ralphw@ius3.ius.cs.cmu.edu (Ralph Hyre) (06/11/88)
Anyone have any thought or experiences on using transputers in Sun
workstations. I recall seeing some literature on a Sun 9U board-level
product with around 8 T800's on it at one time, but can't recall the company.
VME 6U boards might be preferable.
The main competition around here is the WARP systolic array processor,
200 Mflops/box. The main drawback is the lack of an easy-to-use software
development environment.
The support burden of another widly different architecture would need
to be small to even be considered, especially given the technology
window between now and the next generation of Warp (late '89), which will
have twice the performance and C compilers and such.
The Quality of C and Lisp systems needs to be comparable with the SunOS
products as well, and I suspect this might ultimately be the sticking point.
Some around here are doing neural net stuff, but mainly it's just
integer image->image transformations, and random bursts of floating point
(usually at such inopportune times as to make FP subsystems whose performance
is proportional to pipeline utilitization not as useful)
Send email, and I'll summarize to the net.
Thanks.
--
- Ralph W. Hyre, Jr.
Internet: ralphw@ius2.cs.cmu.edu Phone:(412)268-{2847,3275} CMU-{BUGS,DARK}
Amateur Packet Radio: N3FGW@W2XO, or c/o W3VC, CMU Radio Club, Pittsburgh, PAjap@maths.bath.ac.UK (06/19/88)
I am using a box made by a company called Meiko (who hail from Bristol in the UK). It was started by a bunch of people from the original transputer design team. The organisation is that you get an interface card which slots into the Sun card cage and then run a link cable into the Meiko box (called the Computing Surface). They make a number of boards with different configurations depending on what you want to do. We currently have a localhost board (interface at the other end), an early version masstor board (intended to be used for disk interface (SCSI)) with a T4 and 8Mb and a compute board with 4 * T8, each with 4Mb. There is also a graphics board with 24bit (each) RGB output. In the works, I hear, are a compute board with 8Mb per processor and an ether interface board. The hardware seems quite well made. The software is reliable but occasionally shows limitations. Personally, I just don't like the TDS interface. Debugging support is abysmal in this environment too - if it does not work it just goes dead. Nevertheless, we have managed to get KCL working quite happily (single processor) in about one and half man-weeks and T4 performance is bit slower than a Sun 3/160 and T8 is about 20% faster than a 3/160. The benchmark we used was OPS5 running some consistency tests and an extended version of the (canonical) monkey and bananas problem. If we can find the time we will probably put the Reduce algebra system up too. The Sun hosted C compiler (from Renishaw Controls) is not very good, and I would expect better performance from almost any other C compiler. A promising candidate is the Norcroft C compiler, which produces very high quality code on other machines and which Perihelion (see below) are using, but that is ANSI only. We have had various versions of the Helios (Perihelion) file server working between the Sun and the Meiko and this promises to offer a much more satisfactory working environment - the touch and feel, to you and to your programs, are Un*x. The other project here is the porting of Portable Standard Lisp to the transputer to support existing work on concurrent processing for a project (called Concurrent Processing for Advanced Simulation) done in conjunction with the Rand Corporation. --Julian.