Intel disables hot new TSX tech in early Broadwells and Haswells One of Intel's new ways to make software go faster is called Transactional Synchronization Extensions (TSX), an innovation that gives developers fine control over how multi-threaded code uses a CPU's resources. TSX is a reasonably big deal, because as Intel never tires of pointing out it long-ago decided that grunt alone is …
From the sound of it, it's just a BIOS change to switch this feature off at boot time - just like you can switch some other CPU features off through the BIOS. A microcode update which got the feature working properly would be better, of course, but sounds as if Intel can't or won't fix it this way. My guess is that for a feature this performance-critical, the implementation (and hence the flaw) is right in the silicon, too low for microcode to patch.
There was a similar "workaround" for the original FDIV bug as I recall: a little DOS utility which switched the floating point unit of early Pentiums off. In those days, "everyone" was still booting through DOS, and there were enough 486SXs out there that most software could never depend on having floating point hardware, so that was almost workable - except, having coughed up $$$$ for a high-end chip, you were then left with lousy performance on anything maths-heavy!
"except, having coughed up $$$$ for a high-end chip, you were then left with lousy performance on anything maths-heavy!"
Intel did have a free replacement program for the FDIV Pentiums (after people complained loudly) and the replacements cost Intel dearly, hundreds of millions or so.
The Intel CPU erratas are not rare at all, but most seem to be about "CPU locks under some offbeat scenarios" and are in general not a problem. But the FDIV produced unreliable FP calculations and it was really simple to reproduce too. Pentiums didn't support microcode updates so this couldn't have been fixed with a BIOS update anyway. (and 20 years ago getting a BIOS updates was way, way more cumbersome to get and program than today)
TSX was very ambitious, trying to solve something in hardware that is very difficult to do properly in software. So it comes as no shock that there are problems with the version 1.0 of it. It is likely that newer models will have bugs found that may or may not be fixable without simply disabling it. Anyone depending on something like this in its first version in production is foolish to the extreme.
I'm not surprised, nor am I disappointed... Well a litte. ;)
The good thing is that Intel are willing punt it to customers, and smart enough to put their hand up and switch it off. Pain tends to happen along with progress.
The thing is, even if Intel doesn't get TSX out into the wild, I think it's a fair bet some MIPS clone will get something comparable in China (it may already have been tried).
The Register 
Biting the hand that feeds IT
Copyright. All rights reserved © 1998–2024
