Maarten (Coppertop) ICQ'de me deze errug interessante usenet thread met een logisch beredeneerd antwoord op de vraag waarom Rambus RIMMetjes zo idioot duur zijn:
Go back up and read that initial quoted paragraph again. Remember that a Rambus die, when active and running at 800MHz, gets *hot* *hot* *hot*. As in, it needs a heatsink the size of the entire RIMM module to cool it. And that die, the raw die, has its electrical contacts on the rear of the die. Itsy bitsy solder balls. MAMA MIA!Not until earlier this evening did I realize that the individual Rambus dice cannot be tested at 800MHz (or whatever full speed) until the RIMM is fabricated and the heat sink/spreader installed. What this means is, you solder the Rambus dice (2-16 of them, but typically 8), which have been tested (by Samsung) only for functionality at a very slow speed, onto the RIMM substrate. Then you install the spreader. *THEN*, and only then, do you test them at full speed. And if they won't run at speed, you throw the entire RIMM away. Just *one* bad (or slow) die means the entire RIMM is worthless (or won't run at full speed). The fabrication process for ball grids is rather delicate and you ain't gonna take a 150-watt American Beauty soldering iron and repair a RIMM.
The problem is the heat. I can't see any way to test the individual dice at full speed before the fabrication of the entire RIMM. Apparently Micron can't either, because Micron refuses to manufacture RIMMs.
Unless this glaring test problem can be overcome, RIMM prices will *never* drop to anywhere near DIMM prices. *Never*! RIMM production has proceeded, at low rates, for over a year now, and no fix is in sight.
Oftewel: de 400MHz clock frequentie van een PC800 RIMMetje genereert dermate veel hitte dat het niet met zekerheid vastgesteld kan worden dat goedgebakken Rambus chips ook samen met z'n achten of zestienen op een PCB'tje functioneren. Als tijdens de testprocedure blijkt dat een RIMM niet stabiel functioneert moet de complete RIMM weggegooid worden en gaan er dus 8 of 16 chips naar de klote. De schrijver van de post berekend middels een ingenieuze formule dat de yields van 8-chip modules bij 91,7% goede Rambus chips al naar 50% kelderen.