TSX borked in Haswell and Broadwell

[SOFTengCOMPelec]

Yeah, looks like my 4770K wasn't such a bad buy after all, lol.

Unless all 4770 (non-K) owners, are given a free voucher to get an additional Skylake processor (I7), for free.
(A kind of "sorry we messed up", from Intel present).

 

[Compman55]

Was the E3-1245V3 supposed to have it to begin with? I cannot find data on it having it.

 

[SOFTengCOMPelec]

Was the E3-1245V3 supposed to have it to begin with? I cannot find data on it having it.

Yes. (Under the "Advanced Technologies" tab).

Intel® TSX-NI

Source

 

[Phynaz]

Searching round the web about this, some people reckon that TSX is important to the soon to be released Haswel-E/Haswell-EPs, because with lots of cores, and a server which has database software (or whatever they are doing with it) updated to use TSX, they will be rather disappointed. The Wiki about the TSX instruction reckons that TSX makes a huge speed difference, with lots of cores and big databases.
I still have not been able to confirm one way or other, if Haswel-E/Haswell-EPs are affected, but they could well be. But not definitely, because they have an improved or at least bigger cache (I think), which might mean the bug is not present, as TSX has a lot to do with caches, I believe.

Source


So it sounds like a potentially big deal for servers, depending on how much TSX was going to be implemented, on servers.

E/EP is affected. Word is it will be fixed for EX.

 

[krumme]

Unless all 4770 (non-K) owners, are given a free voucher to get an additional Skylake processor (I7), for free.
(A kind of "sorry we messed up", from Intel present).

I actually changed from k to non k because of tsx. Lol - but it sounded good. Ended giving it to one of the kids because it was the same as the one it should replace. Broadwell is the same. This skylake...lol. Man we - desktop - will get it 2 years after the tablets and still quad core - facelifted for us oldies. Unless some miracle happens and we get some competition its just pain. Reminds me of buying a 386 or pentium. Rip offs have returned.

 

[SOFTengCOMPelec]

E/EP is affected. Word is it will be fixed for EX.

That might force them (Intel) to announce that anyone who buys a bugged version of E/EP, can get a free replacement (if they want), when the (bug fixed) stepping, become available.

Otherwise some buyers would simply be put off by this, and delay or not buy their E/EPs.

I actually changed from k to non k because of tsx. Lol - but it sounded good. Ended giving it to one of the kids because it was the same as the one it should replace. Broadwell is the same. This skylake...lol. Man we - desktop - will get it 2 years after the tablets and still quad core - facelifted for us oldies. Unless some miracle happens and we get some competition its just pain. Reminds me of buying a 386 or pentium. Rip offs have returned.

I sometimes like to keep PCs going for many, many years, by which time TSX may be of ever increasing importance. So not having it now, may not matter that much, but a few years down the line, it may be a problem. I hope they offer a free replacement programme for this (for existing owners of affected parts, such as the I7-4770), because if they do, it will make people feel better about Intel, and stop/reduce a possible reduction in sales, in the short term.

 

[WhoBeDaPlaya]

Yay! Now I feel better about my 4770K (compared to the 4790K)

 

[ashetos]

Would this situation teach intel to not differentiate the freaking ISA on their SKUs?

Does one need to be very intelligent to realize that one ISA for all products is the way to go?

Intel put themselves into this mess.

 

[ShintaiDK]

That might force them (Intel) to announce that anyone who buys a bugged version of E/EP, can get a free replacement (if they want), when the (bug fixed) stepping, become available.

Otherwise some buyers would simply be put off by this, and delay or not buy their E/EPs.

There wont be any fixed Haswell or Haswell-E/EP. Broadwell seems to be the first fix with the F stepping. For desktop users the fix is called Skylake.

Buyers wont be put off. They will sell the same amounts, TSX or not.

 

[ShintaiDK]

Would this situation teach intel to not differentiate the freaking ISA on their SKUs?

Does one need to be very intelligent to realize that one ISA for all products is the way to go?

Intel put themselves into this mess.

I doubt that would change anything. People experimenting with TSX isnt exactly buying those SKUs without.

But else I agree, specially in the case of AVX2. Thats just plain dumb.

 

[SOFTengCOMPelec]

Buyers wont be put off. They will sell the same amounts, TSX or not.

