4
\$\begingroup\$

Is there IC fabrication services? You upload the IC design you want and pay. They send you physical IC of what you ordered.

Something like 3D printing service, such as shapeways.com for example, but for Integrated circuit.

Do such service exist? Why not?

EDIT:

I want to build SoC that I would use to build a server/super computer. I'm trying to build a computer like raspberry pi. But want it smaller and faster. If I get to design my own CPU. I could build something that is much better, then existing systems, because server on chips would take less space and power. At least for my programs.

\$\endgroup\$
6
  • 4
    \$\begingroup\$ Yes. TSMC for example. \$\endgroup\$
    – MarkU
    Jan 29, 2017 at 10:29
  • \$\begingroup\$ Yes, there are, see my answer. But it's very likely not for you. Cost and feasibility depends very much on what kind of IC you need – is it something that has both a digital part that runs at GHzes and must not use much power, and therefore needs to be no older than the 45nm generation, and an analog part that requires large currents, or is it just something that is basically easy to model using a couple of the analog building blocks that are common, in a "dated" structure size such as 220 nm? That makes a difference of 0.2-5 M$. \$\endgroup\$ Jan 29, 2017 at 10:42
  • \$\begingroup\$ So, your question is too broad unless you describe what specifically you need. and "a chip" isn't specific. \$\endgroup\$ Jan 29, 2017 at 10:45
  • 2
    \$\begingroup\$ Sure, send TSMC $1M on your credit card with the files and get samples of your part in almost the latest technology. Revise as required. \$\endgroup\$ Jan 29, 2017 at 10:49
  • \$\begingroup\$ I want to build SoC that I would use to build a server/super computer. I'm trying to build a computer like raspberry pi. But want it smaller and faster. If I get to design my own CPU. I could build something that much better, then existing systems, because server on chips would take less space and power. At least for my programs. \$\endgroup\$
    – Mustafa
    Jan 29, 2017 at 11:22

3 Answers 3

11
\$\begingroup\$

Do such service exist? Why not?

Yes, such services exist.

I'll let "why not?" stand anyway, because if you ask "Why can't I use them?":

By the phsyics of the processes involved, however, there's practically no chance that you'll be able to afford this :(

Silicon chips are produced on wafers. Wafers are large discs of highly pure, prepared silicon, that are then exposed to lithography and selective doping and metallization. The masks necessary for that are immensely expensive to manufacture. So having your own wafer will set you back at least a couple 100 k$, depending on the process you need. Sure, then you'll have hundreds of your chip, but that isn't all that great if you needed 3.

This is, however, how everyone in the market (who doesn't have their own fabs) produces ICs: you call your favorite fabs, get an offer for the process, wafer size and yield you want, and then transfer files and a lot of money. The most well-known example of such contracting fabs is TSMC.

Then, there's services that put a couple of your IC on to a "omnibus" wafer together with other customer's ICs. That of course restricts your process to one of the most popular ones, because you need a lot of other people to share your wafer in order to control the cost.

All in all, you don't need custom ASICs, unless you know what that means.

Why can't I use them?

Because IC design still requires a lot of very specific, pretty hard, knowledge that, frankly, you don't seem to possess. Since making a mistake is really expensive here (see above), your 3D printing service is a nice analogy:

Imagine the first 3D print of someone would be a really complicated part. How large is the likelihood that they produced the perfect 3D model file that can actually be successfully printed without someone manually going through that file, adding support where necessary, taking away unprintable parts etc? Practically zero. Now make your 3D print 10,000 times more expensive and you see why this will no option for a beginner.

I want to build SoC that I would use to build a server/super computer. I'm trying to build a computer like raspberry pi. But want it smaller and faster. If I get to design my own CPU. I could build something that much better, then existing systems, because server on chips would take less space and power. At least for my programs.

Aside from my slight doubts that you alone or friends will get anywhere near the performance that CPUs have achieved after 50 years of optimization, this means that "old" processes don't make any sense, since you need high clock rates and thermally manageable losses.

That means you need to use competitive structure sizes – if you want to be faster than current hardware, you can't use transistors that are slower. So something like 22nm or 14nm is your target structure size.

Sadly, a single mask for these cost millions of dollars. So this is totally out of your reach, both from a competence (how many expert on optimizing of memory accesses, pipelines, reordering, high-speed busses, speculative computation, on complexity of implementation, CPU design, ISA design, VLSI, … do you have on your team?) and a financial perspective.

If you want to implement your own CPU, start small, and program it and simulate in software. Then, put it on an FPGA. After that is proven to work, you can port it to ASIC technologies, and not earlier.

\$\endgroup\$
2
  • \$\begingroup\$ There are services that similar to PCBs allow you to buy space on their water, allowing the cost to be as low as a few hundreds per IC, but obviously with lots of limitations \$\endgroup\$
    – PlasmaHH
    Jan 29, 2017 at 11:03
  • \$\begingroup\$ @PlasmaHH that was exactly what I was trying to refer to as "omnibus wafers". \$\endgroup\$ Jan 29, 2017 at 11:03
6
\$\begingroup\$

Although the other answers are good, they are not complete because they are missing some actual resources might get you started. mosis and xfab are groups that let you do shared wafers. The numbers I see of masks and 10 wafers on 130nm are USD85k or so. On 14nm, with dedicated 14nm runs (you get to pick your recipe), you are looking at about USD30M. Keep in mind, once you have your masks, you are printing money at USD3k for a silicon wafer. However, I find that these costs are small compared to the costs of the design software.

I disagree with the desire of other posters and comments to use an aggressive process just of the sake of "speed". You do not have to be at a smaller node to go faster, as the metals do not scale, and the devices are terrible because you hit velocity saturation before threshold. The overdrive gives you barely an exponent of 1.1 and then you actually see this fall to about 1.0 or less at the process Vdd. For instance, the multiply unit on a 65nm Cell processor required 43.1 pJ and an asynchronous multiply unit that we made on the same process only required ~2pJ because we could operate it near threshold. If you want to make things faster, it's often a question of architecture.

I've found that I get the best power performance at 45nm, but I'm not doing clocked systems. If you want to see how terrible the 14nm world is, take a look at the illustration from this answer.

\$\endgroup\$
1
\$\begingroup\$

Yes there is this is sometimes referred to as Application Specific Integrated circuit manufacture. ASIC

The manufacturer typically has a library of building blocks e.g. logic gates,memories, analog circuits which you use to build your design and then they will manufacture it.

This process however is not cheap. If you need just a few parts or to limit the Non Recurring Engineering costs you may want to consider FPGAs EPLDs or a microcontroller.

\$\endgroup\$
1
  • \$\begingroup\$ But what's expensive ? engineering costs or the creation of the mask itself ? for normal PCBs we pay something like $5K USD (excluding engineering), how much more expensive a mask for an IC would be ? \$\endgroup\$ Aug 18, 2020 at 17:29

Not the answer you're looking for? Browse other questions tagged or ask your own question.