"When Overclocking Is Not Overclocking"

[NOX]

Hehehe...for some of you who didn't know this already. It's a good read.

 

[praetor1]

thanks...i did'nt know that.

 

[osage]

Yes, good read NOX. thanks

 

[pm]

This article contains a couple of errors. Yield refers to the ratio of working chips on a wafer to the total number of possible die. It has little to do with frequency aside from the fact that a die will yield ok if it works at the minimum possible frequency. The term the author should have used is "bin split".

Then there's the fact that the basic premise of the article is incorrect. The author clearly states that there is absolutely no difference between a Pentium 600E and a Pentium 800EB and that running a 600E at 800MHz is not overclocking. This is wrong. There is one huge difference: Intel guarantees that the 800MHz part will run at 800MHz for 7 years. Intel has put the 800EB part into it's testers and tested it using a suite of tests that have been built up over the last two decades and using a complex statistical formula have determined that this part will run reliably within spec for 7 years. The author merely supposes that Intel's bin split to 800MHz is 100% all the time. This is definitely not the case, as he even admits in the article. Bin split is a statistical distribution. That means that some will very slow, so you can't say whether or not a 600EB is one of those slow parts. The only source who can is the manufacturer - who have extensive experience, have expensive equipment and who are betting their reputation and a fair amount of money on being right.

Anyone who works as a computing professional would not want to put their paycheck on the line that the processor will overclock reliably for a long period of time. I don't know anyone who works at an IT professional who would say to their boss, "we can say $200 (15%) off of the cost of this Dell system that we are going to put our payroll accounting system by getting a 600MHz processor and tweaking the speed up to 800MHz. It will probably work fine and we can say a little money". No one who values their job would do that. So there is big fundamental difference between a 600E and an 800EB.

 

[pm]

To add one more point - I'm not saying "don't overclock"... my attitude is that it's your processor, you paid for it, you can do what you want. My problem is that the author is basically saying that Intel overclocks it's processors - and this is fundamentally incorrect because Intel is the company who determines what the speed is and who has the accountability if anything goes wrong. So, if Intel says it's not overclocked, then it's not - because they have tested it and said that it's not.

 

[NOX]

<< because Intel is the company who determines what the speed >>


Would you mind explaining the process in which Intel will determine the speed?

 

[NOX]

I remember I while back when it was also speculated the PIII 600 was an overclocked PIII 450 (many reports on the web), because the core voltage was 2.2. Unlike like the rest of its family of CPU's which were all running at 2.0. Change the multiplier and presto!

[EDIT]

BTW, he was also suggesting that Intel underclocks, though I personaly would not call it that. I would say Intel is setting the CPU to run at a certain speed.

 

[urbantechie]

<< I remember I while back when it was also speculated the PIII 600 was an overclocked PIII 450 >>


I think so too. I'm runing one myself, runs a bit warm. But I love the cache.

 

[AndyHui]

Pentium II 300 MHz SL2W8 and SL2YK?

Both set to run at 300, both overclocked perfectly to 450MHz, many even had the cache rated to run at 225MHz.

Well?

 

[NOX]

<< Well? >>


Me?

 

[compuwiz1]

NOX, read between the lines buddy. We don't want to put our friend PM in a compromising position.

 

[AndyHui]

Anyone.

Those two processors seem to support NOX's case.

 

[wj]

I've got a Celeron 566MHz @ 850Mhz I think that would be OverClocking. LOL. I did here that Intel was going to come out with the Celeron at 100MHz FSB, is that true.


wj

 

[NOX]

<< NOX, read between the lines buddy. We don't want to put our friend PM in a compromising position. >>


Opps, never though about that...hahaha. Thanks for bringing it to my attention!


<< I've got a Celeron 566MHz @ 850Mhz I think that would be OverClocking. >>


That's the point, only if you (the person) take a CPU sold to you at 600MHz, and set it to 800MHz with whatever means possible it?s considered overclocking. But if Intel or AMD does the same (which I?ll bet they do) it?s considered determining what speed to market/ship their processors at. Which by the way is what anyone in their right business mind would do.

 

[Stallion]

So if I read that correctly you could say the same thing for a 700E and a 933EB then right? As my 700 runs at 933 at default voltage and if I remember corectly it say 933EB when it boots up.

 

[NOX]

<< So if I read that correctly you could say the same thing for a 700E and a 933EB then right? As my 700 runs at 933 at default voltage and if I remember corectly it say 933EB when it boots up. >>


That's basically what the article is saying, and I find what you said to be even more proof.

 

[Navi]

Even AMD does the same thing as Intel. They run the dies for a speed based upon a statisical probability. Then, they test'm. Based upon their performance they are either stamped one way or another.

I would be willing to bet a small monkey that most of the low end Durons are all from the same die style. IE, they are probally all 700's and were chucked into a specific bin based upon performance. I would also be willing to bet that the GHz and up TBirds are from a 1100 die. They stamped 1000 when they first came out until they got the process perfected and moved their way up and yields improved.

The reason they do this is to save cost. Each die costs Intel or AMD millions &amp; millions of dollars to setup. They probally try to cover as much ground as they can. One die for several proc's is always a good thing to save you the consumer money and to put more profit in the investors hands.