I was seriously considering the possibility of getting Haswell E/EP, but probably was not going to get it, as I was going to go for some type of broadwell/skylake or something, later instead.

But now I (almost) definitely WON'T get haswell-E/EP, as I am put off as a buyer. Because I avoid faulty/bugged stuff, unless the problem is VERY trivial.

I bet others will/would be put off as well. Haswell-E/EP are VERY expensive processors, to buy faulty/bugged ones, and the huge number of cores they can have, make TSX even more important (I think).

 

[ShintaiDK]

I was seriously considering the possibility of getting Haswell E/EP, but probably was not going to get it, as I was going to go for some type of broadwell/skylake or something, later instead.

But now I (almost) definitely WON'T get haswell-E/EP, as I am put off as a buyer. Because I avoid faulty/bugged stuff, unless the problem is VERY trivial.

I get others will/would be put off as well. Haswell-E/EP are VERY expensive processors, to buy faulty/bugged ones, and the huge number of cores they can have, make TSX even more important (I think).

What TSX software will you run that can singlehandle define your purchase? Specially compared to the cons of having to wait a year longer.

Also why would you buy a Haswell-E/EP in the first place? Specially the EP? Everyone buying servers will buy Haswell-EP like nothing was changed and they all cheer on the reduced TCO. Just like they did with every model before and before that and so on.

 

[SOFTengCOMPelec]

What TSX software will you run that can singlehandle define your purchase? Specially compared to the cons of having to wait a year longer.

Also why would you buy a Haswell-E/EP in the first place? Specially the EP? Everyone buying servers will buy Haswell-EP like nothing was changed and they all cheer on the reduced TCO. Just like they did with every model before and before that and so on.

The problem is, that I might do all sorts of things with a Haswell-E/EP, over the next five years. (If I had got one).

For example (amongst the numerous other things I might do), I could play against a many-core, very powerful chess program, to solve difficult chess problems/studies and stuff like that. Potentially some of the Chess software I try out, may indeed use TSX (for all I know), which gets released over the next five years, who knows ?

Given that such a computer, can cost thousands of dollars, depending on the exact configuration, I would worry, that it might be problematic.

You saying to me "it does not matter, TSX is rarely used", could well be good advice, but people worry about things like this, especially us enthusiasts, so it WOULD put me completely off it. Sorry.

 

[ShintaiDK]

The problem is, that I might do all sorts of things with a Haswell-E/EP, over the next five years. (If I had got one).

For example (amongst the numerous other things I might do), I could play against a many-core, very powerful chess program, to solve difficult chess problems/studies and stuff like that. Potentially some of the Chess software I try out, may indeed use TSX (for all I know), which gets released over the next five years, who knows ?

Given that such a computer, can cost thousands of dollars, depending on the exact configuration, I would worry, that it might be problematic.

You saying to me "it does not matter, TSX is rarely used", could well be good advice, but people worry about things like this, especially us enthusiasts, so it WOULD put me completely off it. Sorry.

Chess isnt something you would gain much, if anything from TSX. However if you start to run large databases with heavy load. Then its a whole other matter. But the database software you need is quite some years away.

But talking about your chess. Your chess game may support AVX2. It also supports more cores. So now you actually settled that you are willing to penalize yourself, due to TSX for at least 1 year. Are you sure you still need the E or EP at all? That doesnt sound very enthusiastic.

Even within 5 years, I doubtful that there will even be anything outside the HPC area that would use TSX. However this bugged TSX in Haswell only delays that case even further. And thats unfortunate for the technology adoption.

But something that may support your case somewhat. Specially because there is a software caverat. There are 2 TSX methods. And one of them isnt compatible at all for non TSX CPUs and thats where the real gain is. However the obviously next question would be, who would make exclusive software to a niche market?

Also more information here:
https://software.intel.com/en-us/blogs/2012/02/07/transactional-synchronization-in-haswell

 

[SOFTengCOMPelec]

Chess isnt something you would gain much, if anything from TSX. However if you start to run large databases with heavy load. Then its a whole other matter. But the database software you need is quite some years away.

But talking about your chess. Your chess game may support AVX2. It also supports more cores. So now you actually settled that you are willing to penalize yourself, due to TSX for at least 1 year. Are you sure you still need the E or EP at all? That doesnt sound very enthusiastic.

