"When Overclocking Is Not Overclocking"

[Goosey]

PM: I am not saying that having the lowerend chips is bad. And charging more for the higher-end chips is normal! I just feel that in most cases a 800EB probably came from the same wafer as the 600E.

You can buy dozen and dozen of 600E and probably over 98% will make it to 800EB easily. I wouldn't use the word "Identical", but with a little voltage tweak you are just about guaranteed!

 

[pm]

And I'm not denying that they come from the same wafer in some cases (generally parts in the middle of a wafer are faster, ones on the outside edge much slower).

And whether or not you think a part makes it to 800MHz is mostly due to the fact that when you overclock your expectations are much lower than someone who buys it retail in, say, a Dell computer.

For one thing, the part needs to work at 800MHz for 7 years... you probably don't care about this, but CPU's do slow down with time (I can get you references for this too). So, in 7 years you may not be able to run your 600E at 800MHz, but Intel guarantees that theirs will. Most of the people on this BBS upgrade frequently, but my father still has a working 80286 that they use to do some work at his business...

Second, you may need to increase the voltage on your part to get it to 800MHz (which dramatically cuts the lifetime of the part), or put a bigger heatsink on it (which costs you, the enduser, more money), or you may need to get a special motherboard (one with particularly good bypass cap on it), or you may need special memory, etc. And a lot of threads on here run something along the lines of "I have a Malaysian WW32 SLQUR - I heard this OC's well, but mine won't even do 650MHz. What am I doing wrong." People hunt for specific steppings, SSpec's, production weeks and packaging sites. I know most "normal" people at home wouldn't put up with this stuff. They want to pay money and have it work right the first time.

Third, in your case, if it processor crashes occassionally, well, you tweak it a little and hope for the best. If it crashes a lot, you sell it to someone and buy another one. Hobbyists do it because it's their hobby, but a person who buys a Pentium III 800MHz and has problems with it is going to complain and return it.

My point is that your expectations are fairly low. Intel's corporate customers have very high expectations for reliability. When you are overclocking you are eating into this margin and hoping for the best. It's a risk and you do it because you think it's fun (or at least not totally frustrating) and you think it saves you money.

But the fact that overclocking is possible doesn't change the point of my arguement that Intel tests it's Pentium III 800's and stands by the marking put on the chip. This means that a Pentium III 600E is not the same identical chip as a Pentium III 800EB. In fact, a Pentium III 800EB is inherently worth much more because other people know that it will work reliabily and consistently at 800MHz. So, the author is wrong.

I personally don't overclock my machine. Aside from the fact that Intel gives me fast processors to play with, I use it to work from home. Uptime is hugely important to me. I don't have time to fiddle with my machine, and when I get a blue screen (which fortunately never happens to me with Win2k) I know that I should be cursing Bill G. and not poking around to see if MBM thinks my CPU temp is too high.

 

[NOX]

<< The wafer (a big wide chunk of silicon with a whole bunch of die/chips on it) comes out of the fab and it's slapped into a wafer tester which tests very quickly at various temperatures to basically determine &quot;alive or dead&quot;. The dead are marked dead, and the wafer is scribed and the die (chips) are packaged. Then the real testing occurs. The packaged chips are tested thoroughly in multi-million dollar testers. The tests first use code that tests whether or not every transistor on the chip can turn on and off. Then the processors are &quot;burned in&quot;. This is a test purely designed to kill as many weak/marginal parts as possible and statistically the parts are substantially slower as burn-in than before because the test effectively ages the parts. Then they are tested at various temperatures, voltages and frequencies and the pins are checked for setup and hold time issues. The results of these tests are coupled with statistical data on the processors that is gathered pre-production to determine which parts should be binned in various binning catagories (ie. this part is a 500MHz part, this part is a 1GHz part) such that they will work over the lifetime over the part when run at spec. >>

Thanks PM for that information.

