You are currently viewing SemiWiki as a guest which gives you limited access to the site. To view blog comments and experience other SemiWiki features you must be a registered member. Registration is fast, simple, and absolutely free so please, join our community today!
That doesn’t change the point the Hskuo was making. Systems companies (besides Apple) weren’t making their own chips back in 2015. Google bought off the shelf SOCs for Pixel from merchant chip vendors. Cloud was not nearly as big as it is today and those folks were buying from AMD and mostly intel not making custom chips. GPUs were basically only used for specific highly parallelized HPC workloads/simulations and 3D rendering. It is something of a zero sum game Hist. Okay Amazon makes their own CPUs. Well that means AMD/intel sold less of them. The market didn’t grow or shrink from the perspective of wafer demand.
You also only counting your big sexy companies with their sexy products among the list of TSMC customers. As an example broadcomm is absolutely massive as is marvell. They make chips on N5 but N3 will be a little bit because they tend to wait for wafer prices to come down after wave 1 and wave 2 customers move onto the next shiny thing. Another example is smartTV chips (from the likes of QCOM/MTK/Samsung/LG) or NAND flash controllers (from silicon motion and the like) the volume on any one of those this year is almost assuredly more than the number of TPUs Google needs for the rest of the decade. Will those chips will one day be made on N3 today they are made on nodes like 16FF. Then there are the swarm of companies I have never heard of doing various analog/RF stuff.
With all of that said, I do agree that we are seeing more systems companies making chips. The same factors that made the fabless merchant chip companies even possible (semiconductor foundries, easily licensable IP, and modern EDA tools) are enabling the systems companies to get in on the fun. Throw in the increasing adoption and standardization around disaggregated design, RISC-V, and folks like ARM developing turn-key solutions that are even easier to integrate and the barriers of entry into chip design are becoming even lower. This lower barrier of entry will do the same thing to the fabless merchant chip vendors that they did to the IDM merchant chip vendors which they themselves did to the vertically integrated systems companies. And to be abundantly clear merchant chip vendors can and will still thrive. But they will do so with diminished importance as merchant chip vendors now need to ensure their chips are superior to what systems companies can throw together with off the shelf or even custom IPs (similar to how fabless companies need to have better products than their IDM counterparts to counteract the benefits of margin stacking).
This diminishing of the importance of the merchant chip vendors (be they fabless or pure-play IDMs) presents a unique opportunity only available to the foundries (be they pure-play or IDM-foundries) and the design enablement companies. An opportunity that I wonder if it will make pure-play IDMs move to the IDM-foundry model.
We are actually coming full circle. Systems companies used to have their own fabs and make custom chips. IDMs changed that and fabless changed it again. Systems companies are now taking back control of their silicon and it will in fact shrink the market for fabless companies but the ecosystem will benefit (EDA, IP, Foundry).
The systems companies I work with design chips differently. They are not on a shoestring budget. They can justify huge spends since they are selling high margin systems or services. They really are out designing the fabless companies like the fabless companies out designed the IDMs. Circle of life....
That doesn’t change the point the Hskuo was making. Systems companies (besides Apple) weren’t making their own chips back in 2015. Google bought off the shelf SOCs for Pixel from merchant chip vendors. Cloud was not nearly as big as it is today and those folks were buying from AMD and mostly intel not making custom chips. GPUs were basically only used for specific highly parallelized HPC workloads/simulations and 3D rendering. It is something of a zero sum game Hist. Okay Amazon makes their own CPUs. Well that means AMD/intel sold less of them. The market didn’t grow or shrink from the perspective of wafer demand.
You also only counting your big sexy companies with their sexy products among the list of TSMC customers. As an example broadcomm is absolutely massive as is marvell. They make chips on N5 but N3 will be a little bit because they tend to wait for wafer prices to come down after wave 1 and wave 2 customers move onto the next shiny thing. Another example is smartTV chips (from the likes of QCOM/MTK/Samsung/LG) or NAND flash controllers (from silicon motion and the like) the volume on any one of those this year is almost assuredly more than the number of TPUs Google needs for the rest of the decade. Will those chips will one day be made on N3 today they are made on nodes like 16FF. Then there are the swarm of companies I have never heard of doing various analog/RF stuff.
With all of that said, I do agree that we are seeing more systems companies making chips. The same factors that made the fabless merchant chip companies even possible (semiconductor foundries, easily licensable IP, and modern EDA tools) are enabling the systems companies to get in on the fun. Throw in the increasing adoption and standardization around disaggregated design, RISC-V, and folks like ARM developing turn-key solutions that are even easier to integrate and the barriers of entry into chip design are becoming even lower. This lower barrier of entry will do the same thing to the fabless merchant chip vendors that they did to the IDM merchant chip vendors which they themselves did to the vertically integrated systems companies. And to be abundantly clear merchant chip vendors can and will still thrive. But they will do so with diminished importance as merchant chip vendors now need to ensure their chips are superior to what systems companies can throw together with off the shelf or even custom IPs (similar to how fabless companies need to have better products than their IDM counterparts to counteract the benefits of margin stacking).
This diminishing of the importance of the merchant chip vendors (be they fabless or pure-play IDMs) presents a unique opportunity only available to the foundries (be they pure-play or IDM-foundries) and the design enablement companies. An opportunity that I wonder if it will make pure-play IDMs move to the IDM-foundry model.
I am a little bit confused. Aren't we talking about the same thing? I kept saying the number of companies who are using leading edge process nodes is not decreasing as some people predicted. Depending on how we defined the "leading edge", it's smaller or equal to 7nm at TSMC, do we see less or more companies using leading edge nodes in the past 5 or 10 years?
I am a little bit confused. Aren't we talking about the same thing? I kept saying the number of companies who are using leading edge process nodes is not decreasing as some people predicted.
On that specific note I would say it went down. As an example RF devices used to be on leading nodes but all else being equal smaller transistors give worse performance so those sorts of use cases eventually stopped moving to new nodes (at least as fast) as they hit a good sweet spot. And I have to assume there are far more of the various trailing edge merchant chip companies (TI, microchip, analog devices, broadcomm, lattice semiconductor, ST-M, infenion, Raytheon, etc etc) than systems companies with leading edge chip design.
Depending on how we defined the "leading edge", it's smaller or equal to 7nm at TSMC, do we see less or more companies using leading edge nodes in the past 5 or 10 years?
Personally I don’t know if I’d consider 7”nm” leading edge anymore (especially at year’s end when we finally get non Apple folks running products on N3E). N7 is 6 years old at this point. I don’t think anyone was calling TSMC 90nm leading edge back in 2011, or TSMC 28nm leading edge in 2018. To further support this line of reasoning TSMC is already shipping specialty versions of N7 family like N6RF and is on the verge of having MRAM and ultra low leakage versions available. Those sorts of offerings don’t really get added until fabs are deprecated and your wave 1/2 customers have already moved off to your newer offerings.
