/s links, and an integrated DRAM controller.  A T414-20 had a
20 MHz clock and did about 10 million transputer instructions per second.
(I think.  I am not a transputer expert.)

A pipelined 16-bit datapath for a two-cycled 32-bit RISC requires about
8x9=72 CLBs, and its control unit could be ~50 CLBs.  A full 32-bit datapath
adds another 72 CLBs.

Consider the 2 KB of on-chip RAM.  In an XC4000, each CLB can store 32 bits.
2KB * 8 b/B * 1 CLB/32 bits = 512 CLBs, or about half of an XC4025E.  Or
just 4 of the 512 B embedded RAM blocks on a Virtex part.  I assume you were
hoping to target an XC4000 so let's reduce our on-chip RAM requirements to
128 B (32 CLBs) and move on.

An EDO DRAM controller w/ page mode support requires a 12-bit register and
comparator, a 12-bit mux, and some state machine logic.  Call it 20 CLBs.

The four serial links (10 MHz is slow and easy) are each ~12-16 CLBs,
although if you time multiplex them you may be able to make four links from
one implementation plus a register file.

Totals:

CLBs    What
72-144  CPU datapath
50      CPU control
32      128 B on-chip RAM
20      DRAM controller
20-64   4 serial links
----
194-310 CLBs

An expert might fit it in a 196 CLB XC4005E/XL, more likely you would
require 1) an XC4008E or XC4010XL and 2) hand-mapping and floorplanning
experience.

(The XC4003E, with 10x10 CLBs, will probably prove too small.  You could
build a minimalist 16-bit wide datapath in about 4x9 CLBs and could build a
CPU there, but you probably won't have room for links or memory controller.)

If you are more interested in exactly implementing the transputer
instruction set architecture, and IIRC it is a stack machine, then consider
the example of the MSL16 microprocessor.  See Leong, P.H.W, P.K. Tsang, and
T.K. Lee, "A FPGA based Forth microprocessor", pp. 254-255, in Proc. IEEE
Symp. on FPGAs for Custom Computing Machines 1998, and at
http://www.cse.cuhk.edu.hk/~phwl/msl16/msl16.html, and Paul Lee Wai Lun's
thesis and presentation, "FPGA Implementation of a Forth Processor", at
http://home.hkstar.com/~wail/project-7260/project.htm.

To implement a legacy ISA in an FPGA, I advise building a simplified
implementation hidden behind a binary rewriting system!

Jan Gray

Copyright © 2000, Gray Research LLC. All rights reserved.
Last updated: Feb 03 2001