EVERY TIME WE write something about Nvidia's woeful reliability record and honest communications, the fanbios come out of the woodwork and ask "what about the bad ATI cards?" We have spent months looking for such beasts, and eventually found out the source of those whispers, but it isn't ATI.
There are bad 3870s, no question there, documents seen by the Inquirer show that there was unquestionably a large number of potentially flaky parts sent out by at least one board maker and sold by one OEM. The question of why and how gets really murky really fast however, and there is no smoking gun.
The companies in question are Diamond, Alienware, GeCube and ATI. The parts are the 3870, first and second revision, the 3870 X2, and potentially some 3850s. What happened is that Alienware got a large batch of Diamond cards, and they started failing at higher than normal rates, between two per cent and, in some cases, slightly above 10 per cent. Quite unacceptable.
There were three causes for the problems: a wrong resistor, a questionable BIOS, and a PSU that wasn't delivering anything close to its rated wattage. Users saw dead cards, cards that would artifact, BSODs, and Crossfire that simply didn't work. None of these things – causes or symptoms – are a good thing.
You may or may not realise it, but Diamond doesn't make cards, they buy cards from large board makers. This is very common, most GPU sellers in fact do not make their own boards, they contract out to large Taiwanese OEMs, or in some cases, rival board makers.
Here is where the finger pointing starts. As far as we can trace it, the boards appear to be made for Diamond by GeCube, nothing unusual there. Other makers have also made Diamond boards in the past as well. These were then sold to Alienware to be put in their machines, and again, business as usual.
Some blame GeCube for sub-standard boards. Others blame Diamond for installing a substandard BIOS. The boards that Alienware qualified were not the same boards as were delivered for production, and those changed in the middle of the run.
Did GeCube change production without telling Diamond? Did Diamond order a change? Did some engineer take it upon him/herself to save few dollars? This part is totally unclear, and will likely never be answered fully, but there were several changes, and they had a detrimental effect.
The cards started dying, both in burn-in/power-on testing, and in user hands. Most runs were around fice per cent bad or less, with a single small run of parts much higher than that. In total, a few thousand cards were directly caught, and sources say the number of cards potentially affected are between 10,000 and 20,000.
Keep in mind that of these, about five per cent are bad, likely less. Again taking the worst case of 20,000 bad parts, and it isn't that high, this means there are about 1,000 bad cards manufactured, and most of those were caught and or reworked before release.
To put it in perspective, there were way way more than a million 38xx parts sold, and some sources put the number much closer to two million than one. Even if there were only one million sold, that would put the bad run of parts at two per cent of the total made, and of that five per cent, or one tenth of a percent failed.
Another thing to realise is that these parts are most definitely not the reference design made by ATI. An ODM or AIB changed something, or somethings, that they should not have, and that lead to problems. The stock parts are fine, and no other vendor that we talked to had heard of the problems, so they look to have been contained to a single run, caught and mostly fixed before many got into the wild.
Diamond isn't shirking responsibility like certain competing GPU vendors out there, and if you have one of these cards, you should contact them so they can make things right. In any case, due to the tiny number of bad parts out there, it is unlikely that you will ever see one, much less have one die on you. A tenth of a percent is a pretty small number.
So, if you hear rumors about bad 3870s or X2s, now you know why. They do exist, and they do fail, artifact, and some won't run 3DMark as is whispered around. Unlike certain other companies, if you are one of the unlucky 1,000 or so, likely far less, you will probably be taken good care of, not stonewalled like the other guys. µ
Upcoming flagships might not switch to USB-C after all
Netflix without the chill
The best things come in the same sized package as last time
'Open source' and 'Microsoft' in same sentence shock