what happens when a handfull of transistors on a monster chip are defective? do they throw out the die, or is the hardware smart enough to work around the defects?
pic related: 8 billion transistor, 683 mm2
It's been touched by bare human hands...the ENTIRE thing is worthless
binning
>>51767531
they still use the die, they just don't use the dud core and sell it as a cheaper model
>>51767531
i3s
>>51767550
>amd QA
>>51767531
They're being used for Skylake CPUs
>>51767550
It's not entirely infeasible that the entire wafer was shit.
They do sell whole wafers on ebay from time to time.
Some people like to hang them up like vinyls
>>51767531
>all that processing power
I bet it would be a dream to shitpost on.
>>51767531
Excuse my ignorance, but what are these round things supposed to be? I've seen them a few times now and I can't fathom their use.
>>51767559
let's say it's a 4-core chip with 1 billion transistors per core. what are the chances that one transistor on a core is faulty? close to 0% or close to 100%?
>>51767595
IIRC it's around 80%
>>51767591
nah it wouldn't
You wouldn't be able to split up 4chan into ~800 threads. No algorithms.
>>51767594
>>>/google/
>>51767594
https://en.wikipedia.org/wiki/Wafer_%28electronics%29
>>51767594
It's a silicon wafer on which the processor cores are "printed", more like engraved with UV and shit. Then they cut the cores apart from each other.
>>51767531
Depends on where the defect is.
Some units on die can get binned off, others it may not even impede functionality to any significant degree. Its actually impossible to function check every single transistor on die when they're going through validation.
>>51767550
Virtually no IC is ever left bare, it'll be covered with a few insulating metal layers on top. You can touch them directly without immediately harming anything, though risk of minute ESD, and acid in your oily finger prints will likely kill the chip prematurely.
>>51767617
>photolithography
>engraved
literally the exact opposite
>>51767623
Just wanted to put it in the analogy that plebs can understand
>>51767531
A "bed of nails" tester checks each chip on the wafer and marks it to be thrown out. The whole wafer is not trashed...
Also, in some situations, a processor die will have an extra core that is disabled, but enabled when one of the normal cores malfunctions.
>>51767619
>it may not even impede functionality to any significant degree. Its actually impossible to function check every single transistor on die when they're going through validation.
can you elaborate on both points? what kind of circuitry can be defective and not impede functionality to a significant degree? and why can't the entire circuit (the entire chip) be validated?
>>51767651
>>>51767619
>>it may not even impede functionality to any significant degree. Its actually impossible to function check every single transistor on die when they're going through validation.
>can you elaborate on both points? what kind of circuitry can be defective and not impede functionality to a significant degree? and why can't the entire circuit (the entire chip) be validated?
Faulty transistors can lead to minor things like being unstable while overclocking or require more voltage
>>51767651
If it passes QA, it gets shipped. Most logic circuits have error handling, so if they can account for the unusable circuits, the only thing that's lost is processor speed.
>>51767651
If one cell of SRAM is faulty then an entire bank of cache is not defective. One bit, or even several can be faulty and it won't harm anything, the chip just doesn't address those dead cells. The same is true even with logic. Some transistors within the ALUs in an integer core can be dead, and the unit will still function normally.
Temperature sensors on die malfunction commonly and it doesn't harm functionality at all, though you will see inconsistent temps being reported.
It takes quite a lot for an entire core needing to be binned off on a CPU.
When a chip is put into its test socket and checked they are only looking to see that all of the units on die function, they're not checked every transistor. It would take far too long, and would be an enormous technical hurdle. It would require additional circuitry on die to monitor every single transistor, then you would have no way of knowing if all of your monitoring circuitry was intact. Everything brought up through the BEOL, ultimately theres just no way to check every single gate.
>>51767830
thx. is it possible for the end user to do a full test once the cpu has been installed, by running some sort of validation benchmark software?
>>51767888
Unfortunately no. Theres nothing you can do to look inside the inner workings to see which gates are switching and which aren't.
i work 12 hr shifts in a wafer fab pretty fucking boring i tell ya
>>51767561
intel doesn't die harvest i3's anymore, but yeah, both AMD and Intel try to salvage what they can. that's why AMD made 4 and 6 core variants of the bulldozer and piledriver chips and why they have GPU-less APUs. ditto with intel and no hyperthreading/avx pentiums.
Chips are usually designed with a modular fashion in mind. The aim being that if a bunch of transistors are faulty, they can disable the module that they're in and then sell the chip as a lower-end chip.
This is why different classes of processor exist at all. It's all the same chip but with larger and larger parts of it with broken parts that are then disabled. Like an i3 would be an i7 if it had worked 100%, so they just disable some broken parts and it's now an i3.
>>51767947
what kinds of chips and what are their yields?
>>51767531
There is a certain amount of fault tolerance as part of the design.
The perfect chips are sold as high end models, while the ones with minor imperfections are sold in a lower-cost but still high-end model. Larger imperfections still can be re-used in lower end SKU's with cut down features.
In cases where an unusually large number of chips make it through unscathed they may even go the route of artificially limiting a certain number of chips so they can sell more of the more lucrative non-flagship models.
>>51767975
Really?
I would have thought i7, i5, i3, and pentium are all the same chip with certain bits enabled/disabled?
1 or 2 faulty cores, sell as i3, disable ht and sell as an i5 and etc
>>51768019
Total fab yield is like 98-99% the kinds of chips idk whatever TI makes so anything from Automotive to w/e else they do lol
>>51767975
>3 core phenoms
kek
>>51768042
my guess is that intel makes dual core dies because they're smaller and can be binned/harvested into low power mobile and embedded chips
for -e/-ep haswell they had at least two dies too, 12 cores and 18 cores. the entire haswell-e lineup were harvested from those 12 core dies.
