It is becoming clearer and clearer that we aren’t the only ones calling on AMD to do something about the Barcelona issue. AMD has too much invested in the system to be acting the way it is, damaging its relationship the way it is. Some of the others are simply suggesting a fessing up to the situation, as it appears that spokes people are trying to spin something hard.
[**update: **the register has a take]
No. We don’t want spin. We want chips, and a patch. Thats it.
Well, a few people want AMD to stand up and say “we have a stop ship going on”. I don’t want to hear that, I want to hear “we have resumed shipping to the channels, en-masse, and advise early buyers to get the chip and the patched kernels. Here is a list of companies that will support the patch (the day job will).
I am a little disappointed, as we wanted to show off JackRabbit with these chips. Right now it is a little hard to do, as I can’t easily get the chips.
That said, JackRabbit turns out to be an awesome virtualization engine as well as being an extremely fast storage unit. Given that the bug manifests itself under virtualization workloads (and I have heard, in lmbench workloads), this could be a “blessing” that we don’t have them … but heck, we can put the system to other uses, show off other things, and use the patch, which works around the problem.
Again, there is very little down-side to us for this. There is very little downside to AMD for this.
AMD needs to remember the Intel FDIV bug fiasco. Some really bright person at Intel at first denied their was a problem, then they denied that the problem would impact people, then they issued a patch, and then they issued parts replacements. But the patch solved the issue (well at the source code level … it required adding a tiny quantity to a divisor), and it slowed down divisions for a while. That bug is still famous, as an example of what not to do when you have a bug.
AMD seems hell bent of creating its own version of FDIV here. The bug exists. First step is to acknowledge it. Second is to figure out the workaround for it. Third is to come up with a remediation plan for those for whom the workaround doesn’t fix the problem. Fourth is to continue to ship units with the fixes until the new units arrive without the bug. In the last week, they seem to have sort of/kind of done the first step. They have the second step. Nothing on third or fourth.
As John West put it on HPCwire,
I couldn’t have said it better.