07/10/2004 09:45:02, "David Bisset" 
<david_bisset@itechnic.co.uk> wrote:

>Well Samsung obviously have 512M parts in BGA, 

I don't have any in my hand, so I don't quite believe that. It's 
plausible, though.

>Given that:
>a) Most of us have lived happily with 64M (even TCL).

Partly because TCL haven't released the monster version yet. 
There is also the issue of of using Balloon as a compile box, to 
stop the need for cross-compilers. Similarly, if Cambridge 
computerlab is thinking of letting compsci undergrads near them, 
more memory might be handy, if the project wants big datasets 
for video or whatever.

>b) No one has owned up to ever having stacked up TSOP's, 
something you
>could only do on a few "one offs" given that it wouldn't reflow.

Certainly. It was never proposed as a mainstream solution. 
Imagine the situation where you didn't want to cross-compile. A 
couple of development boards get the stacked-RAM treatment, 
and production boards get what they need. 

>c) You can only work within the technology available at the time 
of
>design.

Yeah. That's why I'm designing in a CPU I don't have, 'preliminary' 
NAND devices, a power controller that's disturbing me by its lack 
of existence, and we're having this discussion about RAM. About 
the only things I'm sure about are the decouplers and resistors... 

>I'd say the choice was obvious go for 4 BGA's.

Certainly. They're all different sizes, even though they tend to 
have the same pinout  - the Samsung part isn't the same shape as 
either of the Micron offerings. Which one shall I arbitrarily pick, so 
that when the manufacturer changes the die size or shape, we're 
stuffed, since I've had to pack them in so tightly to allow 4 sites.

(Note that the PXA26x only supports 4 banks of up to 64Mbytes 
each. 256Mbytes is tops, and that will take 8 16-bit-wide chips. 
32-bit chips aren't commodity, and in TSOP, they're really 
annoying to solder. Double-die SDRAMs from Micron are tempting 
me again. 2 pairs of 
MT48V32M16S2 2.5V 8 Meg x 16 x 4 banks BOTH DIE
would let us fully populate the whole memory space, or any 
increment up to that. 

TSOP does have the benefit that Balloons could be shipped 
memoryless, and hackers could drop on whatever SDRAM and 
NAND devices they fancied / could source. Add to that TSOP54's  
universal shape, size and availability, and it's not ruled out yet. It's 
certainly the safest option.

>If you really want 1G then design and build an expansion board 
via the
>LoonBus (it needs a name...) OK so it will be slower, but the 
capability
>is there (I assume). 

Tricky. It'd need to present as some kind of swap device - 
PXA26x doesn't seem to want to support vast memory spaces.