Building ASICs

[pretendo]

I know next to nothing about computer hardware, but doesn't creating specialized computing devices require an industrial operation? As in a teams of trained experts, large specialized tools, specific materials. How do the no names supposedly creating bitcoin ASICs do it? Am I overestimating the requirements for mass producing computers?

[Blazr]

I know next to nothing about computer hardware, but doesn't creating specialized computing devices require an industrial operation? As in a teams of trained experts, large specialized tools, specific materials. How do the no names supposedly creating bitcoin ASICs do it? Am I overestimating the requirements for mass producing computers?

What you've said is true. We really don't know how BFL has developed a fully custom ASIC miner, but the other Bitcoin ASIC manufacturer's aren't using fully custom chips, which from what I gather cost much less to develop.

[eldentyrell]

What you've said is true. We really don't know how BFL has developed a fully custom ASIC miner, but the other Bitcoin ASIC manufacturer's aren't using fully custom chips, which from what I gather cost much less to develop.

The term is full custom, not fully custom.  Putting neon lights in your computer case makes it "fully custom".  Writing your own software to draw the transistors -- like Intel and AMD do -- makes your chip "full custom".

BFL has not announced a full custom chip.

[Puppet]

Lets clear up some terms.
An ASIC is a Application Specific Integrated Circuit. Most chips could be considered ASICs, although the "application" might be something very vague or wide, like running x86 code in the case of an x86 CPU. Technically it could be called an ASIC. A  bitcoin ASIC by all reasonable definitions would be a chip designed specifically for bitcoin mining, and as such its a "custom" chip. By contrast an FPGA is a field programmable gate array, its a chip that can be programmed to act as whatever you want by just loading a different bitstream.

"Full custom" has little meaning, but usually when a company says it does a full custom asic, it means they not only design the chip, but also design individual custom transistors, rather than relying on standard cell libraries provided by the foundry. More typical for asic's is to use those libraries. Think of the difference as between using C++ instead of assembler. In both cases you end up with custom code (or hardware) specific for your application, but one is a lot harder and lower level than the other but might yield slightly better results because you can optimize more. I imagine AMD and Intel will usually do "full custom" in that sense, but they are exceptions. Very few companies do it, because its terribly expensive, complicated and usually pointless. So its extremely unlikely BFL is doing a full custom design in this sense.

When BFL says they are doing full custom asic, what they most likely mean is a chip designed specifically for bitcoin, and the custom part refers to the chip design, not the transistor design.

As for whats needed; well, a lot obviously. But you dont need to have it all in house. Very few companies have all the expertise and equipment to develop a chip from scratch to working silicon. Intel is one of very few (only?) company that does that. Even AMD no longer operates its own FABs, it only designs its chips, and outsources the rest. SO these companies rely on foundries like TSMC to do the manufacturing, and you can outsource any stage of the design, development, production, testing and assembly you want. So at the other end of the spectrum, you could write an FPGA bitstream and tell altera to make chips based on that, and you need hardly anything at all, just a single developer with the right software (both of which you could even hire if you wanted) and few $100K.

[mobodick]

I imagine AMD and Intel will usually do "full custom" in that sense, but they are exceptions. Very few companies do it, because its terribly expensive, complicated and usually pointless. So its extremely unlikely BFL is doing a full custom design in this sense.

I don't even think Intel and AMD do this in that way.
They propably make custom libraries, but designing something like a whole cpu from scratch (500~1000 million transistors) would be crazy.
There is abstraction on several levels. Transistors are made into subcomponents that are put together into functional units which can be part of bigger pieces.

[SgtSpike]

What you've said is true. We really don't know how BFL has developed a fully custom ASIC miner, but the other Bitcoin ASIC manufacturer's aren't using fully custom chips, which from what I gather cost much less to develop.

The term is full custom, not fully custom.  Putting neon lights in your computer case makes it "fully custom".  Writing your own software to draw the transistors -- like Intel and AMD do -- makes your chip "full custom".

BFL has not announced a full custom chip.

Yes they have.

...but Butterfly managed to skip structured ASIC entirely and create a solution that is built using the full custom approach.

http://bitcoinmagazine.net/bitpay-breaks-daily-volume-record-with-butterfly-asic-mining-release/

[mobodick]

So how about the required design?
SHA-256 is a pretty simple compared to the functionality of a CPU.
I can imagine they can get a lot of performance from making simple units and repeating them on die, maybe a bit like shaders on a GPU but without all the stuff for graphics like rops.
This would make designing it more feasable.

[Puppet]

A bitcoin asic is probably about as "simple" as any asic would get. Almost no IO, and very simple, easy to validate logic, incomparable to say cpu's or gpu's
That doesnt change the fact that building any asic is pretty expensive, particularly on a relatively modern process, mostly because of the maskset.
 I see no reason why BFL wouldnt be able to pull it off though. Aside from the money and tools, all it takes for the design is one or two experienced guys and a couple of weeks/months. The production you simply outsource to some taiwanese or chinese foundry.