My friend who worked at Motorola did the same thing when they were matching discrete transistors and diodes back in the day. Though, I couldn't imagine glaring at a curve tracer all day long. *cringe*

-Navi

 

[pm]

Well, actually, I had typed a big long reply (really) and my wife came home last night and shut her laptop off and zapped it. So rather than type it again, I just went to bed. I don't have a problem with posting some discussion of the testing flow, but it's not a simple short thing. So, it will be later today.

Look, I'm not arguing that different processor speeds aren't the same design - in many cases they are cut from the same wafer. If I did, I'd be lying. My point is simple, the author claims that a 600E and an 800EB are the same thing - that there's no difference between them... But there is, Intel has said that the 800EB will run at 800MHz at the rated voltage, using the rated heatsink, for 7 years. So there's a big difference.

It's like if I took two bungie cords off the assembly line, tested one thoroughly and said &quot;statistically most of our bungie cords work fine, we only have 20% failure on them. You'll probably be fine using the untested one.&quot; Which one would you choose to bungie jump with? (assuming that you are even willing to bungie jump) Are these two cords exactly identical in every respect? If you had to pay more for the tested one, the one that is guaranteed to work, rather than taking a risk with an untested one, is that a rip off? I honestly don't think so.

edit: fixed a bad typo... changed &quot;are&quot; to &quot;aren't&quot; in the 1st line, 2nd paragraph.

 

[spidey07]

Agree with PM, the chips do come from the same die but have one BIG exception. The procs have gone through their respective tests and are guranteed.

I'm sure demand and industry conditions dictate the &quot;bin splitting&quot; as well which is why we come across gems such as the cel 300a, cel566, and the 700e.

Although my two cents about overclocking - I overclock everything I can. Bumping a 700-933 on an 815e board is a nice PROCESSOR overclock. But I prefer the severe memory/AGP/Video clocking my BX gives me. Sure I've got a gig processor now, but memory is faster than Intel gig chip and does indeed perform better than the retail gig.

If I had my kudos I'd slap an 866 in her and rock at 155 cas2. (tested it once and WOW)

cheers,
spidey

 

[NOX]

<< . My point is simple, the author claims that a 600E and an 800EB are the same thing - that there's no difference between them... But there is, Intel has said that the 800EB will run at 800MHz at the rated voltage, using the rated heatsink, for 7 years. So there's a big difference. >>

I agree with you 100%.

 

[pm]

How testing works.

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.

There may be things that I've missed with this, or glossed over or got out of order. I'm a designer and the only time I get out to the test floor is when one of us designers screws and we have to figure out what we did wrong. Usually when this happens I'm too worried that I'm the one who screwed up to take detailed notes of the test flow. But this is the basic flow.

 

[Goosey]

I agree with the fact that Intel overclocks there chips. And also agree with the fact that a 600E and 800EB is probably the same chip.
Why you ask?

Intel has a very good yield up to a certain MHz. When you buy a
600MHz chip you are NOT buying a weaker chip. It simply has to do with having a broad product line and milking the consumer. We all know that to produce a 900MHz costs no more than a 600MHz. So why make a 600MHz, simply for profit!

Not all people can afford a 900MHz. So you take that 900MHz and clock it at 600MHz.

A good example has to do with the multiplier lock. If you are supplying consumers with a &quot;Weaker&quot; chip and give yourself a margin of 10% you should be fine. And a re-marker will not make any money off 10% difference.

Why do we see the higher end chips with high voltages, simple they are overlocked! And why is it that always the lower end chips overclock the best?? Because these chips were never produced to run at that level, they are just there for the consumer demand!

 

[pm]

Goosey: As I mentioned in the first post that I made, the distribution of processor speeds from a wafer is a gaussian distribution (also known as a normal distribution). This is a fact of the semiconductor industry and you can look it up any any book on microprocessor design (Hennessey and Patterson have a section on this, I can get you page numbers if you don't believe me). Some chips on a wafer run at 600MHz, some at 900MHz, and some at 1GHz - this is due to edge effects, polishing issues with the CMP machines, etc. etc. In any case, it is not the step function that you are describing.

As a process improves, the distribution starts to move higher... but then the prices on the low-end flatten out as well. The price differences between 500, 550, 600, 650, and 667MHz processors are all very close together - within 15-20% of each other.

The thing to ask yourself is: everyone charges increasingly more for higher speed grades. Intel, AMD, Transmeta, Motorola, IBM, Sun... every single microprocessor manufacturer does this. They did it before there were multiplier locks and they are still doing it. This practice of charging more for high speed grades has been going on practically since the IC was invented. Do you really believe that this is some form of conspiracy that everyone is participating in? Or is it a little more realistic (and a little less paranoid) to think that this is a reality of the semiconductor industry?

 

[qacwac]

Thanks for the article.

Anyone else think pm is a genius. I read some other stuff he wrote and I see why he's elite at 792. But thanks again.

 

[pm]

Goosey: I recommend reading pages 245-246 of Weste &amp; Eshraghian's &quot;Principles of CMOS VLSI Design&quot; for a basic intro to PVT variation which discuss the effect of process variations on yield and performance and shows a graph of a Gaussian distribution of parts from a fab. If you have a problem finding this book, I can fax you the pages.

Qacwac: thanks for the complement, but I'm hardly a genius (I wish... my job would be so much easier if I were a little smarter). I work in this stuff day in and out and I've been designing microprocessors for about 9 years now, so I know it fairly well. Once you get outside my field of expertise (microprocessors), then I have to bow to the real geniuses. (genii?)