<< I just feel that in most cases a 800EB probably came from the same wafer as the 600E. >>

That's basically what PM is stating.

<< I see why he's elite at 792. >>

What's up with that? No offence, but most bust their butt trying to achieve Elite member, and you have Elite status at 792!?

 

[pm]

NOX, I never asked to be Elite and it came as complete surprise to me when I saw that I'd been up'd a few weeks ago. I was pleased and flattered, but I had nothing to do with it. If you are irritated don't blame it on me.

 

[NOX]

<< NOX, I never asked to be Elite and it came as complete surprise to me when I saw that I'd been up'd a few weeks ago. I was pleased and flattered, but I had nothing to do with it. If you are irritated don't blame it on me. >>

Like I said no offence, though it's strange you don't know why you achieved Elite Member before your 2000th post, let alone 1000th.

 

[NOX]

<< Elite status is given based on contribution to the forum, instead of by post count. >>



That's why.

 

[Eug]

That article is misleading, and it's pretty clear he's talking without true knowledge of Intel's practices. Anybody buying a 600E expecting it to run at 800EB would be naive, and it's unfortunate he's encouraging that.

It WOULD make perfect sense for a company to sell a chip that could do 800 at 600 for supply reasons, but even just a short jaunt over to overclockers.com would suggest that is not the case here (at least not entirely). While most 600E chips hit 800 MHz fine, several cB0 600E chips either do not make 800 stably at all, or else don't do it at stock voltage (1.65 V for both the 600E and the 800EB). (I'm discounting the steppings which are pre cB0 because presumably they would not be as reliable overclockers.) And remember that overclockers.com types exaggerate their speeds, too.

 

[VladTrishkin]

Well, this article is too much screwed up, i wont even go into details, but PM nailed most of the stuff.

PIII 600 same as 800? What planet did he come from? LOL

 

[pm]

Vlad: Most?!? How did I only nail most of it? I practically wrote a research paper on the subject. I can't see what I left out.

 

[MikeHelvey]

PM has certainly explained the testing/categorized/binning process. I would like to point out that it is not always so clean in that market demands do determine some mandatory bin crossing in the production line. It is obvious by the number of AMD Athlon?s marked and set at a certain speed that have processors that have contradictory markings for a higher speed. That leads me to believe that market pressures drove the manufacture to use higher grade chips (or tested to a higher grade) for a lower grade processor. Just makes sense to match the demand instead of loosing customers to a competitor who is producing more chips in the appropriate price range.

My 2 cents . . .not worth a dime.

 

[NOX]

<< PIII 600 same as 800? What planet did he come from? LOL >>

The same one most of us here came from. If you look at what PM has stated in the process of which will determine what speed a CPU is sold at you?ll notice that they come from the same ?wafer? or bin. Which tells me they?re all made the same way, and made to run between certain speeds, lets assume 600-1GHz. Though that?s not to suggest all will do 1GHz, or 800Mhz for that matter. The process of which PM has stated is clear: ?Then they are tested at various temperatures, voltages and frequencies and the pins are checked for setup and hold time issues. The results of these tests are coupled with statistical data on the processors that is gathered pre-production to determine which parts should be binned in various binning catagories (ie. this part is a 500MHz part, this part is a 1GHz part) such that they will work over the lifetime over the part when run at spec.?

So The TechZone (I don?t know his name Moto or something) is some what correct, but misleading. I think for some of us it?s no big secret that both Intel and AMD do it. The problem with the article which I saw is he?s basically suggesting all 600?s are 800?s, which is not true because Intel labeled it 600 for a reason, as the process describes. But it's not to suggest it's not done.

