Array
(
    [content] => 
    [params] => Array
        (
            [0] => /forum/index.php?threads/creating-a-foundry-idm.17384/
        )

    [addOns] => Array
        (
            [DL6/MLTP] => 13
            [Hampel/TimeZoneDebug] => 1000070
            [SV/ChangePostDate] => 2010200
            [SemiWiki/Newsletter] => 1000010
            [SemiWiki/WPMenu] => 1000010
            [SemiWiki/XPressExtend] => 1000010
            [ThemeHouse/XLink] => 1000970
            [ThemeHouse/XPress] => 1010570
            [XF] => 2021370
            [XFI] => 1050270
        )

    [wordpress] => /var/www/html
)

Creating a Foundry+IDM

What design related business would provide a foundry with the greatest benefit?

  • FPGAs

  • ASICs

  • Core IP/ISAs

  • A variety of off the shelf IP that customers can integrate into their designs

  • ARM SOCs

  • Other


Results are only viewable after voting.

nghanayem

Well-known member
Like Morris Chang, Robert Tsao also had a dream of creating a semiconductor foundry. Back when UMC was first founded though they had a harder time getting external contracts than TSMC due to the issues associated with IDM foundries and the small size of the fabless ecosystem at the time. As a result Mediatek was spun out and UMC became a pureplay foundry. This got me thinking about if UMC didn't spin out mediatek while continuing to focus on Tsao's dream of being a foundry, I wonder if we could have seen the Foundry+IDM model. In the typical IDM+foundry model the IDM's would historically sell excess capacity to fabless firms to ensure fabs had high utilization. In the modern day IDM+foundry seems more focused on achieving the necessary scale to continue justifying transistor scaling and amortize the ballooning R&D costs for new nodes.

For a hypothetical Foundry+IDM, foundry is the end goal rather than a means to an end/extra way to monetize the IDM's investments. In this situation IDM would be there to either initiate the yield learning curve, help ensure fab utilization is high, and or help build a better design/IP ecosystem. To do this while still preserving much of the spirt of pureplay foundries is difficult, but I have some ideas.

FPGAs are great products to initiate the yield learning curve due to their simplicity and linear benefit from node shrinks. It is also a small enough market that you might not step on customer's toes very much and can maybe bundle them in at cost with your foundries' PDKs.

ASICs have the benefit of being low volume enough to not steal much capacity and business from customers, but it might make many of those smaller ASIC customers avoid your foundry.

Licensing out core IP/ISAs could be an interesting way to jump start customer designs, and be a big value add over pureplay foundries (we will have to see how this shakes out for IFS to see how this works in practice though). Additionally this approach might be too expensive to do without also selling some of these CPUs yourself (potentially making you an IDM+foundry). You must also be cognisant that you don't force customers to use your ISA over what they want to use as this could push people away (intel custom foundry comes to mind here).

Expanding on the two points above maybe you could create a verity of designs and IP that get licensed out to customers. But once again this seems like a whole lot of work to put in without actually selling said chips.

With so many firms choosing to design ARM chips, maybe it makes sense to sell some ARM SOCs. This obviously risks burning many of your potential customers, but given how off the shelve many ARM designs are, maybe this isn't that big of a concern. If nothing else this approach would be great for improving yield learning and your design ecosystem for ARM chips.

While I am skeptical that in practice Foundry+IDM would be a competitive with the pureplay or IDM+foundry models, it was certainly interesting to think about. Any thoughts on how the IDM part would look and how big it would be for this hypothetical business would be appreciated.
 
Last edited:
Fortunately UMC spun off MediaTek and several other companies and many of them (including UMC) are strong players in their own market. I can't think about any successful case in the past 30 years based on a IDM+ Foundry or Foundry+IDM model. As long as an IDM and a foundry are sharing a single wallet, they will keep getting back to the same problems again and again. It's like a left pocket vs a right pocket on the same pant with only one wallet/purse.

Pure paly foundry industry can achieve what they are today because they don't try to confuse themselves and their customers.
 
Last edited:
Fortunately UMC spun off MediaTek and several other companies and many of them (including UMC) are strong players in their own market. I can't think about any successful case in the past 30 years based on a IDM+ Foundry or Foundry+IDM model. As long as an IDM and a foundry are sharing a single wallet, they will keep getting back to the same problems again and again. It's like a left pocket vs a right pocket on the same pant with only one wallet/purse.

Pure paly foundry industry can achieve what they are today because they don't try to confuse themselves and their customers.
I'm in a similar camp, although I suppose for what it is worth nobody has ever tried foundry first with a small IDM on the side whose sole purpose is to do the bleeding for the foundries' customers. People have only ever tried IDM with foundry as a side hustle (which I suppose Samsung logic and memory are pretty successful at). The fact nobody has ever tried it is what makes it the most interesting for me. Although as you said I think that model might always grow to the IDM side dominating because that is generally where the money is at, Or the foundry jettisoning the design side.

