Phenom VS. Intel Quadcores (Q6600) ?

There's rarely any commercial code that will trigger this bug. Most multi-threaded programming today is vastly parallel in nature, which given the nature of the description, wouldn't set it off. So I see people trying to "duplicate" the bug by attempting to load all 4 cores via Orthos (even in a VM environment) and claiming there is no bug when it passes after 24 hours. Thats a wrong way of attempting to recreate the issue. It's not creating any kind of cache thrashing at a serious level (because the general assumption for programmers just 2+ years back is that you need to avoid hitting another core's cache because it'd go through a slow bus). The issue is about multiple cores hitting and competing on the same allotment of the shared cache, not about stressing the cores in parallel.

If you want to trigger the TLB bug consistently, use software that is has large amounts of cache thrashing, generally application servers or hpc research.

As for whether or not this matters, I think this is on the order of the FDIV Pentium issue in terms of seriousness and exposure. It won't happen to the vast majority of the customer base, but is generally unacceptable in comparison to previous marketed hardware. I think its fairly asinine to think "if it doesn't happen to me, its not a problem" - it just lowers the bar for future hardware.
 
Why are you getting defensive?
I'm not making anything out of this bug. Nowhere have I claimed anything about how often it occurs or how severely you will be affected.
In fact, I'm here to try and find out exactly that, because there seems to be no data on this subject.
Why does everyone with a Phenom get so defensive (and in effect are making a big deal out of nothing... just test the thing and get it over with)?
Must be the psychological aspect to the bug. But I'm not interested in psychology, I'd just like a scientific approach to determining how rare this bug really is.
I'm not trying to hurt AMD. In fact, it may actually help AMD if the results are positive enough.



I am far from 'getting' anything about it.And my comment was far more directed at what I have seen others say on other forums,not you here.I dont have a Phenom yet,nor am I getting defensive about it.I respect your opinion about it,and enjoy many of your posts.I was not trying to take a jab at you,seriously. :) If you thought that,I apologize.Maybe sometimes I am not as clear as I think I am in my posts.

I know its hard sometimes with forum posts,but try not to read inbetween the lines of what I post to heavily.After everything I have said and posted about AMD this past 12 months,no one here will ever be able to call me,or label me a fanboi or apologist for eithier side.I have laid into both companies,quite heavily on this forum over the years.. :eek:
 
http://www.xtremesystems.org/forums/showpost.php?p=2648015&postcount=11

There it is. Thats the only thing you need to know. I think we can all move on from this silly conversation.

There's rarely any commercial code that will trigger this bug. Most multi-threaded programming today is vastly parallel in nature, which given the nature of the description, wouldn't set it off. So I see people trying to "duplicate" the bug by attempting to load all 4 cores via Orthos (even in a VM environment) and claiming there is no bug when it passes after 24 hours. Thats a wrong way of attempting to recreate the issue. It's not creating any kind of cache thrashing at a serious level (because the general assumption for programmers just 2+ years back is that you need to avoid hitting another core's cache because it'd go through a slow bus). The issue is about multiple cores hitting and competing on the same allotment of the shared cache, not about stressing the cores in parallel.

If you want to trigger the TLB bug consistently, use software that is has large amounts of cache thrashing, generally application servers or hpc research.

As for whether or not this matters, I think this is on the order of the FDIV Pentium issue in terms of seriousness and exposure. It won't happen to the vast majority of the customer base, but is generally unacceptable in comparison to previous marketed hardware. I think its fairly asinine to think "if it doesn't happen to me, its not a problem" - it just lowers the bar for future hardware.
Thanks to both of you for the explaination. So I guess that a single threaded application and a multithreaded application that didn't access the same cache won't be affected at all? So there's rarely any commercial code that will trigger this bug and luckily for non-commercial code that will trigger it, there is a fix for that.
 
