Don't want to ruin the party but, how much improvement can be done in just a few months?
Depends entirely on what's causing the bug. If it's just a matter of rerouting a signal that's right on the limits for valid timings (so sometimes it works, sometimes it doesn't) then it could easily be fixed in a couple of months; one chip I worked on had serious problems which turned out to be due to the routing software having a brain-fart and putting in a wire that was almost half an inch long when it should have been a couple of millimeters... fixing the chip was trivial once we realised that.
Even if it's a logic issue, that can often be fixed without respinning the silicon (which takes much longer than metal layer changes).
Using virtualization, something becoming increasingly commonplace in Data centers, makes the TLB bug much more likely to happen.
And yes, there was a correction to the article:
Quote :
(Correction: AMD is shipping samples of the Barcelona processor to channel and distribution partners not production versions of the chip.)
Well that and the current workarounds are hurting performance, so fixing the TLB Errata should increase performance as well!
I fancy a change from my 6000+ to a nice speedy Phenom!
Well that and the current workarounds are hurting performance, so fixing the TLB Errata should increase performance as well!
I fancy a change from my 6000+ to a nice speedy Phenom!
The article is about the Barcelona, which is a K10 Opteron. Currently there are no work arounds for defective K10 Opterons, but rather AMD sells the defective chips to customers that aren't affected by the bug.
However, once AMD gets around to fixing Agena (Phenom, K10 quad-core), then we will see a performance boost in comparison to patched systems. However, keep in mind many of the benchmarks used today are UNPATCHED Phenom systems. So in comparison to those systems, simply fixing the bug won't increase performance.
Message edited by TechnologyCoordinator on 03-02-2008 at 06:46:34 PM
Depends entirely on what's causing the bug. If it's just a matter of rerouting a signal that's right on the limits for valid timings (so sometimes it works, sometimes it doesn't) then it could easily be fixed in a couple of months; one chip I worked on had serious problems which turned out to be due to the routing software having a brain-fart and putting in a wire that was almost half an inch long when it should have been a couple of millimeters... fixing the chip was trivial once we realised that.
Even if it's a logic issue, that can often be fixed without respinning the silicon (which takes much longer than metal layer changes).
I was talking purely about performance improvements, sorry if I haven't mentioned that.
What TC said. Most benches we see are without the patch for the TLB bug since most every desktop user will not see it so even with the B3 fixing the TLB errata, there is no guarantee(and more than likely not) for a performance boost.
Although it should help sell more Phenoms to those who think the bug was a show stopper.
