What's your bus? Summary

["Alan Deikman" <Alan.Deikman@znyx.com>]

Well I got some interesting responses, which I will summarize.

Kevin from Phoenix Technologies wrote:

>  Nothing to stop PCI Express from moving all the way
> down.  Makes a lot of sense in the desktop and mobile
> market.

It does seem that Intel has thrown their lot in with PCI Express "over"
Infiniband.  They terminated their chip support to cede the market to
Mellanox, without whom IB would be DOA it seems.

David O'Shea from Intel wrote:

> PCI Express, because my employer says so...
> (Goes fast, is PCI space compatible, thus fast S/W
> adoption, and has a decent speed improvement roadmap).
> [I could talk about its bad features, but then I'd get
> fired.  Lets just say its not perfect.   :-)  ]

Wouldn't engineering be so much more straightforward without those pesky
employers?  Just because they have money they think they get to decide
everything.  Sheesh.

Then Peter Marek from MarekMicro suggested:

> I think Hypertransport will have its intermediate
> market. There's working silicon, even from vendors other than
> AMD and the PC chipset companies (e.g. PMC Sierra).
> But in the long run, PCi express will make it.

I have noticed Hypertransport on some chips in addition to plain-old-PCI.
Anyone care to comment on what the software impact of HT is on the OS and
applications?

Martin Olsson Flextronics Design commented:

> Out of technical merits, I am an Infiniband believer.
> Even though Intel partially seem to have killed the motherboard
> chipset incarnations and replaced it with other stuff, the
> layering and abstractions supported by Infiniband, together
> with the potential of replacing a broad variety of other buses
> for cable/backplane/switch based networks such as Fiberchannel(RAID);
> Quadrics/Myrinet(IPC/MPI);Ethernet(TCP/IP);etc. makes Infiniband a
> very strong candidate for communication in for example backplanes
> for Inter-Process-Communcation aggregated with other services.

I have heard IB has had problems with latency and overhead.  I don't work
with IB much, but is there any basis for this?  It does appear that IB will
have it's "day in court" finally with the advent of PICMG 3.2 aka
Infiniband-over-AdvancedTCA.  If it really does have merits over the
competitors and the software hurdles can be crossed, this should make it
ascendent.

Paul Walker had the presence of mind to point out:

> A common feature of the ones you mention is that they
> are not buses, they use switch fabric.

Most true.  I think it is commonly held now that the next generation busses
are really point to point links all connected with switch fabrics of some
kind.  This is due to the rapid price/performance improvments of SerDes
technology as well as the incredibly low cost of state machines that make
them usable at a system level.  "Busses" in the sense of a lot of parallel
traces on a board seem to be passe.

Then:

>  So does SpaceWire, a derivative from IEEE 1355, that is
> simpler and more flexible than any of those you mention.

Egad, I hadn't heared of this yet!  Time to hit the web.

Now bertg@visioncal.com:

> Near term, Starfabric - it has the features of PCI-EX AS,
> but is here now.

StarFabric is interesting, and in a large part because of the people behind
it.  A lot of them are from the Digital Semiconductor group that brought us
PCI-PCI bridge technology.  The other promise it has is a low software
impact, although if you have a bunch of processors all with access to the
same PCI device (say a video controller) how do the disparate operating
systems deal with that?  StarFabric also looks to be prominent in PICMG.

I heard that StarFabric is merging with another standard.  Which one was
it?

Last but not least Danny Lobo:

> 3GIO in my opinion would be the ultimate bus.

Dem's fightin words.  (How could I have forgot about 3GIO?)  Anybody care
to post an argument supporting this?

That's it so far!

Alan Deikman
ZNYX Networks, Inc.