Obviously TSMC would never do this because their whole brand is that "we don't compete with our customers". But humor me and say they bought xilinx or altera. In theory this gives them a captive customer that they can use to accelerate yield learning and to help refine their PDKs/DRs. In our timeline TSMC kind of has Apple for this, but at the end of the day Apple will not buy wafers from TSMC until yields on their SOCs are excellent. In theory a Xilinx or Altera under TSMC's banner don't have to be profitable. As long as they improve the foundry even breaking even on their N-1 products is acceptable. My guess is something like broadwell wouldn't have happened had intel not been an IDM, and this begs the question would Skylake have done so well had there not been broadwell to fix 14nm? Same deal with Samsung's "early" versions of their nodes. Another benefit of the FPGA is like I mentioned they are low enough in demand that you could never fill a fab on just FPGAs (in theory preventing the capacity shafting that occured with IDM+foundries in the late 80s thru 2000s).

In Samsung's case if you forget about the rest of Samsung electronics and only look N or N+1, then Samsung foundry kind of looks like the above model with an ARM soc IDM attached to it. Almost all of their customers are other people making ARM socs. Would this be the case if Samsung used PPC in their socs? My guess would be no, because their process and PDKs are probably highly optimized for high density low power applications and ARM.

In another hypothetical what if Samsung foundry was spun out and had acquired say Nuvia? From there maybe they sell some of their own Nuvia chips, and licenced out their core designs to the likes of Qualcomm and Mediatek under the stipulation that it is made on Samsung nodes? This sort of situation would technically be pure play (if they only licensed rather than sold), would be a big differentiator, and it would make their nodes much stickier.

Like I said I don't think most of these ideas would work in practice, but it is interesting to wargame things out. I think it is also profitable to occasionally question the status quo/conventional wisdom, and learn the "why nots?".
 
I'm in a similar camp, although I suppose for what it is worth nobody has ever tried foundry first with a small IDM on the side whose sole purpose is to do the bleeding for the foundries' customers. People have only ever tried IDM with foundry as a side hustle (which I suppose Samsung logic and memory are pretty successful at). The fact nobody has ever tried it is what makes it the most interesting for me. Although as you said I think that model might always grow to the IDM side dominating because that is generally where the money is at, Or the foundry jettisoning the design side.

Obviously TSMC would never do this because their whole brand is that "we don't compete with our customers". But humor me and say they bought xilinx or altera. In theory this gives them a captive customer that they can use to accelerate yield learning and to help refine their PDKs/DRs. In our timeline TSMC kind of has Apple for this, but at the end of the day Apple will not buy wafers from TSMC until yields on their SOCs are excellent. In theory a Xilinx or Altera under TSMC's banner don't have to be profitable. As long as they improve the foundry even breaking even on their N-1 products is acceptable. My guess is something like broadwell wouldn't have happened had intel not been an IDM, and this begs the question would Skylake have done so well had there not been broadwell to fix 14nm? Same deal with Samsung's "early" versions of their nodes. Another benefit of the FPGA is like I mentioned they are low enough in demand that you could never fill a fab on just FPGAs (in theory preventing the capacity shafting that occured with IDM+foundries in the late 80s thru 2000s).

In Samsung's case if you forget about the rest of Samsung electronics and only look N or N+1, then Samsung foundry kind of looks like the above model with an ARM soc IDM attached to it. Almost all of their customers are other people making ARM socs. Would this be the case if Samsung used PPC in their socs? My guess would be no, because their process and PDKs are probably highly optimized for high density low power applications and ARM.

In another hypothetical what if Samsung foundry was spun out and had acquired say Nuvia? From there maybe they sell some of their own Nuvia chips, and licenced out their core designs to the likes of Qualcomm and Mediatek under the stipulation that it is made on Samsung nodes? This sort of situation would technically be pure play (if they only licensed rather than sold), would be a big differentiator, and it would make their nodes much stickier.

Like I said I don't think most of these ideas would work in practice, but it is interesting to wargame things out. I think it is also profitable to occasionally question the status quo/conventional wisdom, and learn the "why nots?".

One of the biggest challenges is how do we tell who is the winner from whatever segments of the semiconductor industry? Intel chose Altera and made it a captive customer. But Xilinx seems to be the best FPGA player and probably took in much more business than the Alter/Intel FPGA.

Pure play foundry's approach, such as TSMC's, is much simpler. It offers itself as everyone's foundry and let the market to decide which TSMC customer's product is the best for a particular application. I think this business model makes more sense and easier to execute.
 
To be fair Altera was always number 2 behind Xilinx (even before the intel merger). It seems like they went to intel to use their process advantage to try and become number one, but in the end that didn't work out that way.

In my head the foundry+IDM firm also views their only path to success as their own customers succeeding. I have to assume that pureplay firms have a sizeable internal design team to make test chips and debug the PDKs/DRs. But by the same junction it is hard to imagine that these smaller internal teams can catch everything that someone making a real product might run into. As we saw at most 28nm nodes, 16FFL, and N3 when things are ramping slow it is real hard to get DD down with how few data turns you have. Heck if memory serves it wasn't all that long ago TSMC would take a loss on new nodes just to get the yield learning curve moving. For these sorts of situations having a more advanced design team might allow you to have a better pipe cleaner, and smoother ramp. When you put in all of that effort it almost feels like a waste not to monetize it in some way though (be it licensing the IP or selling chips), hence the idea. Unfortunately there is nobody left to try it. There is only one leading edge pure play foundry left, and two IDM+foundries where foundry is kind of just a moonlighting gig.
 
Back
Top