I believe you are making the common mistake of only counting interesting engineering and I agree, that is almost nonexistent in here. However in a proper ASIC engineering there is huge amount of non-interesting engineering especially when testing is involved. I also agree that BFL is probably skimping on the testing capabilities and that is another reason why you do not want to take on tasks that rely heavily on the testing to get them to work.
Also when talking about the technique of switching off cores I was not talking about BFL, but mainly thinking about applications where this kind of technique could theoretically make some sense meaning designs 1000 times bigger and more complicated and in there adding stuff in there is also 1000 times more complicated than in simpler designs but again in such small designs this makes absolutely no sense from economical point of view. You simply do not lift a finger trying to chase a savings of grand total of $150. There are other things 100 times more important to worry about. If we would be employed by BFL, even this discussion we were having in here would have cost the company more than $150 and we haven't even started actual work.
I believe normal ASIC manufacturing process involves testing the chips either before or after packaging and shipping only known working chips to customers. Many customers do not do inhouse testing the ASIC chips separately and neither tests the 100 other kinds of non-ASIC chips that go on to the PCB, but just solders them all on there and after that performs the testing for the full PCB. It is also possible that the defect on the ASIC is such that is totally unfunctional and goes up puff in smoke and you really do not want to trash many fully populated PCBs by having such chips soldered on so you need to separate somehow the faulty chips you want to use from ones you do not want to touch.
Topic: Butterfly Labs is going to give lifetime warranty (Read 7544 times)
Switching cores off is a useful technique to be used on chips that have huge die sizes ( 200mm2 and up ) where defect probabilities go up and for each defect you either have to throw a very expensive chip to bin or do something else that also might cost big money. The disabling a core is also an expensive solution because it needs a lot of engineering and testing.
I believe this is not useful for BFL because the die sizes are so small making each chip so cheap to produce that is much more economical to just throw the faulty ones to bin rather than trying to do something clever with them. The small die size also means that defect probabilities are much smaller than on bigger dies as the defect probability is some constant multiplied with the die size.
Pulling some numbers from my ass: Not being able to switch cores off could cost them 50 chips each worth $3. How much engineering would you like to do to save $150? The situation is totally different if we are talking about saving 10000 defective chips each worth $200.
The engineering required is near zero. Basically a gated clock buffer instead of plain clock buffer. And a single register wide enough to store those "clock enable" bits for each hashing core.I believe this is not useful for BFL because the die sizes are so small making each chip so cheap to produce that is much more economical to just throw the faulty ones to bin rather than trying to do something clever with them. The small die size also means that defect probabilities are much smaller than on bigger dies as the defect probability is some constant multiplied with the die size.
Pulling some numbers from my ass: Not being able to switch cores off could cost them 50 chips each worth $3. How much engineering would you like to do to save $150? The situation is totally different if we are talking about saving 10000 defective chips each worth $200.
The overall design is extremely repetitive. The first constraint is thermal. The second one is power distribution rail bounce caused by the huge number of simultaneous switches. Gating the clock would be a standard hardware debugging technique for such a project. I could say that not including clock gating would be a design mistake. The primary objective is to facilitate debugging. Defect tolerance is an additional benefit obtained for free.
Again, the chip is so repetitive and so self-testing, that standard debugging aids (like JTAG) are nearly worthless. The chip is almost an analog or mixed-signal power chip: the primary constraints are thermal and parasitic impedances.
Edit: Furthermore, I think none of the Bitcoin ASIC manufacturers can afford to invest time and money into a proper chip testing. I'm thinking that all packaged chips will be soldered into the mining boards and the final testing will be in-situ. I don't even think that an investment into the proper test equipment would be worthwhile from the engineering point of view. All in all, those chips are just lottery ticket printing machines, it doesn't make sense to test if some rare tickets are missing or mangled. Each winning ticket is worth something for just a couple of minutes maximum.
Switching cores off is a useful technique to be used on chips that have huge die sizes ( 200mm2 and up ) where defect probabilities go up and for each defect you either have to throw a very expensive chip to bin or do something else that also might cost big money. The disabling a core is also an expensive solution because it needs a lot of engineering and testing.
I believe this is not useful for BFL because the die sizes are so small making each chip so cheap to produce that is much more economical to just throw the faulty ones to bin rather than trying to do something clever with them. The small die size also means that defect probabilities are much smaller than on bigger dies as the defect probability is some constant multiplied with the die size.
Pulling some numbers from my ass: Not being able to switch cores off could cost them 50 chips each worth $3. How much engineering would you like to do to save $150? The situation is totally different if we are talking about saving 10000 defective chips each worth $200.
I believe this is not useful for BFL because the die sizes are so small making each chip so cheap to produce that is much more economical to just throw the faulty ones to bin rather than trying to do something clever with them. The small die size also means that defect probabilities are much smaller than on bigger dies as the defect probability is some constant multiplied with the die size.
Pulling some numbers from my ass: Not being able to switch cores off could cost them 50 chips each worth $3. How much engineering would you like to do to save $150? The situation is totally different if we are talking about saving 10000 defective chips each worth $200.
Remember, the cores don't need to operate perfectly.
The occasional false positive or false negative will not really impact things much when it comes to mining.
The occasional false positive or false negative will not really impact things much when it comes to mining.