There's rarely any commercial code that will trigger this bug. Most multi-threaded programming today is vastly parallel in nature, which given the nature of the description, wouldn't set it off. So I see people trying to "duplicate" the bug by attempting to load all 4 cores via Orthos (even in a VM environment) and claiming there is no bug when it passes after 24 hours. Thats a wrong way of attempting to recreate the issue. It's not creating any kind of cache thrashing at a serious level (because the general assumption for programmers just 2+ years back is that you need to avoid hitting another core's cache because it'd go through a slow bus). The issue is about multiple cores hitting and competing on the same allotment of the shared cache, not about stressing the cores in parallel.

If you want to trigger the TLB bug consistently, use software that is has large amounts of cache thrashing, generally application servers or hpc research.

As for whether or not this matters, I think this is on the order of the FDIV Pentium issue in terms of seriousness and exposure. It won't happen to the vast majority of the customer base, but is generally unacceptable in comparison to previous marketed hardware. I think its fairly asinine to think "if it doesn't happen to me, its not a problem" - it just lowers the bar for future hardware.

Ah, so there is someone else who understands the bug at a deeper level and agrees with what I said about stress-testing :)
 
Thanks to both of you for the explaination. So I guess that a single threaded application and a multithreaded application that didn't access the same cache won't be affected at all?

So why don't I get thanks?
I was the first to explain it here. But apparently it isn't true when I say it.
In fact, I think an apology is more in order than a thanks.
 
I am far from 'getting' anything about it.And my comment was far more directed at what I have seen others say on other forums,not you here.I dont have a Phenom yet,nor am I getting defensive about it.I respect your opinion about it,and enjoy many of your posts.I was not trying to take a jab at you,seriously. :) If you thought that,I apologize.Maybe sometimes I am not as clear as I think I am in my posts.

Well I thought about this line in particular: "I did my job,and feel no guilt or worry over not testing for some beyond incredibly obscure
'bug'"

As if you thought I was accusing you of not testing the system properly, and you defending yourself for that.
That wasn't what I meant to say. I don't normally expect people to stress-test any PC that they build... not beyond the usual "play some games and run some Prime/Orthos" or anything.
I just meant to say that you and Kyle didn't really provide any more data other than "It didn't crash on me yet", because alg7_munif constantly mentions you as 'proof' that there is no bug in Phenom.
 
So I guess that a single threaded application and a multithreaded application that didn't access the same cache won't be affected at all? So there's rarely any commercial code that will trigger this bug and luckily for non-commercial code that will trigger it, there is a fix for that.

That's not what he says.
It has nothing to do with whether code is commercial or not.
It has more to do with the fact that a single process only uses a single address space, even if it has multiple threads (so loading 4 cores with 1 application makes no difference really).
Look at his example of an application server. That's when you will be running many processes side by side and switching between address spaces a lot etc, and trashing the cache constantly.
All these applications may still be commercial and even single-threaded... But together they might form the 'deadly cocktail' for the TLB bug.
 
So why don't I get thanks?
I was the first to explain it here. But apparently it isn't true when I say it.
In fact, I think an apology is more in order than a thanks.
You said:
The equivalent of this for the TLB bug is to run virtualization and a 100% system load for any length of time. Why don't you and Manny run that for a few days, and see if the system crashes?
Then we will at least have a rough idea about the worst case scenario for this bug.
However dexvx said
So I see people trying to "duplicate" the bug by attempting to load all 4 cores via Orthos (even in a VM environment) and claiming there is no bug when it passes after 24 hours. Thats a wrong way of attempting to recreate the issue.
I just meant to say that you and Kyle didn't really provide any more data other than "It didn't crash on me yet", because alg7_munif constantly mentions you as 'proof' that there is no bug in Phenom.
No one has ever said that the bug is not there, what others are trying to say that the bug is so rare that it won't affect most user and should not be a concern to most user even with the fix disabled because like dexvx said
There's rarely any commercial code that will trigger this bug.
I think that narrows it down much further. Tell me how many softwares that most user use are multithreaded, and how many of those multithreaded codes will use the same cache? To narrow it down much further, let's recall what JumpingJack said on XtremeSystems:
But a strong point needs to be made.... that window of time is likely very small, so small that the probability of happening is slim to none ...

