The silicon lottery, what causes it?

[Maximilian]

Why are some chips better clockers than others? For example with Sandy Bridge 2500k's they have the exact same design, the same transistor layout, the same number of transistors, what causes some chips to be able to do say 4.8ghz at 1.25v and others need 1.4v, a few may not even be able to reach that speed at all. Is it some impurity in the silicon that causes these differences or what?

I heard someone say that chips closer to the center of the wafer are better, dunno why or if that's true or not.

 

[Ferzerp]

The wafers can't be completely uniform, the manufacturing process can't be 100% uniform. You have to realize we're dealing with some of the smallest scale things in the world. 28nm is 28 *billionths* of a meter. The center to center distance of features of a processor have only around 56 silicon atoms of width between them at that point. (that's center to center, the space in between edge to edge would be fewer than 56 atoms). 22nm is only around 44 Si atoms.

Things are so small that an extra atom in the wrong place can produce different characteristics. It's amazing that we get things as consistent as we do.


edit: does anyone know what the feature width for Intel's 22 nm process is?


edit2: combine this with how many amperes of current are flowing through the chips (50+ amperes easily for higher end desktop chips under middling to high load). How can we *not* have variations?

 

[veri745]

There are lots of parametric differences between one chip and another, one of the most important of which is SIDD, or static current, also referred to as leakage current. If you just power up a chip, don't power-gate anything, and leave all the clocks off, it will draw a certain amount of current.

Higher SIDD chips are hotter, but depending on the specific process, the transistors also tend to be able to switch faster, so you get a better Fmax.

Differences in SIDD are primarily caused by slight differences during fabrication, which can lead to things like varying gate thicknesses or other differences in dimensions of the individual transistors. Differing transistor dimensions causes the performance of each transistor to vary slightly.

 

[Bman123]

No 2 things in life are exactly the same, end of story

 

[BrightCandle]

Its so many factors that go into determining the clock speed and at what voltage. In short there are tolerances in the manufacturing process just like anything else. The size and shape of the transistor itself, the metal layers, the alignment of the layers etc all have variations which combine to produce either golden chips or ones that never even pass Quality assurance.

Its not uncommon in the beginning of a new process or design for more than half of chips to never pass quality assurance at all. The bigger a chip is the more likely a faulty will occur within the chip. Due to the tiny feature size and high probability of failure within a single unit we see today most chips are designed to tolerate faults by turning off cores or cache so that the majority of the chip can still be used when only one part doesn't work. It only takes a single faulty transistor to break a CPU, its got billions of them.

The centre of the wafer benefit is about the lithography step as the light is less distorted compared to the edges. It was the case 15 years ago but they might do it very differently now and there may be little to no benefit in the centre now.

 

[KingFatty]

Growing Si crystals, how does it work?

 

[veri745]

Growing Si crystals, how does it work?

http://en.wikipedia.org/wiki/Czochralski_process

 

[Leyawiin]

The motherboard its dropped in can have an effect as well.