Root scientific causes of TSMC's leadership / Intel and Samsung's failure

[Portland]

Intel struggles with the trends, they're still trying to catch up with gpus when dpus exist.

 

[benb]

Few more random thoughts, science starts with a hypothesis informed by interesting observations, IMHO.
TSMC dominated and still dominates the last planar node, 28nm, which is now the least costly and most attractive for many applications, and they’re building out new fabs for it. Intel and Samsung don’t have cash cows like this.
The thing that happened to Intel, at 14nm, happened to TI, at 90nm. Same thing at Global Foundries at 7nm. Same thing at Motorola/Freescale at sub-130nm.
What if semiconductor businesses age like people, and slow down and become less competitive with the youths in their 40s-50s? It hit the oldest semi businesses first and TSMC will be the last.
China completely new semiconductor organizations (Yangtze Memory, etc) may be seen as the youths today.

 

[semiman]

I was more expecting technological decisions in companies(for example, how cobalt and COAG affected Intel 10nm), but somehow people are talking about cultures.

 

[benb]

Why all the discussion of culture, rather than decision A led to outcome A, etc? Culture is upstream of the decisions and outcomes. It also informs resource allocation.
Samsung memory business is 10x larger than Foundry, so the culture is to develop Foundry processes with resources “borrowed” from Memory. There is no dedicated R&D for Foundry. Samsung got EUV DRAM in production first, that was more important than progressing EUV Foundry nodes to keep up with Intel and TSMC.
Likewise, Intel chip design competes with (and overwhelms, I think) resources for TD. Intel chip designs for graphics are coming to the market soon, after a decade of work, and that may explain the decade of slow progress with TD.

 

[tooLongInEDA]

I was more expecting technological decisions in companies(for example, how cobalt and COAG affected Intel 10nm), but somehow people are talking about cultures.

Perhaps they both matter. And may be connected.

I'm not a process expert, but my sense is that Intel pushed for too much at 10nm - the area shrink they quoted from 14nm was (I think) something like 2.6x instead of the usual 2.0x for a node shrink. Perhaps they just took too many risks/short cuts.

 

[Xebec]

Why all the discussion of culture, rather than decision A led to outcome A, etc? Culture is upstream of the decisions and outcomes. It also informs resource allocation.
Samsung memory business is 10x larger than Foundry, so the culture is to develop Foundry processes with resources “borrowed” from Memory. There is no dedicated R&D for Foundry. Samsung got EUV DRAM in production first, that was more important than progressing EUV Foundry nodes to keep up with Intel and TSMC.
Likewise, Intel chip design competes with (and overwhelms, I think) resources for TD. Intel chip designs for graphics are coming to the market soon, after a decade of work, and that may explain the decade of slow progress with TD.

The problem here is there are too many variables that led to this situation.

One could argue AMDs choice to pursue a bulldozer type architecture (losing ground to Intel as a practice, per Jim Keller interviews) led to Intel making choices that let them fall behind TSMC on the node race (for now).

You can also argue the recovery after the 2008 global crisis and mass migration (finally) to hyper scale clouds caused Intel to focus only on certain portions of their business at the expense of others.. cash cows are great short term, but long term they tend to make you focus on the cash and nothing else. (See GE Aerospace and Lockheed Martins successor Space Systems get complacent and lose ground to SpaceX and others).

Obviously there is also just the view that TSMCs decision to take a more incremental approach meant long term success... However Intel had something like this for years (copy exactly and Tick Tock) but it broke down at some point did to various reasons.

That Intel ran into cobalt related problems at 10nm that took a really long time to resolve is most likely an effect from other causes rather than a major cause unto itself.