dataflow (was: RTX-2000 processor PC/AT add-in card (any takers?))

Jecel Assumpcao Jr. jecel at merlintec.com
Tue Apr 11 20:24:07 CDT 2017


Chuck Guzis via cctalk wrote on Tue, 11 Apr 2017 18:05:01 -0700
> On 04/11/2017 04:53 PM, Jecel Assumpcao Jr. via cctalk wrote:
> 
> > I consider the heart of any modern high performance CPU to be a
> > dataflow architecture (described as an "out of order execution
> > engine") with a hardware to translate the macrocode (CISC or RISC) to
> > the dataflow graph and tokens on the fly.
> I wouldn't characterize an out-of-order execution scheduler as
> "dataflow", at least not in the traditional sense.

I have never seen anybody else, including people whose research in the
late 1980s was dataflow architectures, do so either. But I see an engine
with 24 "in flight" instructions plus all the register renaming circuits
and it sure looks the same to me.

> Certainly, nobody that I was aware of ever categorized, say, a CDC 6600
> as a dataflow machine.

I was not aware that there had been any out of order implementations
after the IBM ACS until the second half of the 1990s. Given Cray's
passion for simplicity, I would not expect any of his designs to use
o-o-o (specially one as early as the CDC 6600).

> At least not in the same sense that I'd categorize a NEC uPD7281 as a
> dataflow device.

That is the one I am most familiar with, along with the Manchester
Dataflow Machine and the MIT Tagged Token machine. An interesting modern
dataflow architecture is the TRIPS:

https://en.wikipedia.org/wiki/TRIPS_architecture

-- Jecel


More information about the cctech mailing list