How are processors clock speeds determined?

[Special K]

I suppose someone would have to fill me in on some of the details on how processors are made for me to understand the answer to this question, but basically what I want to know is how does Intel, AMD, or other CPU manufacturers clock their processors? I know the question is worded kind of awkwardly, but let me try and explain. Let's say that a wafer of coppermine processors is produced (aren't they called wafers before the individual processors have been cut apart?) Is the wafer of processors designed to run at a certain speed to begin with (866 Mhz, for instance), or do each of the processors have to be tested to see what speed they will safely attain, and then are marked as that speed of a processor? How do they test the processors while they are still on the wafer (for speed, defects, etc.)? If each processor is tested individually, do the speed ratings follow a normal distribution (I think that is the correct term)? What factors during manufacturing affect a processor's chance of being able to attain a higher clock speed?

 

[GoSharks]

each processor is tested and put in their respective categories of how fast they can stablely run

 

[Dulanic]

They actually go through and test them and then bin them into different speeds. Now Im sure some of the Intel employees on the forum can go into more detail, but I can answer another question. You asked what makes them attain a higher clock speed, that would be their yield. If lets say AMD creates a a bunch of CPUs and they have all the 1.4Ghz CPUs they want to make already they are going to put CPUs that really are capable of 1.4 and label them 1.33 or 1.2 or 1.1 etc... So good yeilds are good in many ways, both for overclockers and for price. Oh BTW yes they are called waffers.

 

[Noriaki]

It's called "speed binning", or something to that effect.

They make bins according to clock speed, and they test the CPUs to see how high they will reliably run.

<EDIT>Bah to slow Jon, to slow </EDIT>

Also market demands can call for processors to be marked and sold below what they can actually run at.

For example, back near the beginning of the Athlons life, there was a batch of 500 and 550s that had 750 cores in them. This is because the Athlon yields were very good, and 750s were coming out, but there was demand for 500s and 550s so AMD produced a bunch with the higher tested cores. Not a big loss to AMD becuase most people don't overclock (I'm sure Intel has done the same thing many times, I just don't know any specific examples).



<< do the speed ratings follow a normal distribution (I think that is the correct term)? What factors during manufacturing affect a processor's chance of being able to attain a higher clock speed? >>

You'll have to look beyond my knowledge for the answer to those to, but we do have at least one Intel fab technician on hand I don't think those would be classified information.

 

[jonnyGURU]

Right on both accounts. Speed binning is done based on the CPU's capabilities and market demand. They are not tested as wafers, but as processors prior to having their multipliers marked, speed etched in the die, etc.

I would imagine that all of the CPUs from a given wafer should all be able to accomodate the same FSB, though, nevermind end result clock speed.

 

[Sohcan]

<< If each processor is tested individually, do the speed ratings follow a normal distribution (I think that is the correct term)? What factors during manufacturing affect a processor's chance of being able to attain a higher clock speed? >>

Yep, the yield at different speed ratings follow a normal distribution. During design, the target clock rate for a CPU will be determined, based on the speed of the transistors used and the critical path length. During production, most CPUs will be at this target speed, but because of anomalies that can occur, some dies on the same wafer may be able to reach different speeds (I don't know exactly why some transistors on the same wafer can reach different speeds....talk to Wingznut ). Thus, for a high-volume CPU like those produced by Intel and AMD, there are a wide range of speeds available, from the cheeper, more common lower speeds to the lower-yield (and thus more expensive) higher speeds. Check out page 10 of this PowerPoint presentation. (Watch out, it's around 500K....It's by the guy who started ArtX, who made the graphics system for Nintendo's Gamecube...he came to talk to a class that I took last semester).

 

[CTho9305]

johnny - why would the proc care about fsb? shouldn't only its internal speed matter? or are there some transistors that only work on the fsb speed, and might work at 200 but not 266 mhz?

i'm sure wingznut [pez] will provide some excellent answers...

 

[Dulanic]

<< I would imagine that all of the CPUs from a given wafer should all be able to accomodate the same FSB, though, nevermind end result clock speed. >>



That is true, FSB has no direct effect on the CPU. It does have the indirect effect of raising clock speed if you leave the multiplier the same. Its the motherboard that has to handle the FSB.

 

[jonnyGURU]

<< That is true, FSB has no direct effect on the CPU. It does have the indirect effect of raising clock speed if you leave the multiplier the same. Its the motherboard that has to handle the FSB. >>



Gotta gotcha for you then....

Remmeber the K6-2 300? How come then, were there some of them that AMD deemed INCAPABLE of running at a 100 MHz FSB, and thus &quot;binned&quot; them as 66 MHz CPUs running at 66 MHz x 4.5 instead of 100 x 3. Remember, back then, CPUs were NOT multiplier locked, yet many of the 300-66 CPUs simply WOULD NOT run stable when run at a 100 MHz FSB.

 

[DeeK]

<< why would the proc care about fsb? >>



It doesn't. The CPU chip contains an on-board PLL to boost the FSB frequency up to the CPU frequency. Which multiplier lines on the PLL are attached depend on what FSB and CPU frequency the CPU is supposed to run on. A CPU cannot tell the difference between running on 8x100MHz or 6x133MHz, since either way it's running at 800MHz.

 

[CTho9305]

Deek - your answer conflicts with the one I got in the super-incredible-overclocking-motherboard thread... in there, I was given the impression the actual processor has no PLL and the motherboard handles multiplying the fsb.

edit: jonny - are you sure they weren't multiplier locked? maybe some pins weren't connected, like on pentium 133s that dont do over a 2x multiplier because of disconnected pins

 

[DeeK]

<<
Gotta gotcha for you then....