[firefop]

Intel hasn't done it inhouse for years, they contract through tektronics - same a motorola and pretty much everyone else.

ATI used to grow their own - AMD used mitsubishi to grow silicon... not sure if AMD taking over ATI was to get the manufacturing in house again or what.

[Puppet]

Intel hasn't done it inhouse for years, they contract through tektronics - same a motorola and pretty much everyone else.

?

Intel is one of the largest fab operators:
http://en.wikipedia.org/wiki/List_of_Intel_manufacturing_sites#Fab_sites

tektronics just makes test equipment and has nothing to do with this. Perhaps you meant ASML, which is a big supplier of fab equipment to just about any fab, but that doesnt change the fact intel owns and operates its own fabs.

ATI used to grow their own - AMD used mitsubishi to grow silicon... not sure if AMD taking over ATI was to get the manufacturing in house again or what.

Growing silicon? No one is talking about the production of raw wafers here. And you got every one of your "facts" wrong. ATI was a fabless design house that used (and still uses) TSMC for wafer processing and production. AMD used to have its own fabs in Desden, but spun them off in to a separate company called Globalfoundries.

[BlackHeartFund]

I hope that BFL comes through with their ASICs, but I admit I am not totally sold that they have done it and will deliver as promised in a reasonable time frame.

I have some BTC in mining operations that are waiting for the BFL ASICs, but not much. If they come through I will go crazy buying interests in companies with pre-orders on the first day they start being delivered. I'd probably order as many rigs as they can deliver... but I'd be late, because I wasn't and still am not sold enough that the technology will be delivered as promised to pre-order. If BFL comes through, those who took what I consider to be the risk of pre-ordering, will cash in, as they should.

If they have the technology and can deliver, great. I really hope that is the case. I haven't seen any evidence of it yet, but I have also seen no concrete evidence that BFL will not honor its contracts. Slow deliveries and growing pains aside, as far as I can tell they are a legitimate company that delivers on their contracts. I have no evidence to suggest they will not deliver their promised ASICs. I really hope they do, but...

If they lied about the technology and aren't going to deliver, it was a great (if unethical) business strategy. Obviously everyone would be furious and  wouldn't do business with them again, but with the pre-orders and people waiting to trade up, they have a complete monopoly on the mining hardware industry. They could release a slightly upgraded FPGA or poor performing ASIC and basically force everyone to buy it. I'm not saying they are lying, this is 100% mildly informed speculation, but if they have been pretending to have this technology coming very soon, when they really have no chance of ever doing it, that 100% buyback guaranteed that no one purchased anything from their competitors. They cornered the market and increased their business what, ten fold? With a promise. Not bad.

It seems quite clear that BFL are either going to revolutionize BTC mining and become the Intel of the Bitcoin community, or they are going to piss off everyone and make a much smaller fortune by becoming another in the long list of Bitcoin failures and frauds. I seriously hope they are going the Intel route, but to me it's a coin flip and I'm not putting money on them delivering, YET.

Good luck BFL, hope you guys are staring at silicon right now and not piles of cash!

[bitmar]

I found something interesting:

If you want to design ASIC chip you can use the Synopsys Design Compiler - having chip design, you can run the simulation to know  specification without building. Here an example, but they had existing 2 ASIC designs: http://csrc.nist.gov/groups/ST/hash/sha-3/Round2/Aug2010/documents/papers/SCHAUMONT_SHA3.pdf

[Bogart]

Do y'all reckon any of the mining ASICs currently in development use tri-gate transistors or memristors?

Are these technologies likely to already be fully exploited in existing SHA256 IP cores that the makers of standard-cell based ASICs are likely to have licensed for use in these mining ASICs?

Or if a Soft Core was licensed, are the synthesis tools likely to fully exploit these technologies?

[firefop]

Do y'all reckon any of the mining ASICs currently in development use tri-gate transistors or memristors?

Are these technologies likely to already be fully exploited in existing SHA256 IP cores that the makers of based ASICs are likely to have licensed for use in these mining ASICs?

Or if a Soft Core was licensed, are the synthesis tools likely to fully exploit these technologies?

very unlikely.

We're in the 'old and slow' camp for these asics (at least we'd better be) I'd expect everything to me on 120nm (ish) - those gains (may or may not be slight) won't be in this round of asics.

[Sitarow]

I found something interesting:

If you want to design ASIC chip you can use the Synopsys Design Compiler - having chip design, you can run the simulation to know  specification without building. Here an example, but they had existing 2 ASIC designs: http://csrc.nist.gov/groups/ST/hash/sha-3/Round2/Aug2010/documents/papers/SCHAUMONT_SHA3.pdf

Intresting

[Bogart]

I found some open source SHA cores including SHA256. 

http://opencores.org/project,sha_core

They only mention a 180nm process.

I wonder if any ASIC miner products will include a copy of the LGPL.