HN Explorer

Ask HN: Will two similar processors process data in the same exact way?

5 points by RevRal ↗ HN
Friend and I got into a debate, the simple question: will two computer processors, one finished manufacturing right after the other, compute the same data in the same exact way, down to each transistor?

Opinions? Facts?

I appreciate the help here.

10 comments

[ 5.1 ms ] story [ 37.7 ms ] thread
kasey_junk
Well you didn't mention what memory bus the processors would be installed in...
zerohp
No, they don't always achieve the same results but there is a lot of engineering dedicated to correct the errors.

You can find academic papers on the subject of GPGPU soft errors around 5 years ago when people started doing general purpose computation on graphics processors. GPU's did not dedicate as much logic to error correction because it doesn't matter as much for real time graphics.

chrisBob
No, many of the current Intel processors come off of the same fab and are intended to be identical, but the devices that test worse become the lower spec versions. A quad core i7 with a core that performs significant'y below the others becomes a dual core model. When you are making 1.7 Billion of anything there are going to be some errors, but a CPU is designed to compensate for this.
caw
It was my interpretation that "binning" is because of the manufacturing and imperfections in the etching or underlaying wafer cause certain chips to not perform up to spec. It's not because there's 1.7 Billion of something, but rather because you're drawing at 22nm. The way a processor does math is identical in every single processor, but processor A may result in more heat discharge or higher voltages required than processor B at the same clock rate.

Where you get into "it may not compute identically" is the revisions. At Intel at least, going from A0 to B0 means the transistor layout on the wafer changed. Going from B1 to B2 means the wiring layout changed (though the underlaying metal layer remained the same). Again that's a "technically", whether that's followed to a T is something I'm not sure about. Sometimes features are added going between metal revisions.

chrisBob
Thanks. It looks like I was wrong. I understood that there was some redundancy build into chips, but that appears to be false. After some more research I think you are right that only the voltage/clock rating vary die to die, and that the transistors must be 100% working.

That means that the answer to the OP's question should be that if the data is stored in the same memory location then two chips in adjacent server racks will use the same transistors to do the same math.

caw
Transistors may not need to be 100% working on 100% of the die to be shipped. If the fault is contained within a subsection, the manufacturer can set fuses to disable portions of the chip. This is generally used to distinguish features between product lines. Processor A could have VT-D failed in manufacturing, so it's fused off and sold as a cheap processor. Processor B could be fully working, but Intel needs to sell more cheap processors, so they intentionally disable VT-D on B. Sony did something similar with shipping Playstation 3s since they couldn't get the yields high enough initially. They disabled 1 SPU of each processor since they could reliably get 7 working ones in manufacturing. Those 8 SPUs are more interchangeable though, they're more like graphics chips in that regard than general x86 CPUs.
GotAnyMegadeth
There is often redundancy built into the larger internal memories of chips.