How do companies like Intel or AMD test their silicons???

[mutex]

How do big companies like Intel test the chips (CPU/chipset)? How do they make sure all the transistors/interconnects can operate properly at desire frequency??????

 

[CTho9305]

One of the classes I took last semester covered this... at least the basics.

You first find all possible faults (e.g. a line is stuck at 1, two lines are shorted, etc.), then collapse these down to eliminate things that are functionally the same (an inverter input stuck at one is indistinguishable from its output being stuck at zero). Then you have to come up with a test vector that would have its output affected by a given fault (for the inverter example, the text vector would just be puttying a 0 on the input, and checking the output). Once you have a bunch of vectors, you'll find that many of them cover multiple possible faults, so you can combine them to get the number of required tests down. Then, when you get the manufactured part, you apply the test vector to the inputs, and watch the outputs.

On a circuit as complicated as a processor, it would be impossible to test every part from just the pins, so ALL the latches (or just registers? I'm not sure) in the chip are tied together and can be put into a "scan" mode, which lets you treat them like a giant shift register. This way, you can observe the internal state of the chip at every point. You scan in the values you want, clock the chip, then scan the registers back out and see if they are all correct. This lets you only worry about testing a bunch of circuits the size of the path between registers instead of one gigantic circuit.

edit: To make sure the part works at a given frequency, you run it at that frequency when doing the tests .

 

[Lynx516]

Well I have just finished doing some work at ARM and was shown how they do this. They use something called PFL (I think cannot remember). This is a program into which you input the net list that has all the gate lenghts and widths in aswell as slew rates rising edge rates, temperature and what type of silicon. This is then run and it takes a while (ARM7 CPU took 12 hrs). Then it gives you the speed it would work on decent silicon at the voltage and temp specified.

 

[sgtroyer]

Mutex, can you clarify your question? Are you asking:

1. During design, before silicon, how does Intel know how fast a processor will run?

2. After fabrication, how do they verify that the part functions correctly and fast enough?

3. Something else.

 

[pm]

As sgtroyer pointed out, there are several parts to this question:

1. Checking during design that the design will run correctly at the design target frequency.
2. Checking to make sure that the manufactured chip will run correctly at the labelled frequency.
3. Checking that all of the transistors on the chip work correctly.

#3 has been covered by Chris (CTho).
#1 has been answered by Lynx.

I tend to read the question as #2, however. Mutex, as sgtroyer requested, can you clarify your question?

Here's a somewhat technical link to an article at Intel giving an overview of the microprocessor manufacturing testing flow and future directions.

 

[f95toli]

Another thing worth remembering is that in most chips there is some redudancy, important parts are built in such a way that they don't fail if a single transistor is not working according to specs.
I am not sure but I think there might even be "spare paths" in some chips, if a path is complettey "blocked" due to a faulty part there is an alternate signal-path that can do the job.