Bitcoin Forum

I know it must have been asked before,

Why can't we group buy fpga and get a discount ?

I would invest in any effort to get an ASIC made, currently the only thing that would be holding me back on FPGA is mhs/$ and the threats of an ASIC batch being made.

Can someone in the semi industry describe a best case scenario for a cheap ASIC + packaging,  For arround ~300 000$, what are possible options ?

still most of this discussion is in another language for me but i am looking forward to the day that normal humans can use fpga's.

It's either:
A. There's no one in this community that we can trust $300,000 with.
B. People don't have enough trust to give this person/organization $300,000.

Given Bitcoin mining's anonymous nature, there's really no way of knowing whether the chip designer/chip maker/chip packager/PCB assembler is ripping you off or not. The designer and foundry could secretly start their own cluster. The packager and PCB assemblers could be profitting from the "rejects". And when we're getting ripped off I doubt anyone in their right mind would donate another $300,000 to assemble a legal team to sue the offender.

The designer ? He does not have to touch any chip, it can be open-sourced, and taken from good a fpga design.

The foundry can already do this, I doubt foundries would be running with the money.  They usually do what they're paid to do, it's not of their concern what the chip does, except checking for IPs.
Packager ?  They could be required to give back every bad chip they get.
Suing is not usually the way bicoiners solve their problems.

Anyhow, that doesn't stop us from elaborating technical details.


A handful of people would benefit as much from securing the network/investment that they would mining for coins with this hardware.

Obviously this can't be open-sourced due to its very nature, because if we pooled together $300,000 for a group buy, most of it will go into the IP. Someone who didn't participate in the group buy can just run to the fab with the open source design and spend his $300,000 entirely on the hardware.

If AMD's fab, TSMC, started selling a CPU that perform exactly like Bulldozers, everyone would notice immediately.
If this group buy goes through and the global hash rate sky rockets, no one will suspect that TSMC is behind the scenes pumping out more chips than what we ordered. Even if such suspicions do arise, nothing can be proven due to the anonymous nature of mining. So TSMC has nothing to lose and everything to gain.

That would work if you're willing to patrol the floor and audit every shift. Packagers have no incentives to steal now because selling the stolen goods would be very difficult like I mentioned above. However mining chips are different because they don't need to be sold. They can used to generate profit legally and anonymously.   

RE the FPGA group buy, it's a good idea but IMO will need to hit qty 250-1,000 per order
to even get 10 minutes attention from the senior execs with the power to approve a
real discount [30-50%+]

-rph

Higher than that actually.  1k parts isn't even close to being considered enough volume to get any real discount.

Even if every single Bitcoin miner were to participate in this GB, the volume doesn't justify any real discount.  In the industry, you have to have part volumes > 10-100k (depending on many factors) to get any real leverage, and that applies only to customers who agree to long term business relationships as well.  Bitcoin is a drop in the bucket compared to commercial designs.

Exactly. And to make an ASIC you need to invest a lot of money in development. Maybe $1 million?

This (http://bitcoin.stackexchange.com/questions/103/whats-the-difference-between-an-asic-and-an-fpga-which-is-better-for-mining) may be relevant to the discussion

For graphic cards, 100 may seem a lot but for FPGA, the quantities are in production quantity and can be tens of thousands.  If you just need some for pre-production or R&D purposes, some of the companies will want to know where estimated time line.  I doubt we can form a group large enough to take on a substantial quantity for industrial pricing.

FPGAs are fairly low-volume devices to begin with. qty 1000 of 6s150 is $150k+
at list price, and for that kind of business, you could probably get a nice discount
directly from Xilinx. The problem is getting enough reliable buyers to reach 1k.

-rph

well someone(maybe me if i have time) needs to put in a request to a FPGA manufacturer to get a real number for the price we could expect on a 1k volume of chips, once we have a target we can think of the best and safest way to approach it

Yup go for it, I would be in for at least 5-10 if you can get 30%+ off the Avnet qty 1 pricing. I think all the guys
selling or building boards should pool together.. it's not a big enough market to save much on our own..

-rph

If you can get a 30% discount I would take at least 50.  But I'm not holding my breath.

940 to go..  8)

-rph

Maybe this would help?

http://www.kickstarter.com/

Dunno if this would be seen as very "Creative", though... hmm.

AVNET is SHI.T, I repeat, AVNET is SHI.T

I read that and saw a link to largecoin.com --- Not many details yet however.

In the US they're the cheapest for 6s150 -3 qty 1 last I checked. Worked out OK for me.

-rph

So who is not shit?  Who has the lowest prices?  How much lower are they then AVNET?  10%? 20%?

On edit: still waiting...

As far as a quick google search goes, Digikey has the same price as AVNET?

Yeah, most disties have the same list prices for "big ticket" single source items like FPGAs.

It's sort of like OPEC.

-rph

I'm still interested in a group buy..

What are you guys thinking for the exact part #.. xc6slx150-3fgg484c? Or -2?
And does the -3N even exist?