If you think this bug really bothers you, don't buy a Phenom, or when you need to use a Phenom, just enable the fix. I would like to follow W4ngsta suggestion from now.
 
So why don't I get thanks?
I was the first to explain it here. But apparently it isn't true when I say it.
In fact, I think an apology is more in order than a thanks.

because you don't explain anything you just make a claim and everyone else is wrong without any shred of an actual explanation :p
 
You said:
However dexvx said

You must be forgetting the posts where I mention that Prime95 is not a good stresstest period: http://www.hardforum.com/showpost.php?p=1031896692&postcount=146
I also say to use many context switches (meaning multiple processes running at the same time).
And I probably mentioned similar things in various other posts.
All in all, I've already said everything that dexvx said in this thread.

I never said that there is no bug in the CPU but I said the bug is so rare that it won't affect most user and should not be a concern to most user

What is most users? 6 out of 10?
 
because you don't explain anything you just make a claim and everyone else is wrong without any shred of an actual explanation :p

Nobody who didn't understand it actually asked any specific questions.
Also, I never made any claims, I just pointed to information coming straight from AMD.
Let's face it, you and alg7_munif just don't want to admit that you don't understand it, that's why you didn't ask me to explain in more detail. I offered to explain, just didn't see what was left to explain. And as you can see, I already mentioned that Prime95 is not a good stresstest because you want lots of context switches, and Prime95 doesn't deliver those. Same as what dexvx says.
If you're going to act as cocky and condescending towards me as you guys have been doing, I'm going to assume you know what you are talking about. It's such a loss of face when it turns out you're talking out of your arse with such a big mouth, isn't it?
Don't blame me, you were pretending to be more knowledgable on the subject than me, and refused to ask questions when I offered.

In fact only just now I had to correct alg7_munif's interpretation of what dexvx said, so even his explanation wasn't understood very well.
 
Actually you suggested running virtualization and a 100% system load before mentioning anything about context switches.
http://www.hardforum.com/showpost.php?p=1031896530&postcount=138

Tell me what you really want? I don't think that what Manny, Kyle, Pepsiennis and me have said before is wrong, even JumpingJack on XtremeSystems and Dexvx have made it clear that the bug won't affect most user so it is not much of a concern for most user. What is the point you want to make actually? To quote someone:
If you could sum up your point into a few sentences, what would they be?
 
Actually you suggested running virtualization and a 100% system load before mentioning anything about contact switches.
http://www.hardforum.com/showpost.php?p=1031896530&postcount=138

Contact switches, lol.
Besides, I didn't realize how little you understood, so I didn't think it was necessary to explain the 100% system load at that point. I explained later, when I realized you know much less about it than you pretend.
This "contact switches" is yet another example of how you still really are completely clueless to the whole matter.
 
it shouldn't be too hard to write an app that will stress the l3 cache didn't you write some code that ran much better on a core 2 due to the cache structure? you should have most of it there then ^^?
 
I'll admit that I don't have much knowledge about programming, but I'm more clueless about the point that you want to make? At least with my limited programming knowledge, I have sum up my points in this post, and I still don't think that it is wrong. To write it up in a sentence, even with the fix disabled, I think that bug won't affect most user so it should not be much of a concern for most user
 
it shouldn't be too hard to write an app that will stress the l3 cache didn't you write some code that ran much better on a core 2 due to the cache structure? you should have most of it there then ^^?

Geez, don't you understand anything?
It's the TLB of the L3 cache, not the L3 cache itself.
You can use the L3 cache all day, but if the TLB never gets updated, the bug never gets triggered.
 
Scali2 can you answer me?
I'll admit that I don't have much knowledge about programming, but I'm more clueless about the point that you want to make? At least with my limited programming knowledge, I have sum up my points in this post, and I still don't think that it is wrong. To write it up in a sentence, even with the fix disabled, I think that bug won't affect most user so it should not be much of a concern for most user
So can you tell me what is the point that you are trying to make exactly?
 
when is the TLB updated?

