A CPU Manufacturing Question

[jadinolf]

I've been wanting to ask this question for a long time, so here goes.

To begin with, I spent most of my working life in semiconductor manufacturing BUT old technology, namely zeners and rectifiers. In rectifiers we shot for our highest voltage- 600 volts and then used the fallout for lower voltage rectifiers. With zeners we targeted for the customer required voltage. These gave us the typical bell shaped curve, with the customer's diode hopefully heavily in the middle of the curve. The zeners on either side of the curve would usually be saleable as well.

Now my question: is CPU manufacturing pretty much the same (namely using fallout) or can the target speed be hit quite accurately?

 

[rimshaker]

Yes, i believe it's called binning, where they sort out the cores into speed grades. Yield rates and statistics are common to all semiconductor fabrication lines.

 

[jadinolf]

Thanks, rimshaker. I suspected that it was the case but it's nice to have it confirmed.

I appreciate the fact that you responded.

 

[bizmark]

not only are they binned, they are sometimes artificially binned, or so it would seem. Like the new P4's that Intel is selling as 1.6GHz seem to run pretty predictably at higher than 2GHz. So even though their process is good and produces high quality yields, Intel can somehow make more money selling them as 1.6s than 2.2s.

I'm sure someone with some actual knowledge of this could give more detail. this is just from what I've read. But as far as your original question goes, yes, all processors in the same product line come from the same fabs, and maybe two chips that were right next to each other on the same silicon wafer would be rated at different speeds based on individual testing. It's this process that makes overclocking worthwhile, as it means that there's usually a little bit of extra performance headroom for a chip.

 

[Eskimo]

<<

Now my question: is CPU manufacturing pretty much the same (namely using fallout) or can the target speed be hit quite accurately? >>



Well if we've done our part right as engineers we can generally hit our target distribution set according to strategic planning. There is some distribution to the speed grades but unlike the products you work with we are constantly pushing the edge of our speed envelope. Often this means that if you are looking a plot of yield versus speed it will appear as something of a cliff as you approach higher speeds. After a certain speed there is a dramatic falloff. Our challenge is to push this curve outward with process and design tweaks in order to enable further speed grades. Any processor will run slower, but faster is our problem. Sometimes we push the envelope too hard and have a fall off, that's when management and marketing start yelling .

 

[Wango]

<< not only are they binned, they are sometimes artificially binned, or so it would seem. Like the new P4's that Intel is selling as 1.6GHz seem to run pretty predictably at higher than 2GHz. So even though their process is good and produces high quality yields, Intel can somehow make more money selling them as 1.6s than 2.2s. >>



It's a form of price discrimination that Intel uses to increase profits by selling the same chip at different price points. The vast majority of people don't overclock, so they'll pay extra for the high-rated chips. By making lower-rated chips available for less, Intel brings more consumers into the market and thus increases profits.

Economics makes the world go round!

 

[pm]

<< Now my question: is CPU manufacturing pretty much the same (namely using fallout) or can the target speed be hit quite accurately? >>

It's pretty much the same.


<< not only are they binned, they are sometimes artificially binned, or so it would seem. Like the new P4's that Intel is selling as 1.6GHz seem to run pretty predictably at higher than 2GHz. So even though their process is good and produces high quality yields, Intel can somehow make more money selling them as 1.6s than 2.2s. >>

Intel's bins can usually be accurately gauged by simply looking at the cost per MHz on a graph. As soon as you do this, you'll notice that you can get more MHz (or GHz in this case) without spending any money at all up to a point. At that point, the price starts to rise.

Prices from Newegg.com

Pentium 4 1.5GHz $130
Pentium 4 1.6GHz $130
Pentium 4 1.7GHz $135
Pentium 4 1.8GHz $177
Pentium 4 1.9GHz $207
Pentium 4 2.0GHz $249
Pentium 4 2.2GHz $378
Pentium 4 2.4GHz $530

So somewhere around 1.8GHz the price starts to rise and somewhere around 2GHz it starts rise quickly. This shows the Guassian distribution of frequencies through economics.

People will say, as wbwither and Wango have said in this thread, that really the bin curve is much higher than this because overclockers can achieve high gains with low frequency parts. One point to keep in mind is that Intel (and AMD, and everyone else who designs CMOS semiconductors) designates a CPU to operate at the specified frequency over the life of the part under worst case operating conditions. So the CPU should work at the highest spec'd temperature, under the lowest spec'd core voltage, across a range of motherboards of varying quality, on every possible program that can be run, for the lifetime of the processor. I see people in various threads saying that Prime95 is not a valid test of stability because it will fail CPU's that are otherwise fine. I see people talk about what power supply to get and a lot of threads about what heatsink to get and how to cool your case effectively. I would like to see anyone take their amazing overclocked 1.6 CPU's and run them at 70C with the core voltage turned down to 1.3V (the minimum spec), run Prime95 on a no-name motherboard from some manufacturer that no one has ever heard of and then tell me that it will continuously operate like this for seven years (CMOS semiconductors slow down with time).

I'm not arguing whether or not parts are binned below their rated spec. I really don't know either way for a fact, and, being an employee of Intel I'd hesitate to speculate. But what I am saying is that people are using overclocking results to back up their speculations without considering that these specific systems are carefully optimized. I'd suggest run them at the worst case spec and then substract out some for the eventual slow-down of the electronics with time and then check against the rated speed.

Patrick Mahoney
Microprocessor Design Engineer
Intel Corp.

 

[RSMemphis]

I see people in various threads saying that Prime95 is not a valid test of stability because it will fail CPU's that are otherwise fine. I see people talk about what power supply to get and a lot of threads about what heatsink to get and how to cool your case effectively. I would like to see anyone take their amazing overclocked 1.6 CPU's and run them at 70C with the core voltage turned down to 1.3V (the minimum spec), run Prime95 on a no-name motherboard from some manufacturer that no one has ever heard of and then tell me that it will continuously operate like this for seven years (CMOS semiconductors slow down with time).

I'd like to see that even for a 2.4 GHz CPU at stock speed. Well, maybe it *does* work.

 

[jadinolf]

Thank you all for the comments.

Nice to know that there are so many knowledgeable people around.

 

[ttn1]

Just got back last week from a precision engineering meeting. There is a lot of work going on in the semiconductor industry as they slowly shrink the feature size on chips.

As far as i know, a very educated guess is made for the center of the normal distribution for a specific design. Then some run-offs are performed to get a baseline. Since your using a lens system to expose an image of the CPUs on a chunk of silicon, you have the abberations of the lens system to worry about. Along with a lot of other process variables. Usually the first batch of CPUs you make have a normal distribution and you sort tham into speed ratings. Then you tweek the process a bit to get better yields at higher clock speeds. Most of the stuff they do to tweek is the same as in most other manufacturing. You measure the product, figure out the errors and then software correct to get a better product.

Since the speed ratings are usually decided upon long before the first mass produced silicon goes out the door, sometimes your process is just too good. This usually happens when CPUs manufacturing goes through a die shrink. Then you need to fill out you required speed categories whether they are better than that or not. Like has been said before it is economics and forward projections.

 

[jmitchell]

anyone interested might find some additional/informative info in this archived post - especially those comments posted by PM

Old long post on similar topic

 

[jadinolf]

Interestingly we made far more profit on fallout than we did on the target because the fallout was usually ordered in smaller quantities.