Re: NetBSD/sparc64 and GbE performance

[Dave McGuire <mcguire%neurotica.com@localhost>]

On 5/8/11 6:17 AM, Matthias Scheler wrote:
> > > >       Look at those pictures a bit closer.  The Cassini is a 64-bit card
> > > > and the Intel is only a 32-bit card.
> > >
> > > So? The Cassini is still bulky even compared to 64bit Intel card.
> > > And it consumes a lot of power and produces a lot of heat.
> >
> >    It is physically larger and therefore technologically inferior?
> >
> >    WTF?
>
> Yes, larger chips means more transistor, inferior manufacturing or both.

  No.  It's the chip *packaging* that you're seeing, and chip package 
size has little to do with die size.  Package size is primarily 
determined by the mounting style (BGA, QFP, etc) and the connection 
requirements of the die (number of pins)...not the size of the die.  In 
this case, the Cassini has greater pin requirements right off the bat 
with its wider data bus and the external PHY.

  About that external PHY, Cassini's (and many others) use of this 
setup is for flexibility, not due to "inferiority".  There are different 
PHY chips available for copper and a few different flavors of optical, 
with no need to have a completely different die, packaging, and PCB 
design. (I don't recall whether Cassini falls into this group or not, 
but most such boards have all possible PHY configurations designed into 
the PCB layout, and just populate the components for the specific 
configuration being built)

> And it is also the number of extra chips they need. The Intel card
> contains the chip, the PHY and a little eeprom.

  ...and the Intel card is a wimpy desktop-grade 32-bit card that is 
known to be a mediocre performer in anything but a light-duty desktop 
environment.

  By your logic, my Timex/Sinclair 1000 (only four chips!) must be the 
most amazing computer around.  Perhaps my employer should replace their 
IBM z890 with it (2100lb, thousands of chips), and save all that space 
and power.

> >    I'm running a few of these in higher-end (V480-range) Sun machines
> > running Solaris.  While I've not benchmarked them, I know I get better
> > than 500Mbps out of them on those machines.
>
> I never said they couldn't do 500Mb/s. But I'm reasonably certain that
> an Intel card (later updates of Solaris 10 support those for SPARC
> as well) would perform better. A friend of mine used an i82541 based
> card in a SB1000 running Solaris 10 and it worked and performed
> very well.

  A newer server-grade Intel card certainly might, yes, I've found them 
to perform quite well.  However, they are much larger and contain more 
chips.  There are several available now that are considerably later 
designs than the Cassini.

  But not the wimply little 32-bit desktop-grade board that you've 
asserted is a superior performer simply because it's smaller and 
contains fewer components.

             -Dave

--
Dave McGuire
Port Charlotte, FL