What you're describing is certainly possible but I don't think it is practical for BFL and I'm surprised nobody else in this thread has called you out on it yet. They're making small batches of a niche product - there is no big volume here. The "switch a core off if it didn't work" model is only really practical if you're making chips by the boatload. BFL needs a low chip failure rate to survive.
Please remember, that the mining software doesn't access the hashing chip directly, but through the the secret firmware. There will be no need for expensive lasering or fusing on the die. Just the firmware has to have a bitmap of the verified working hashers. So the whole situation is very different from the things like CPUs with whole sections of them disabled at the hardware level, because BFL can disable the defective cores at the level equivalent to BIOS. I'm also assuming that they didn't do fully-unrolled hashers, but much smaller, iterative ones. Therefore I assumed that the chip will have at multiple tens or even hundreds of hashing cores.
On the FPGAs the hashers were unrolled primarily because the deficiency in the place-and-route software heurisics: for the rolled designs the P&R failed to converge or produced spectacularly bad layouts. ASIC designers use way more advanced layout-optimization software.
And vice versa: the device can be populated with partially defective hashing chips. If the testing discovers a fault in some section of some hashing chip that section could be simply powered down and the whole device will remain useable.
What you're describing is certainly possible but I don't think it is practical for BFL and I'm surprised nobody else in this thread has called you out on it yet. They're making small batches of a niche product - there is no big volume here. The "switch a core off if it didn't work" model is only really practical if you're making chips by the boatload. BFL needs a low chip failure rate to survive.
Well, probably because electronics are only truly useful for a short period of time, then people would purposefully break them to get a newer/better model.
People already routinely do that, where they "accidentally drop" a friends smartphone, then ask their insurance to pay for that friends new phone.Craftsman tools also come to mind.
Well, probably because electronics are only truly useful for a short period of time, then people would purposefully break them to get a newer/better model.
People already routinely do that, where they "accidentally drop" a friends smartphone, then ask their insurance to pay for that friends new phone.
Well, probably because electronics are only truly useful for a short period of time, then people would purposefully break them to get a newer/better model.
People already routinely do that, where they "accidentally drop" a friends smartphone, then ask their insurance to pay for that friends new phone.
There are ways to somewhat prevent this, with water-detecting stickers on the insides of electronics and tamper-proof warranty stickers on the casings, but I'm sure I could kill just about any electronic device without raising any of those flags fairly easily.
So, I can understand why electronics companies do not offer true lifetime warranties, but STOP CALLING IT A LIFETIME WARRANTY!
Well, probably because electronics are only truly useful for a short period of time, then people would purposefully break them to get a newer/better model.
People already routinely do that, where they "accidentally drop" a friends smartphone, then ask their insurance to pay for that friends new phone.
Figures... how come lifetime warranties never actually last a lifetime? They should be called product life warranties.
Lifetime of the product, not your lifetime. That's how a lifetime warranty should be.
Idk why, but electronics just aren't like that.
Figures... how come lifetime warranties never actually last a lifetime? They should be called product life warranties.
Lifetime of the product, not your lifetime. That's how a lifetime warranty should be.
Idk why, but electronics just aren't like that.
Figures... how come lifetime warranties never actually last a lifetime? They should be called product life warranties.
Lifetime of the product, not your lifetime. That's how a lifetime warranty should be.
"Sorry Sir, lifetime warranty means the lifetime of the product, since your product is now technically dead, that means the warranty has expired"
"aaaand it`s gone! this line is for people with the money only please step aside!"
Figures... how come lifetime warranties never actually last a lifetime? They should be called product life warranties.
Lifetime of the product, not your lifetime. Lifetime of the product line. Lifetime of the product would actually be an even smarter and more evil Warranty...
"Hey my product broke but it's under lifetime warranty so please replace it"
"Sorry Sir, lifetime warranty means the lifetime of the product, since your product is now technically dead, that means the warranty has expired"
Regardless we should stop using Lifetime warranties, as it is misleading, doubly so in an industry that will have ridiculously short product line lifetimes.
Figures... how come lifetime warranties never actually last a lifetime? They should be called product life warranties.
Lifetime of the product, not your lifetime. Figures... how come lifetime warranties never actually last a lifetime? They should be called product life warranties.
But then people would know what's up and no one would be fooled.
Figures... how come lifetime warranties never actually last a lifetime? They should be called product life warranties.
There is no loophole.
Of course there's a loophole. Here it is:
https://forums.butterflylabs.com/showwiki.php?title=FAQ:What+is+the+warranty+period+on+your+devices&highlight=lifetime
Quote
Lifetime is defined as the length of time we are manufacturing that product.
https://forums.butterflylabs.com/showwiki.php?title=FAQ:What+is+the+warranty+period+on+your+devices&do=comments
Quote
Once we stop producing a product, the warranty will no longer apply
World's shortest "lifetime warranty". You'll get one year out of it, tops.
have you ever had memory fail? I thought I did once but it was incorrectly inserted, lol.
Yes, I've lost about 1% of the RAM I buy for our data center, sometimes it just stops or starts corrupting data and has to be replaced. At home I've had a few sticks fail over the years, but the last one was 1st generation BGA PC150, so I've been luckier recently. All good memory makers give lifetime warranty.
FYI, if you have random crashes that don't trace back to drivers, try some different RAM, you might be surprised.
Jump to: