loading

How to make an ASIC from a FPGA?

I am starting to play around with FPGAs, and I want to eventually make my designs into a permanent IC. Is there any company that will od that for me? All that I've heard of are companies that take designs to mass produce them, and cost hundreds of thousands of dollars. But I only want one or two ICs, and I want to pay a price that makes sense with the design (simple designs cost only a little, then the price goes up as complexity increases). Is there anyone that does that?

sort by: active | newest | oldest
kelseymh1 year ago

No. As Steve says, the whole point of FPGAs is to get the speed of chip-based logic (i.e., faster than microcontrollers) without the cost of going to a foundry. Chips are made by the WAFER, not by the chip.

The setup cost for multilayer lithography is tens to hundreds of thousands of dollars, while the subsequent run is about a kilobuck per wafer. You can get a few dozen ASIC chips per wafer (more if the circuit is simple enough to get a small form factor).

So if you only need a few chips, they are going to cost you ten grand each (full setup cost divided by a small number). But if you need a million chips, then they're maybe fifty bucks (for a large ASIC), since the setup cost is amortized over the whole run.

FloppidyDingo (author)  kelseymh1 year ago

Of course they don't, why would they. Don't you ever get the feeling that companies don't like tinkerers like me? This is why I hate bug business. Oh well. Thanks for the straight response.

I think you might be missing the key point. "Setup cost" is a real thing, not a conspiracy.

1) Each separate layer of the chip has to have its own mask (that's a piece of material which blocks UV in some places while allowing it in others). Each mask is unique to the particular chip being made; for an ASIC, that means the mask is only for *you*; the company can't reuse it later for other chips.

2) All of the different pieces of equipment used in lithography have to be configured for each job (filaments or boats for sputtering, power settings, etc.), and the whole job run needs to be completed before that equipment gets reconfigured for a different job.

3) Configuring and testing the setup, and validating the resulting chips, is both material and labor intensive, which costs real money.

Those three things are a large up-front cost, but a flat cost which is the same whether you want ten or a million chips. The part of the cost which scales with job size is mainly for the materials (the wafers, the material used for sputtering or ion-deposition of dopants, insulating layers, etc.), and for the electricity use to run the machines.

Those unit costs are small (for the company and for the client) per individual chip, so the setup dominates in small runs, while the materials dominate in large (mass production) runs.

By the way, exactly the same thing is true for other kinds of manufacturing, like injection molding or vacuum forming. There, the "setup costs" involve actually fabricating the molds and jigs used for the specific item being made. That fabrication involves precision machining by expert technicians, among other things, so it's expensive. Once the jigs have been made, the unit cost for materials is small.

I was talking to the guys from E2V in England at an SPIE conference in Baltimore back in May. E2V are the leading manufacturer of CCD imaging sensors for professional scientific applications. I don't think there's a science imaging mission in orbit not taking pictures with their chips.

Some of THEIR CCD chips have a yield of ONE per wafer....there's a reason why big chips cost megabucks.

Wow. Yeah, I didn't even mention the fractional yield issue (it's essentially a bump in the piecerate cost rather than setup). I didn't think anyone made individual CCDs that large (~6-8 cm), because of the consequent slow readout. For scientific (or national security) imaging applications, you really don't want dead pixels out of the gate, so the yield is going to be < 100%.

Not really, the process is the problem: it only works economically at very large volumes.

FloppidyDingo (author)  steveastrouk1 year ago

I guess you do make a valid point. I guess I'll just stick to FPGAs then.

the whole point of the FPGA, is to eliminate the costs of ASICS ! There are still cplds which, depending on what you're doing MIGHT work.

FloppidyDingo (author)  steveastrouk1 year ago

I highly doubt I'll find a CPLD large enough, or fast enough, to become a GPU. I just thought it would be cool to have my own IC that I made myself.