-rph

Here -3N  is much expensive than -3C

I don't know Y.

I use -2I.

I'm working on a discount.. should know more in a couple days.. if you're interested
in a group buy please feel free to PM me..

-rph

I doubt that we will see high volume fpga designs or cell based asics any time soon.

I even think that we can completely skip this step and go all the way to full custom in a few years... but not as most of you think.
Bitcoins popularity is still declining and it will take some time before it's gonna go back up. After that bitcoin will be much more mature.

But there is a trend independent of bitcoin in electronic hardware which is larger than bitcoin itself:
Within a few years there will be hackerspaces which are equipped with a clean room and lithography equipment. The goal will not be to stay there but to have Open Source Designs for the machinery itself.

It will be during this transition period when the first low integration chip designs pop up. One of them will be some sort of bitcoin appliance. (It makes much more sense to make it do more than just mining.) This Development will dwarf everything we have seen so far and bitcoin will be seen as sidekick development and can ride this trend.

Until this comes into reality there will be a tough road ahead though...

I hope you are right - that would be sweet - but there is simply no way a hackerspace IC fab is
going to compete with a 45nm FPGA or GPU for bitcoin mining.

It's not physically possible to build 1200 32 bit adders @ 150MHz on 10 micron.

-rph

I don't think that it would be just 10 micron. If assuming the first ones will be getting used university research equipment I think that 250-180nm is reasonable in the long run. The old stuff was always upgraded and the mechanics can handle it. At this integration a custom design would seriously surpass existing tech.
It wouldn't make sense otherwise to keep the equipment around at all otherwise at all which is still done.

Something like that would obviously be one of the later projects but how I see it the development has been made, and I see no reason to start from scratch in terms of available techniques... They just didn't know how to get to a smaller feature size back then what we are do now and this holds true until we need special laser sources, other atmospheres or phase shifting so replicating anything up to this point should be doable as the working principles are the same.

You can do a full-custom 250nm design on MOSIS today for under $100k -- which is definitely way cheaper than building a whole fab --
but I am really skeptical that it will beat a 45nm FPGA. The FPGA overhead seems much less than (250/45)^2 to me.

-rph

Well you can approximate the overhead by using some assumption for the necessary number of 32 bit operations like 3600, the gate count of your fpga and the effective mhash. And you are right the overhead is probably smaller than the disadvantage in feature size.
But if you add the trading margins... well, I have no idea where to start.  ::)

One can probably buy a wafer, develop it, slice it, test the chips and package them for the price of just one Spartan-6 150. (Not including labor of course :P)

But back to the main point of TT...
The problem I think is the quasi monopoly of fpga manufacturers, not that there isn't enough competition between them but their unique position relative to their consumer base. Its like jewelery diamonds, everybody knows that they could be dirt cheap but a few entities control the supply and make sure they are sold for a good price.

Also most consumers don't really care.

Take a hypothetical $200 FPGA.  It likely is either used for
a) a prototype.  Cutting the cost to $80 is immaterial compared to entire prototype phase cost
b) for niche high end unit likely selling for 10x to 100x the FPGA cost.  Sure cheaper FPGA might be nice but the companies profits are more based on the IP of their design than the cost of FPGA.

So you have a tightly controlled supply combined with most "consumers" who are not needing to cut prices as low as possible.  That means margins are going to be and stay high.

Interesting idea here. Any thoughts on how to organize this?

There are two major challenges: the first is actually getting a discount from Xilinx, and the second is
getting paid reliably by people in the group buy. On the first - none of the authorized distributors
are budging for me. A real discount may only come after a couple large orders to establish credibility.

There are brokers that sell surplus/old inventory at a discount (for example, 2010 ES parts).
I haven't looked into that much for this specific device; in the past, after the broker fee, it's usually
not that much cheaper. And you risk getting damaged/relabeled/stolen/counterfeit parts, etc.

So, in conclusion, meh. DeathAndTaxes pretty much nailed it.

-rph

This is one time I wish I was wrong.  However it has been my experience pricing of FPGA is very firm.  Only highly valued (i.e. high profit repeat customers with HUGE volumes) get reduced pricing. 

Why?  I don't know.  Maybe not enough players for a true free market.  Maybe because most of the time FPGA are either used in prototypes or high end expensive devices where cost of FPGA is immaterial to project cost.

Your best bet is repeat business.  Not sure if there is enough demand for it but if you can buy 200 units, then next month 200 units, then next month try to push for a discount.  If you can convince them this isn't a 200 unit buy but part of a 24,000 unit buy over the next year they might be able to budge.

Once you've already bought 400 parts, they know it's a pain for you to switch to $OTHER_VENDOR,
if you are a small/mid-sized company..

You have to be so large/rich that you can tell your engineering team to change vendors on a whim,
just because the sales guy annoyed you somehow. Bought the $100 champagne instead of the
$1000 champagne at lunch. Then -- only then -- you have some actual pricing power.

-rph

Brutal.  Yeah there are just too few players in the space both suppliers and consumers.  Not a very efficient market.