 

[VladTrishkin]

LOL PM, ok you nailed right about what i was about to say

 

[dZeus]

Hi, sorry for bringing up this topic again, but I have some questions that members here might be able to answer:

The article that was linked in the first post is claiming that because the CPUID of the P3-600E and the P3-800EB are the same, the CPU actually is identical. Now poeple here responded by pointing out that obviously Intel doesn't guarantee that a 600E with the same CPUID will run at 800, but that they do tests.... Of course this is very logical, but I am interested in some other things:

If the CPUID is the same... when is it determined? Before or after the process of manufacturing the wafers? If before, does this implicits that Intel locks the multiplier before manufacturing the wafers? Wouldn't this cause a lot of 'overhead' when trying to get the most out of a CPU?

I will try to explain the last thing with an example (all based on the assumption they lock them before the manufacturing process). In this example, everything is hypothetical (so what I say may not be true, or the statements might not even be supported by me when talking about this in 'real life'):

Intel locks a certain coppermine wafer with a 7x multiplier. Intel also has a very high demand for 800MHz+ CPUs. 60% of the CPUs on the wafer test at 933MHz, 10% is dead and 30% is guaranteed at 700MHz. However, if Intel could test the CPUs that failed the 933MHz test without a multiplier lock, they would have found that 70% of that 30% (sorry for the confusion ) will run guaranteed at 800Mhz, and 30% at 750Mhz.

So because of the manufacturing process (locking before determining final speed), they relatively loose a lot of higher MHz.

As you can probably imagine, I find it unlikely that Intel would lock the multiplier before determining the final speed. But there also is a problem if they don't: that would mean they can 'burn fuses' or whatever it is called, that will change the multiplier of the CPU after it is tested for a certain speed. If that would be the case, then why use the same CPUid for the 100Mhz FSB as the 133Mhz FSB 'version' (read: multiplier) of a CPU?

Some other possibilities: Intel can not change the core after producing the stepper, but can 'lock' the multiplier by doing that in the CPU packaging. But even in this case, why wouldn't they be able to put the CPUID in the packaging (thus being able to use a different CPUID for different FSB)?

And what about Celeron-2 chips? are the CPUIDs of the P3 and Celeron-2 chips with the same stepping and multiplier the same?

Final question: I myself bought a Coppermine-700E cC0 stepping 2 weeks ago, and I am running with at default voltage at 933Mhz on a BX mobo. I downloaded Intel Processor Identification Utility, and it reports a 700Mhz CPU overclocked to 933. How does it detects the overclocking if the CPUID would be the same? And would it report the same on a i815 chipset if everything else ran within specs?

A lot of questions here, but I hope someone might be able to answer them

 

[dZeus]

update: according to this Intel document: http://support.intel.com/design/PentiumIII/qit/update.pdf

all the CPUs in the same stepping have the same CPUid, so a CPUID says nothing about its speed. The question now is, how Intel's frequency id util finds the speed the CPU was supposed to run at

 

[pm]

I'm not familiar with this program - I'll download it and play with it tonight.

I was under the impression that Intel was fusing the tested frequency onto the part along with the serial number. I'd heard about this a long time ago from a friend who worked on coppermine, but I never heard any more about it. I presume this is how the program knows the correct frequency for operation.

Let me look into this tonight.

 

[pm]

So, I fired off an email yesterday and received the reply today. The certified frequency is fused into the chip (along with the serial number) after packaging and testing. The CPUID, FWIW, isn't - this is hardcoded with a ROM.

Thanks for pointing out the program, dZeus.

 

[SolTek]

so does this mean that the 800E (100) chips are really 1ghz chips (133) ????

 

[Dulanic]

PM is right here. The only thing I have to say about is lets take a Duron for example.... Now as we all know the yields on these things have to be damned near 100%. Now since they have basically no chips that wont run max speed (800Mhz) they test them.... 99.5% run at 800+ so what do they do? They just start spliting up all the chips into different bins... so even though your getting a 600 it probably could have been put in the 800 bin because the yields are so high the odds of finding a Duron 600 that couldnt have been sold as a 800 are very low.