Remmeber the K6-2 300? How come then, were there some of them that AMD deemed INCAPABLE of running at a 100 MHz FSB, and thus &quot;binned&quot; them as 66 MHz CPUs running at 66 MHz x 4.5 instead of 100 x 3. Remember, back then, CPUs were NOT multiplier locked, yet many of the 300-66 CPUs simply WOULD NOT run stable when run at a 100 MHz FSB. >>



Hmm. Maybe there were problems with either the PLL or the pads on the chip. I've heard many times that quality pads (the places where the I/O pins are attached to the silicon) are a real PITA to design.

 

[DeeK]

<< Deek - your answer conflicts with the one I got in the super-incredible-overclocking-motherboard thread... in there, I was given the impression the actual processor has no PLL and the motherboard handles multiplying the fsb. >>



If the PLL were not on the CPU, there'd be no way for the CPU maker to lock the multiplier.

 

[Wingznut]

<< What factors during manufacturing affect a processor's chance of being able to attain a higher clock speed? >>



While manufacturing wafers is a precise exercise, it certainly isn't an exact science. There are literally hundreds of ways a process can vary, but I'll give you one from Lithography.

To lay patterns (for the circuits) on the wafer, photoresist must be applied (spun) to the wafer and baked. Light is shown through a mask (reticle) onto the resist, which turns it into an acid. So, whatever part of the resist which is blocked from the light (by the reticle) stays hard, and whatever is exposed to the light becomes an acid. The acid is then developed (rinsed) away, leaving a pattern of photoresist. After etching, polishing, plating, etc... These patterns become the circuits. (See here, for more info.)

So, many things during spin, exposure, and develop can change the width. Exposure time, light intensity, and resist thickness. We'll talk about resist thickness.

To apply the resist, the wafer is spun at several thousand rpm, and resist is added to the center of the wafer and spun outwards. The resist needs to be a specific thickness to achieve the target line width. Folks, we are talking about thickness measured in angstroms (1 angstrom = .0001 micron). Some layers require 5000 angstroms of thickness, other layers need much, much, much less. As you can imagine, there will always be some variance in resist thickness, across the wafer. And as a result, some variance of circuits on the wafer.

So... I could go on and on about the other areas, but that's one example for you.

 

[jonnyGURU]

CTho9305: Positive. Intel Pioneered the multiplier lock on the Slot 1 platform. AMD didn't play that until the Athlon came along.

DeeK: Could be. Don't know. All I know is if the heat plate was designating the chip as a -66 chip, it sure as hell didn't do 100 MHz FSB.

 

[Wingznut]

And yes, it's very possible for a cpu to run at say... 10x100 but unable to run at 7.5x133.

I don't think it happens very often any more. But yeah, it is very possible.

 

[Evadman]

It is very possable to get a CPU that will run at 66mhz bus and not at 100. ( or 100 not 133 ) When you up the bus and lower the multiplyer, then the cpu gets fed more information in the same amount of time. Also, the CPU can output more information than be carried on the bus at one time, so the 133 frequency helps the CPU send that data where it needs to go. Theoreticly, a CPU running at 7.5 x 133 can output and recieve %33 more data in the same amount of time that a 10 x 100 can do. But the CPU has to have the extra processing power to use it. The theoretical 33% actually becomes about 10% in real life. The big thing is getting the data to and from the registers. Thats where the 33% boost comes in, but parasidic lag ( the other components slowing down the CPU ) drags the CPU performance down.

 

[Degenerate]

About this bining. So would Intel make P4's from 1.3 to 2.0 (or where ever its going) all at the same time? and similarly the P3s coppermines from 600 to 1.0ghz too?



<< To apply the resist, the wafer is spun at several thousand rpm, and resist is added to the center of the wafer and spun outwards. The resist needs to be a specific thickness to achieve the target line width. Folks, we are talking about thickness measured in angstroms (1 angstrom = .0001 micron). Some layers require 5000 angstroms of thickness, other layers need much, much, much less. As you can imagine, there will always be some variance in resist thickness, across the wafer. And as a result, some variance of circuits on the wafer. >>



by spining, if like a little bit of resist didnt go on a fraction of the chip, would it be useless? If just one of the resistors was not complete, would the CPU be useless?

 

[Wingznut]

Yes.

And yes, it would render the die inoperable. Although there are several metrology (inspection and taking measurements after Spin/Expose/Develop) and analytical steps to ensure that those such things are kept to a minimum.

There are literally thousands of microscopic things that can go wrong and cause a die to be dead. It's actually very amazing that so many of them work so well.

 

[dowxp]

i cant wait to learn all of this in college! come on computer engineering major. give me life.

 

[ST4RCUTTER]

About this bining. So would Intel make P4's from 1.3 to 2.0 (or where ever its going) all at the same time? and similarly the P3s coppermines from 600 to 1.0ghz too?-- Degenerate

Yes.-- Wingznut

Wow! So there could be as much as a 700MHz difference between the best die on the wafer and the worst(that is still able to be sold)? That's quite a range of values.

 

[Ladies Man]

there could be
could being the key word

I doubt that they vary that much from the same wafer

 

[Wingznut]

Oh, I guess I didn't take that question literally enough. I didn't mean to imply that there is that much range on one wafer.

 

[ST4RCUTTER]

Whew, for a second I thought, &quot;Man he isn't kidding&quot; when he said,

&quot;There are literally thousands of microscopic things that can go wrong and cause a die to be dead. It's actually very amazing that so many of them work so well. &quot;


That would, after all, explain such disparity.

 

[Wingznut]

Well, I wasn't kidding in the least about that part.