Even within 5 years, I doubtful that there will even be anything outside the HPC area that would use TSX. However this bugged TSX in Haswell only delays that case even further. And thats unfortunate for the technology adoption.

But something that may support your case somewhat. Specially because there is a software caverat. There are 2 TSX methods. And one of them isnt compatible at all for non TSX CPUs and thats where the real gain is. However the obviously next question would be, who would make exclusive software to a niche market?

Also more information here:
https://software.intel.com/en-us/blogs/2012/02/07/transactional-synchronization-in-haswell

This is getting more and more like the original Pentium divide bug.

All sorts of experts, advised that the probability of the divide error, ever occurring was minimal, and practically non-existent, so please don't worry. Not surprisingly, Intel liked this news.

If you are old enough to remember, a big storm brewed up in the computing world, and people became furious with Intel, and their faulty chips. So eventually, under immense public pressure, Intel relented, and changed their policy, to allow anyone with an effected (bugged) chip, to get a free replacement. The debacle cost Intel millions (if I remember, correctly).

To varying extents, the same thing could happen all over again.

The problem is, ShintaiDK, technically speaking and in practice, you are probably right. But human nature, may well rule here.

My worry is that one or more software packages that I want to use both now and in the future, WILL need TSX, to get the best out of it. So I am put off.

As regards my Haswell-E/EP replacement, it could be Skylake, but I would have to think about it more. Or wait for Broadwell-E/EP, which I was tempted to anyway, because it should be better.

 

[gdansk]

No one, except computer scientists, use TSX. Seriously. It is surprising there exists a bug, but clearly the only correct solution to prevent errors is to disable it. Intel announced the bug, as they should.

What more do you want? You can always go back to the STM solution that you used before Haswell. As far as I'm aware, most transaction memory usage is STM anyway. Any common application will need a software fallback, as TSX won't be in ARM or AMD chips for a few years.

 

[TerryMathews]

No one, except computer scientists, use TSX. Seriously. It is surprising there exists a bug, but clearly the only correct solution to prevent errors is to disable it. Intel announced the bug, as they should.

What more do you want? You can always go back to the STM solution that you used before Haswell. As far as I'm aware, most transaction memory usage is STM anyway. Any common application will need a software fallback, as TSX won't be in ARM or AMD chips for a few years.

And if someone bought their chip because it supported TSX?

I suspect those of you who see nothing wrong with Intel disabling TSX would be singing a different tune if we were talking about Nvidia disabling buggy DisplayPort ports.

 

[ShintaiDK]

This is getting more and more like the original Pentium divide bug.

All sorts of experts, advised that the probability of the divide error, ever occurring was minimal, and practically non-existent, so please don't worry. Not surprisingly, Intel liked this news.

If you are old enough to remember, a big storm brewed up in the computing world, and people became furious with Intel, and their faulty chips. So eventually, under immense public pressure, Intel relented, and changed their policy, to allow anyone with an effected (bugged) chip, to get a free replacement. The debacle cost Intel millions (if I remember, correctly).

To varying extents, the same thing could happen all over again.

The problem is, ShintaiDK, technically speaking and in practice, you are probably right. But human nature, may well rule here.

My worry is that one or more software packages that I want to use both now and in the future, WILL need TSX, to get the best out of it. So I am put off.

As regards my Haswell-E/EP replacement, it could be Skylake, but I would have to think about it more. Or wait for Broadwell-E/EP, which I was tempted to anyway, because it should be better.

TSX is being disabled. So you are not gonna have any errors from it.

Why buy Broadwell-E/EP when Skylake-E/EP got AVX3.2. How can you live without?

It doesnt sound like you was to buy Haswell-E/EP anyway, TSX or not.

 

[podspi]

Ouch -- it does lend some credence to Intel's crazy lag-time between consumer and professional products though. In this case it still wasn't long enough, but it definitely doesn't pay to be on the cutting edge when you're doing Serious Business (TM).

Also, the first Broadwell chips have no need for TSX, so really not a big deal at all that my broadwell tablet won't have TSX.

I feel sorry for Intel's QA dept. These things happen, but it is always better when it happens to somebody else

 

[Fjodor2001]

No one, except computer scientists, use TSX. Seriously.