 

[dZeus]

Soltek: absolutely NOT. If you carefully read the posts about frequency testing, than the only way a 800E can be a 1000EB is when the yields on the Coppermines are so good, that all the chips on a wafer, minus the ones rated for 550MHz - 750MHz would do 1GHz.... don't you think that is a bit unlikely?

Does Intel have a seperate product line for Celeron-2 chips, or are Celeron-2 chips nothing more than a Coppermine of which a part of the cache isn't working properly. And if so, does the Coppermine have some extra (spare) cache on board, in case a part of the cache doesn't work after manufacturing the CPU? Is there a similar way to deactivate the amount of not-working L2-cache blocks + working L2-cache blocks that is needed for 256KB cache, respectively 128KB cache? What about SMP on a Celeron-2... is that fused out as well?

Is it possible that some cores that have 256 (or more) working L2 cache, and despite that are rated as Celeron-2?

Multiplyers are fused out... is this (or: can it be) done after the packaging process? Are the fuses for the multiplyers chosen, in such a way, that if you blow more fuses, the multiplyer will increase or decrease? Why hasn't anyone else done research into this? It can be a very usefull tool for overclocking (for chips that have more or less headroom).

hmm... not really easy answerable questions... I don't think anyone can (and in case can, won't) answer these questions

 

[pm]

Ok, bear in mind that I err on the side of caution when talking about my company's flow, so if I skip questions, they are ones that I'm worried about talking about, and I'm skipping them on purpose.



<< ...does the Coppermine have some extra (spare) cache on board, in case a part of the cache doesn't work after manufacturing the CPU? >>


Yes, there are redundant L2 cache lines in Coppermine parts, so that if a cache line is found to fail, a secondary line can be enabled. This is standard on most CPUs made in the last 5 years or so.


<< Multiplyers are fused out... is this (or: can it be) done after the packaging process? >>


Yes, it is done after packaging, burn-in and testing.


<< Are the fuses for the multiplyers chosen, in such a way, that if you blow more fuses, the multiplyer will increase or decrease? >>


I personally have no idea. Intel is not disclosing any information about the fuses or the fuse circuitry - or the multiplier circuitry for that matter - due concerns about remarking.

 

[IaPuP]

I'll fill in the questions that pm can't answer (because of his job).


Yes, the PIII and celeron are on the same die. They have the same die size and same transistor count.

Intel merely burns out the SMP capabilites and half of the functioning cache lines.

It is possible that many Celeron II chips have a full working 256KB of L2 cache, however it is also possible that Intel splits out the broken Coppermines and removes non-functioning cache sections by disabling half the cache lines.

probably some of both- it is economical to save &quot;damaged&quot; coppermines but when Celeron demand goes up they can take working coppermines and burn half their L2 out for economic reasons.

I don't believe there is a way to re-enable the cache since it is probably burned in either by a laser cut or more likely by an internal fuse which is burned out after testing.

Intel has taken a stand in denying this information but it is very obvious when looking at the CPU die. Note that with AMD chips, the Duron is visibly smaller than TBird because they are fabbed seperately.

Eric

 

[IaPuP]

Here is my response to the author article linked here:

Let me preface by saying I'm a strong proponent of overclocking. I authored most of the OC Tips at Bxboards.com and I frequent the forums around to help out fellow OCers.

I'm also a computer engineer and have to correct some small mistakes that could be VERY MISLEADING to the average reader.

You said:

In chip manufacturing, you will encounter a term known as the Yield Curve. The Yield Curve is expressed as a percentage. When Intel say their yield is 80%, it means that 80% of the chips coming off the line can run at their current maximum speed. As the manufacturing process improves over time, the Yield Curve gets higher and higher until nearly all the chips coming off the line can run at maximum speed.

Ok- mostly correct but there is one VERY IMPORTANT flaw in this. The yeild percentage defines the number of chips that can run above the MINIMUM speed. This means that if the yeild is 80% that 20% of the CPUs on the wafer will not run at a sufficient speed to be retailable- they are thrown out. the remaining chips fall on the &quot;yeild curve&quot; as you have called it. Depending on manufacturing process, this is generally a &quot;bell curve&quot; shape with the mean and median lying at the center of the curve.