In short: whenever virtual memory is addressed whose physical memory is not currently in the TLB.
The TLB is a buffer, a sort of cache. It stores recently used memory addresses for quick virtual->physical address translation (hence the name Translation Lookaside Buffer... the CPU 'looks aside' in the buffer to see if the address was recently translated... if so, use the buffered address, if not perform translation and buffer the address).

Since each process has its own virtual address space (which, together with some other process-specific data like register contents etc is known as a 'context'), the TLB is generally updated after a task switch aka context switch.

This also explains why virtualization is the most common situation where the bug may arise. Because you are actually running two or more OSes on the same physical CPU, each in their own virtual environment, and all their virtual address spaces, there's a lot of translation going on.

But as dexvx mentioned, an application server (where hundreds of users may each run their own processes side-by-side) is also a case where you'll experience lots of context switches and TLB updates.

For more info, see http://en.wikipedia.org/wiki/Translation_Lookaside_Buffer
 
In short: whenever virtual memory is addressed whose physical memory is not currently in the TLB.
The TLB is a buffer, a sort of cache. It stores recently used memory addresses for quick virtual->physical address translation (hence the name Translation Lookaside Buffer... the CPU 'looks aside' in the buffer to see if the address was recently translated... if so, use the buffered address, if not perform translation and buffer the address).

Since each process has its own virtual address space (which, together with some other process-specific data like register contents etc is known as a 'context'), the TLB is generally updated after a task switch aka context switch.

This also explains why virtualization is the most common situation where the bug may arise. Because you are actually running two or more OSes on the same physical CPU, each in their own virtual environment, and all their virtual address spaces, there's a lot of translation going on.

But as dexvx mentioned, an application server (where hundreds of users may each run their own processes side-by-side) is also a case where you'll experience lots of context switches and TLB updates.

For more info, see http://en.wikipedia.org/wiki/Translation_Lookaside_Buffer

Thanks. I still don't understand why its hard to reproduce the problem, if it's known what causes it.
 
Thanks. I still don't understand why its hard to reproduce the problem, if it's known what causes it.
About the errata:
“The processor operation to change the accessed or dirty bits of a page translation table entry in the L2 from 0b to 1b may not be atomic. A small window of time exists where other cached operations may cause the stale page translation table entry to be installed in the L3 before the modified copy is returned to the L2. In addition, if a probe for this cache line occurs during this window of time, the processor may not set the accessed or dirty bit and may corrupt data for an unrelated cached operation. The system may experience a machine check event reporting an L3 protocol error has occurred. In this case, the MC4 status register (MSR 0000_0410) will be equal to B2000000_000B0C0F or BA000000_000B0C0F. The MC4 address register (MSR 0000_0412) will be equal to 26h."
From what I understand, to reproduce the problem, a few things need to happen in this small window of time.
 
Thanks. I still don't understand why its hard to reproduce the problem, if it's known what causes it.

Mainly because it is related to L3-cache, and the actual timeframe in which a problem can occur is only a handful of cycles.
Since each core also has its L1 and L2 caches, generally they won't need to address L3-cache at all.
So it's just an odd combination of circumstances...
One core must be updating a TLB entry in L3 cache AND another core must be accessing that same entry within a few cycles...
That just depends on a whole lot of 'coincidence'.
The situation WILL occur sooner or later, but it may take a long time.
It was actually the major OEMs that found the bug and reported it back to AMD.

It's not all that hard to create a synthetic test that tries to generate as many TLB updates on the L3 cache as possible, and subsequent accesses from different cores within a short timeframe... But doing that just with regular OSes and applications, that's near impossible. There are so many things going on in the background on a PC, that you can never really predict what is in a buffer at any particular time. Any kind of multitasking or multithreading is quite hard to reproduce exactly... and this bug requires pretty much cycle-exact timing to induce... Either that, or coincidence.
So that's the problem here. You may never see the bug, doesn't mean it isn't there. You can never tell when the bug occurs either. So in theory it is possible that you use a system for years without a problem... and then the bug decides to hit just at a crucial point where your entire HDD is corrupted... or whatever other doom scenario you can think of.
And that is why AMD recalled the server CPUs altogether, and enables the TLB fix by default.
 
