Newsgroups: comp.arch,comp.sys.transputer
From: DELETEMEas@comlab.ox.ac.uk (Andrew Stevens)
Subject: Re: PPRO + VIA NIC == T9000 ?
Organization: Oxford University Computing Laboratory
Date: Tue, 27 Jan 1998 10:36:35 GMT
Mime-Version: 1.0
Content-Type: text/plain; charset=us-ascii
Content-Transfer-Encoding: 7bit
Message-ID: <34cdb675.691509197@news.ox.ac.uk>

On Thu, 22 Jan 1998 13:07:18 -0700, "Robert A. Wipfel"
<rawipfel@novell.com> wrote:

>Suppose Intel casts a VIA NIC into silicon,
>then suppose they glue this device close to a PPRO die,
>via an AGP look-alike, perhaps called the RIOP (remote I/O
>port)??
>
>What do you get?
>
>Answer: something that looks like a T9000 transputer -
>	processor plus reliable low latency embedded comms
>	controller designed for clustering (er, parallel
>	processing).
>
>Next, a few new instructions. Hey, if there's space for
>MMX micro-code, why not build an alternate variant - a PPRO
>with extra communication instructions, ones that interact
>with VIA NIC hardware; queues etc. Finally, all you need are
>a few more instructions to support the O/S' integration
>of VIA communication events with its scheduler.

Sort-of except you'll be missing one key bit of the
whole Transputer idea - very low latency context switching /
hardware scheduling.   High-speed Inter-CPU comms alone aren't
the whole story: you need  low latency and fast context switching to
keep the CPU busy doing useful work.

IMHO: the T9000 project was rather misguided.
The whole point of the Transputer architecture was
to support parallel/sequential coding as a "native"
style so as to allow scaling by adding more connected CPU's.

The T9000 (a big expensive uniprocessor) was not really a sensible
way to scale that.    What price links etc etc if the CPU is so big
and fast its too expensivefor it to be sensible  to have more than a
few.

A much smaller low-power low-cost CPU core with the emphasis on
improving link numbers, functionality, and bandwidth and modest
up-clocking would have been much more in keeping with the concept.
Stuff to support multiple cores on die / communcation with
co-processors would have been the way to get the big-FPU/DSP bang
the T9000 was intended to deliver.

Andrew