GCC-3.4.2 PCH failures (segmentation faults)

Declan Moriarty declan.moriarty at gmail.com
Fri Nov 12 02:20:34 PST 2004


On Thu, 11 Nov 2004 18:37:11 +0000 (GMT), Ken Moffat
<ken at kenmoffat.uklinux.net> w
> > >
> > I never knew Amigas had got up to 933Mhz. I presume you have a generous
> > helping of heatsink paste in there? That looks like the old sytyle 486
> > heatsink. I would like to see a heatsink with double the area of a cpu
> > top.
> >
>  This is new (ok, bit over a year old) kit, but when all is said and
> done it's pretty much an evaluation board designed for the 750FX (G3)
> and 7450 (G4) - the 7450 was eol'd before I got my board so I'm on a
> 7455 which is probably hotter (7450 datasheets were also eol'd).  My new
> heatsink is 60x60x(35 or 40, can't remember) and yes, I put some silver
> paste on it.  The cpu itself is the same sort of size as a PIII, and
> anything bigger than 60x60mm will get in the ways of what's on the
> daughterboard (dip switch for clock multiplier, contact pins for who
> knows what, fixing bolts).
> 
> > You can actually work this sort of thing out mathematically. Here are
> > the basic sums:
> >
> > Volyage applied x current consumed = watts. Reach for the cpu datasheet.
> >
> 
>  Surprisingly, the max dissipation on a 7455 is (from memory) 45 watts
> which has got to be into PIII territory.
> 
> > The second one is more complicated: For a central junction, 150 C is
> > failure point and about 115 C (or mebbe lower) is 'instability' at which
> > point builds will fail. Every thermal resistance is provided in ohms.
> > Add the following resistances
> >
> > junctions - case; case - heatsink.
> >
>  The public documentation doesn't go into that level of detail, but the
> cpu idles if it reaches 65 degrees (7455s don't have the TAU register to
> read the temperature, it was discontinued because it was unreliable).
> 
Bah, there's got to be a datasheet for the hardware guys. Give me the
full manufacturer and part number (I've never peeked inside an Amiga),
and I'll have a look in my own sources.
We have ways & means.... and friends :-D

Anyhow, the thing then seems like 


>  The big question is practical - how many volts does it need to run
> reliably at the rated speed ?  And that also depends on what workload
> you feed it - toolchain compiles and (perhaps) doing network backups
> with ssh seem to be the main hot tasks.  Before this week mine was
> running at 800MHz and 1.59V with the old cooler.  Some people recommend
> 1.79V for 933MHz, mine locked up with that.  Seems as if 1.64V might be
> enough.

So you are hopping the voltage around, have upped the frequency by
50%, and are looking at /dev when the thing starts acting funny?? Look
in a mirror :-)

Can your ram do 933Mhz??

The formula for power consumption in a capacitative circuit is 0.5 x
Voltage x Frequency squared! All fets are capacitors from a switching
viewpoint. So there's a lot of capacitance switched there.

I would take this approach. Fix the voltage at 1.59V. It worked there.
Now increase the frequency until it stops working. Then back off a
little.
> 
> > These may be split even further in the datasheet. Your heatsink will be
> > specified in degrees C per watt. So if your cpu uses 10 watts, and you
> > have a 5 degrees C per watt heatsink, your core will be 50 degrees C
> > above ambient. Get the idea?
> >
>  Actually, I did know that, but my heatsink comes from a pile of things
> saved from discarded/broken kit.  A specification was absent.  There is
> also the question of airflow from the cooling fan - these days, most cpu
> coolers don't seem to come with degrees-per-watt specifications (at
> least, none that I've seen in the last four years had useful information
> like that, mostly you're lucky if you get a specification for the fan's
> airflow).
> 
> > Don't forget ram in all of this. That could be the issue here too.
> >
> 
> <fx> sobs </fx> What wouldn't I give for a reliable ppc version of
> memtest86. ;-)  Thanks for your comments, at the moment it looks like
> I'm nearly ready for another go with gcc-3.4 built according to
> -testing, in which case (assuming it doesn't overheat again) I might
> actually get enough results to get back on track for this thread!

Why not try porting it? You throw out all the code for the A20
nonsense in PCs and have a go. You're up to that, aren't you?
-- 
     All the Best

     Declan.



More information about the blfs-support mailing list