>>51767531
What do they do with the corner pieces that aren't a complete circuit?
>>51767531
>8 billion transistor
we may have gone too far in a few places
>>51768108
Straight into the garbage
>what happens when a handfull of transistors on a monster chip are defective? do they throw out the die, or is the hardware smart enough to work around the defects?
> do they throw out the die, or is the hardware smart enough to work around the defects?
No, of course not. Don't be so silly. This is where AMD comes in building on top of rubbish then selling it to poorfags.
>>51768121
That seems wasteful. They can't re-melt them?
>>51768147
>They can't re-melt them?
nah, it's cheaper to throw the partial chips out than to try salvaging them. the materials are cents per chip while the r&d and time to do something like that would be billions
>>51768118
7-8 billion is the current estimate for knights landing. nvidia and amd next gen gpus will be around 16-17b.
>>51767601
I already shitpost in thousands of threads a day just fucking watch me
Well, all this will have no effect in your daily use if you remember to [spoiler]refresh thermal paste[/spoiler]
>>51768042
>I would have thought i7, i5, i3, and pentium are all the same chip with certain bits enabled/disabled?
They make a couple of variations.
Hyperthreading is disabled to make the non-hyperthreading models.
But they don't disable entire cores. (ie: 2 core chips really only have 2 cores)
The different clock speeds are mostly artificial.
The vast majority of chips could run at the fastest models clocks, if it wasn't for cock-blocking overclockers.
>>51768118
>we may have gone too far
It will probably end soon.
Or at least the rate of increase will very likely slow down.
We're at 14nm now.
7nm requires completely different machines to fabricate (EUV rather than immersion lithography).
Those machines exist but are problematic.
Also quantum tunneling becomes an issue.
>>51767975
>GPU-less APUs
i don't even
>>51767947
Me too! for what company? I work for NXP now.
>>51770528
Hasn't IBM managed to make a working chip at 9nm recently?
>>51768019
1.7%
>>51770762
I meant commercial scale fabrication is currently at 14nm.
With the same machines they can do 10nm, and perhaps 9nm.
But anything smaller and you need EUV.
Several companies have EUV machines (all made by ASML, nobody else can make them) for testing but none have used them for production yet.
>>51770762
7nm.
>>51770821
EUV isn't a requirement, its a cost reducer because triple and quad patterning with these current light sources is incredibly complex.
IBM's 7nm node was quad patterned. You wouldn't want to bring that to market for cost reasons, but it would be entirely possible.
>>51770840
Fairly sure quad patterning is already required for 10nm on ASML's immersion lithography machines.
10nm production will be a mix between immersion and EUV.
7nm production will be EUV only.
>>51769204
ebin!
>he doesn't have a wafer full of 16MB RAM chips to masturbate on
>>51768050
Send me some TMS320s family
i understand that 14nm, 10nm, etc refer to the process, but do these lengths actually correspond to some physical feature? i heard that it used to be the width of the gate but now are just labels for process names.
>>51767583
holy shit and they're cheap as shit
gonna buy one i think
didnt realize this was a thing thanks
>>51772544
More efficient (higher speeds, lower power) and better heat management.
>>51772587
yes, but what LENGTH is XXnm (nano meters) a measue of?
>>51772544
The name of the process node only roughly correlates to gate length. The names are actually based off of ASML published guidelines, but this is over the head of everyone here.
>>51772587
Way to not read.
>>51772672
Thickness of the die? No fucking idea tbqh m8.
>>51770582
The Athlon 860k is a great example of this. The CPU is still in perfect working order but the IGP is cut off. This gives you a bargain bin quad core that overclocks like a demon while being price competitive with the likes of the anniversary edition pentium.
If you want four cores and decent single threaded performance, then you get the 860k. If you just want single threaded performance then you'll want the anniversary pentium. Some games will refuse to run on less than four cores though.
>>51772700
Thought it was supposed to be the size of one transistor
>>51772793
Again, it is roughly. Actual node names are based on ASML guidelines.
This ST Micro 28nm FD-SOI process produces a 25nm gate. Same deal with GloFo's 28nm bulk process. TSMC's 28nm bulk produces 33nm gates.
>>51767531
My a-level tutor has four of them that they got for peanuts because there was abit of dust on them in the process of making
>>51772679
Are there any good intro books to IC manufacturing?
>>51772972
Principles of Lithography
Can anyone post some videos on how these waters are made? There was one posted here a while ago from a hackers conference but I can't find it on YouTube anymore
>>51772574
They are just industrial waste, so why should they no be cheap
>>51767619
>Virtually no IC is ever left bare
Muh camera snsors
>>51773015
Thanks. Are you a researcher or do you work in a fab? Are you an EE?
that thing cost only $25
>>51767531
If you're AMD you sell them as tri cores
>>51773759
Not even those familia. They put all kinds of bullshit micro-lenses on those badboys.
>>51773664
https://www.youtube.com/watch?v=NGFhc8R_uO4
This one?
It's a pretty nice watch.
>>51774489
my a6000 i believe has a layer of glass then a layer of film over the sensor
>>51773744
>implying you wouldn't be the asshole who would sell them at the power the robots it took to produce it rather than scrap value
i was expecting them to be about 3x more expensive
You sell it as a lower tier part.
Like, those defective 980 that were sold as 970 with 3.5gb
>>51773744
They're cool and I assume they're pretty rare. How often do they fuck up so badly they through out the whole wafer?
Maybe I'll use them as part of a tabletop? That might look cool.
>>51774519
>the asshole who would sell them at the power the robots it took to produce it
I guess they want to get rid of them while making a bit profit, instead of not making any profit at all with them because nobody buys them
>>51772587
That didn't answer the question at all.
>>51772679
thx. was this done for technical or marketing reasons?