For Intel's current process, I would have to estimate that this mean/median line for the cB0 core would be about 800MHz. I think their process is slightly skewed to the higher end but with the yeilds falling off sharply within 3-sigma (sigma being about 115MHz - probably making a range from 600 to 1.1Ghz) with a tiny number falling outside 3-sigma. This means that approximately 50% of chips would not pass Intel's tests beyond 800MHz.

Some 600mhz processors may possibly be rated to higher than 600 but it is much less likely that they will be fully rated to 800. I would make NO CLAIMS that all 600MHz processors will run 800. Keep in mind that Intel rates its chips conservatively to account for variations in environmental conditions and transistor stability. Intel rates its chips at 600Mhz for a reason. Just because a chip is on the same manufacturing process and on the same stepping, there are still variations on the maximum rated speed. Buying a big cooler can decrease the maximum temperature reached- that can help, so can increasing voltage. That can give the 'impression' of having a stable chip, but your chip's lifespan and margin for error wil be greatly decreased with a higher voltage.

I think you need to tell your friend that his 800eb is actually LIKELY to be much better than a 600e, however it is POSSIBLE that SOME 600e chips would pass Intel's 800mhz spec.

Technically speaking, your article is false and mis-informs the readers. While some of Intel's &quot;ratings&quot; are determined by economics, much of their manufacturing is determined by actual yeilds and I felt you needed correcting.

 

[pm]

Actually, FWIW, the term yield refers to working parts per wafer. It is a ratio of the number of working chips on wafer to the total number of chips on the wafer. So 100% yield means that every possible chip on a wafer is working correctly.

Bin split is the number of chips which work at a various frequency bin. So a 100% bin split to 500MHz means that 100% of all the working chips on a wafer run at or faster than 500MHz. Yield and bin split are two different (although related) things.



<< Intel has taken a stand in denying this information but it is very obvious when looking at the CPU die. >>


I'm not sure what the corporate &quot;stand&quot; is on the subject. I know that I value avoiding what I call &quot;negative management attention&quot;, and so I avoid answering topics that may come back to haunt me when my raise is being calculated.

 

[IaPuP]

Yeah. yeild and bin slpit are different.

I guess I was meaning that when a chip does not work at a specified frequency it is thrown out. I was under the impression that chips thrown out because of low operating frequency were essentially considered &quot;non functional&quot; and thus were not considered in yeild calculations.

That's why I stated 'minimum frequency'.

Am I incorrect?

Even chips that yeild significantly below minimum tolerance but still function at some frequency are considered positive yeild?

I'd love it if you cleared this up.

Eric

Edit: Intel corporate has vacillated between denying that the Celeron II and PIII are off the same line and refusing to answer even though it is quite obvious that they are.

 

[pm]

<pm reads the post again>

Yeah, I guess I read what you wrote incorrectly. What threw me off was the talk of &quot;yield curves&quot; - the original author was using the term incorrectly, and since you picked up on it, IaPuP, I thought maybe you didn't understand the difference too. I re-read what you wrote and you are right.

Still, yield is definitely not a curve. It's pass or fail - and since there would only be two points on whichever axis you used to define yield, it wouldn't make a very nice curve at all. Frequency distribution on a normal part is usually a normal/Guassian distribution as you described (I've seen bimodal distributions, but this is a bad sign of a parts health). This is not a comment directed at you, but just to elaborate on the original article.

Do you work in the industry, IaPup? You definitely know what you are talking about.



<< Intel corporate has vacillated between denying that the Celeron II and PIII are off the same line and refusing to answer even though it is quite obvious that they are. >>



No comments here. I have a family, I like my job, and I like getting big raises every year. I'm not going to elaborate on this issue just in case we are on the waning side of the vacillation.