Mainly because it is related to L3-cache, and the actual timeframe in which a problem can occur is only a handful of cycles.
Since each core also has its L1 and L2 caches, generally they won't need to address L3-cache at all.
So it's just an odd combination of circumstances...
One core must be updating a TLB entry in L3 cache AND another core must be accessing that same entry within a few cycles...
That just depends on a whole lot of 'coincidence'.
The situation WILL occur sooner or later, but it may take a long time.
It was actually the major OEMs that found the bug and reported it back to AMD.

It's not all that hard to create a synthetic test that tries to generate as many TLB updates on the L3 cache as possible, and subsequent accesses from different cores within a short timeframe... But doing that just with regular OSes and applications, that's near impossible. There are so many things going on in the background on a PC, that you can never really predict what is in a buffer at any particular time. Any kind of multitasking or multithreading is quite hard to reproduce exactly... and this bug requires pretty much cycle-exact timing to induce... Either that, or coincidence.
So that's the problem here. You may never see the bug, doesn't mean it isn't there. You can never tell when the bug occurs either. So in theory it is possible that you use a system for years without a problem... and then the bug decides to hit just at a crucial point where your entire HDD is corrupted... or whatever other doom scenario you can think of.
And that is why AMD recalled the server CPUs altogether, and enables the TLB fix by default.
So how is your point different from what I've said before in this post? We never said that the bug is not there, what we said before is even with the fix disabled, the bug is so rare that it won't affect most user and shouldn't be much of a concern for most user. How is running without the fix different from running an OCed CPU? If you are talking about crucial point, it is just like using a computer without an UPS, you never know exactly when you are doing something important and suddenly the power station had a malfunction.

Please tell me what is exactly the point that you are trying to make, from what I can see from your previous posts, you just want to make the issue looks bigger than what it really is. You said before that I should ask you to explain when I don't understand anything that you said but I guess that this is the third time I asked you to explain what is the point that you are trying to make exactly.
 
You guys are just too easy??? Fight the good fight!!!!!!!!!!!!!!!!

Maybe you should start a new thread about tlb errata and how to reproduce it ??????????
Way over my head though, or maybe not??

KISS!!!!! Keep it simple ---------??????????

Kind of a triple post by now but I believe it still pertains.

The reason I started buying amd is not because they were faster than intels but at a lower price and with some work (OC'ing) they could catch or overtake them and at a much more reasonable price. Oh yeh I definitely can keep par with the e6600's with my brisbane. The only ones scoring better than me on pcmark05 are mostly the ones clocked 3-3.5. So you can take your intel and???? I've got a $40-50 brisbane keeping pace with $200 cpu's. If we all had a grand to blow we'd all have yorkies @ 4ghz????? Yeh you know what intel can blow at that price!!!!! And yes I bought an EPoX board knowing very well there would probably be no future support for it but still hoping they'll throw a bone my way from supox?? They never were to quick on updates but were/are great boards.
 
You guys are just too easy??? Fight the good fight!!!!!!!!!!!!!!!!

Maybe you should start a new thread about tlb errata and how to reproduce it ??????????
Way over my head though, or maybe not??

KISS!!!!! Keep it simple ---------??????????

Kind of a triple post by now but I believe it still pertains.

The reason I started buying amd is not because they were faster than intels but at a lower price and with some work (OC'ing) they could catch or overtake them and at a much more reasonable price. Oh yeh I definitely can keep par with the e6600's with my brisbane. The only ones scoring better than me on pcmark05 are mostly the ones clocked 3-3.5. So you can take your intel and???? I've got a $40-50 brisbane keeping pace with $200 cpu's. If we all had a grand to blow we'd all have yorkies @ 4ghz????? Yeh you know what intel can blow at that price!!!!! And yes I bought an EPoX board knowing very well there would probably be no future support for it but still hoping they'll throw a bone my way from supox?? They never were to quick on updates but were/are great boards.
What are on mate? Would like to have some :D

Btw I guess that Scali2 doesn't actually know what is the point that he is trying to make other than making AMD looks bad.
 
alg7_munif,

No offence, but trying to compare an overclocked Q6600 to a Phenom without the TLB fix is laughable.