I think TSX has been quite a common feature to consider when getting a new CPU. At least among the enthusiasts, which is a crowd you do not want to piss off, since it results in them spreading bad PR to the masses. I know a lot of people even let it sway their buying decision from e.g. a 4770K to a 4770 to get TSX if they did not intend to overclock anyway.

Regardless of to what extent TSX is used or not in SW today, I think people expect to get what they have payed for.

Also, if TSX is that pointless, why include it in the first place in those CPUs? And why use it as a differentiating factor between SKUs?

 

[SOFTengCOMPelec]

TSX is being disabled. So you are not gonna have any errors from it.

Why buy Broadwell-E/EP when Skylake-E/EP got AVX3.2. How can you live without?

It doesnt sound like you was to buy Haswell-E/EP anyway, TSX or not.

I am partly over-reacting here. One can/could buy haswell-E/EP now (soon), and live without TSX, as an option, if one wants to.

For a long time now, buyers of the Intel ####K overclockable versions of the cpus, have mostly had to do without a whole bunch of options, such as the full virtualization stuff and TSX (lol).

So it can't be that bad.

If I was to talk to someone in the street, and mention "TSX" (Intel), they probably would not have a clue what I was talking about, in 99.999% of the population.

 

[ShintaiDK]

I think TSX has been quite a common feature to consider when getting a new CPU. At least among the enthusiasts, which is a crowd you do not want to piss off, since it results in them spreading bad PR to the masses. I know a lot of people even let it sway their buying decision from e.g. a 4770K to a 4770 to get TSX if they did not intend to overclock anyway.

Regardless of to what extent TSX is used or not in SW today, I think people expect to get what they have payed for.

Also, if TSX is that pointless, why include it in the first place in those CPUs? And why use it as a differentiating factor between SKUs?

If you are an enthusiast you would know the purpose of TSX. And not just buy it for a checkbox marker.

You have to include it before someone can develop for it.

 

[Idontcare]

That is a bummer. It is hard to define and justify a given level of QRA expenditure when validating a new design, everything is risk vs reward and the risk of the unknown is a very difficult thing to quantify.

I'm sure with hindsight there is a handful of engineers or project managers who can rightly claim this could have been avoided had their validation budget been enhanced by a couple million bucks, or given an extra 2 weeks at the time, etc etc.

But eventually you need to ship product and pay bills, and someone has to draw the line between being overly conservative (and costly) in QRA versus being appropriately thorough. Risk versus reward.

 

[SOFTengCOMPelec]

That is a bummer. It is hard to define and justify a given level of QRA expenditure when validating a new design, everything is risk vs reward and the risk of the unknown is a very difficult thing to quantify.

I'm sure with hindsight there is a handful of engineers or project managers who can rightly claim this could have been avoided had their validation budget been enhanced by a couple million bucks, or given an extra 2 weeks at the time, etc etc.

But eventually you need to ship product and pay bills, and someone has to draw the line between being overly conservative (and costly) in QRA versus being appropriately thorough. Risk versus reward.

Given that it has taken over a year, for anyone to "noticeably" find/report this bug/issue, it could be EXTREMELY technically difficult, to find it beforehand. Hindsight is a wonderful tool.

I'm convinced that there is probably no such thing, as a 100,000,000 line piece of software, which is absolutely 100% bug free. So I guess a multi billion transistor cpu, is at least partly similar.

Whenever I've seen/reviewed cpu errata, there can be (usually are) hundreds of tiny little bugs/faults, most of which, most people never know about. But this seems to be a bigger bug, than usual. They are usually amazingly obscure things, which would almost never occur in practice.

E.g. If you are in the middle of a DecA instruction, and reset occurs, and the cpu voltage is above 5.41 Volts (still within spec), a race hazard within the cpu, may cause the xyz register to NOT reset to all 0's correctly. It has been decided that it is highly unlikely this would affect anyone in practice, because the xyz register, is almost always set to a value before use. (I just made that up!).

 

[Fjodor2001]

If you are an enthusiast you would know the purpose of TSX. And not just buy it for a checkbox marker.

You have to include it before someone can develop for it.

I agree, but nothing of what you wrote contradicts what I said.