An overclocked CPU is exactly that - an overclocked CPU. An overclocked CPU doesn't suddenly develop a TLB bug. ;) The things that can make an overclocked CPU unstable is insufficient core voltage, excess heat, or chipset limitations (FSB wall etc.)

A buggy CPU and an overclocked CPU are totally different, in the case of Phenoms the bug may be rare but it CAN be triggered by a certain sequence of events. With an overclocked CPU, there is no random bug that can cause it to crash. With adequate cooling and voltage, an overclocked CPU can be just as stable as a stock CPU.
 
alg7_munif,

No offence, but trying to compare an overclocked Q6600 to a Phenom without the TLB fix is laughable.

An overclocked CPU is exactly that - an overclocked CPU. An overclocked CPU doesn't suddenly develop a TLB bug. ;) The things that can make an overclocked CPU unstable is insufficient core voltage, excess heat, or chipset limitations (FSB wall etc.)

A buggy CPU and an overclocked CPU are totally different, in the case of Phenoms the bug may be rare but it CAN be triggered by a certain sequence of events. With an overclocked CPU, there is no random bug that can cause it to crash. With adequate cooling and voltage, an overclocked CPU can be just as stable as a stock CPU.
So how do you draw the line between a stable OC and an unstable OC? Can you guarantee that an OC that can pass a certain period Prime95 won't crash in any other circumstances especially when your OC is pushing the limit (>4.0GHz for an example)?

No offense but by going with your logic, are you saying that with sufficiently good cooling and sufficiently high voltage, there shouldn't be any limit to an OC? I think that most OCer know that it is not true, for example when look at this thread, you can see yourself that even with a very high voltage and negative temperature, there will still be a limit for an OC.

Then you should read this article regarding to a stable OC:
Imagine our confusion as we desperately struggle to understand why our system is Prime95 stable for days yet continues to crash under absolutely no load.

Well I dont think he was comparing the processors with respect to bugs, but comparing them stability-wise.

I think its true that you can never guarentee an OCed cpu to run ABSOLUTELY stable. Sure, you can probably prime stable it for days on end, but there is always a small possibility that it could crash (could happen in weeks). A cpu that runs within its specified speed will, however, guarentee 100% stability.
Exactly what I mean.
 
Can someone post the errata (I know it's somewhere in the earlier posts but I've only found quotes to it)? I would like to look further into why the dirty bit isn't set atomically.

According to the quote of it by alg7_munif, this seems like a RAW data hazard with multiple cores instead of on a pipeline.
 
CoW]8(0);1031899057 said:
Can someone post the errata (I know it's somewhere in the earlier posts but I've only found quotes to it)? I would like to look further into why the dirty bit isn't set atomically.

According to the quote of it by alg7_munif, this seems like a RAW data hazard with multiple cores instead of on a pipeline.

It's not official from AMD though but better than nothing:
http://www.lostcircuits.com/cpu/amd_phenom/2.shtml

Can you post your findings here?
 
So how do you draw the line between a stable OC and an unstable OC? Can you guarantee that an OC that can pass a certain period Prime95 won't crash in any other circumstances especially when your OC is pushing the limit (>4.0GHz for an example)?

You have to realise that there is no such thing as an absolute 100% stable overclock. There is also no such thing as an absolute 100% stable stock CPU either. ALL CPU have errata that can trigger in very obscure situations. The vast majority of erratum in CPUs are left unfixed, since they are deemed VERY rare or simply aren't significant enough to warrant a fix. However, it does appear the TLB errata happens to be more severe than your average errata.

In regards to Prime95 stability, it's certainly not the be all end all of stability testing, there are programs that can push a CPU harder (IMO OCCT is more CPU intense than Prime95/Orthos), in fact I've found that intense gaming is often a more accurate judge of system stability, I've had games crash on supposedly 'stable' overclocks according to stress test programs. So, to answer your question, I would draw the line of a 'stable' overclock as one that never crashes on YOUR preferred applications, over a prolonged period of time.
 
You have to realise that there is no such thing as an absolute 100% stable overclock. There is also no such thing as an absolute 100% stable stock CPU either. ALL CPU have errata that can trigger in very obscure situations. The vast majority of erratum in CPUs are left unfixed, since they are deemed VERY rare or simply aren't significant enough to warrant a fix. However, it does appear the TLB errata happens to be more severe than your average errata.

Does having an option to disable the fix means that the errata is less severe than an errata that doesn't have the option to disable the fix?

Btw before this turns out to be another long argument, I would like to ask you the same question that I asked Scali2 before, what is the point that you want to make with this errata argument? My point would still be the same and that is "the bug is so rare that even with the fix disabled, it won't affect most user and shouldn't be much of a concern for most user". I made this conclusion based on my own experience working with two Phenom systems without the fix, by reading other Phenom users experience without the fix and also a few explainations on the errata by Dexvx, JumpingJack and even Scali2.
 
Well it seems that the errata is just that quote. I'm guessing AMD won't be revealing any of their trade secrets.

The issue is quite simple. A CPU isn't writing to the L2 cache fast enough before the L2 cache updates the data in the L3 cache.

In other words, the L2 cache is updating the L3 cache with old data because the update occurs before a CPU actually writes to the L2 cache.
 
Does having an option to disable the fix means that the errata is less severe than an errata that doesn't have the option to disable the fix?

No. Why?
Simple: This fix has no negative side-effects like say 10-15% performance drop across the board. There is no reason not to have it installed, because there are no disadvantages to it.

In fact, if there was no performance drop, AMD probably wouldn't have the option to disable it either. It'd be like most minor errata fixes... You just flash your BIOS, get the new microcode and that's that. You probably don't realize how often you've already updated your microcode on various systems by flashing a newer BIOS.

The severity isn't in the bug itself, but rather in the fix that degrades performance significantly. Fixes rarely have such negative side-effects, and therefore are rarely highly publicized.
 
CoW]8(0);1031899277 said:
Well it seems that the errata is just that quote. I'm guessing AMD won't be revealing any of their trade secrets.

The issue is quite simple. A CPU isn't writing to the L2 cache fast enough before the L2 cache updates the data in the L3 cache.

In other words, the L2 cache is updating the L3 cache with old data because the update occurs before a CPU actually writes to the L2 cache.

How about the second part:
In addition, if a probe for this cache line occurs during this window of time, the processor may not set the accessed or dirty bit and may corrupt data for an unrelated cached operation. The system may experience a machine check event reporting an L3 protocol error has occurred. In this case, the MC4 status register (MSR 0000_0410) will be equal to B2000000_000B0C0F or BA000000_000B0C0F. The MC4 address register (MSR 0000_0412) will be equal to 26h.
 
Yes. Why?
Simple: This fix has no negative side-effects like say 10-15% performance drop across the board. There is no reason not to have it installed, because there are no disadvantages to it.

In fact, if there was no performance drop, AMD probably wouldn't have the option to disable it either. It'd be like most minor errata fixes... You just flash your BIOS, get the new microcode and that's that. You probably don't realize how often you've already updated your microcode on various systems by flashing a newer BIOS.

The severity isn't in the bug itself, but rather in the fix that degrades performance significantly. Fixes rarely have such negative side-effects, and therefore are rarely highly publicized.
So what is the point that you wanted to make with this long errata argument?

Tell me how many of those Dell, HP, and other OEM system users that have actually updated their BIOS?

Does this means that someone, who bought an OEM C2D system in the early days before the microcode update was out and he has never flashed the BIOS before, is running a computer with a bigger risk than someone who is running a Phenom without the fix?
 
How about the second part:

Basically that means if a processor tries to access or modify that piece of memory (I'm not sure if they're referring to L2 or L3, probably L3 since they say processor), it won't be able to to write to the accessed or dirty bits of that piece of memory. The accessed and dirty bits indicate whether that piece of memory was modifed or read by a processor.

This means that there's no record of what that processor did to that piece of memory, it could have modified it or simply read, either way its considered corrupt.
 
Back
Top