6 ms·

I'm usually a big fan of stratechery, but for once I disagree with the explanation of Intel's decline. I think it's much simpler: being number one in fab techn

I'm usually a big fan of stratechery, but for once I disagree with the explanation of Intel's decline. I think it's much simpler: being number one in fab technology is mostly driven by volume: whoever has the most volume wins. During the PC era, Intel had the most volume, so they also had the best fabs. It's a winner-take-all, self-reinforcing ecosystem. But in the last 10 years, mobile has had the highest volume. That's what allows TSMC to have the best fabs, and Intel is struggling. That's all there is to it. Nothing to do with services or software.

klelatti 4y ago
The CISC vs RISC point Ben makes is key here. Intel could have been a big player in mobile: it put too much focus on x86 though when it could have been making (its own or others) Arm based SoCs. Edit: not that Arm is necessarily RISC - but it certainly takes an approach that is different to x86 and better suited to mobile.
- marcosdumay 4y ago
  Intel tried to make Arm SoCs. But their customer management sucked, so in the end they only manufactured the SoCs that they designed, and didn't sell enough of them to be sustainable.
- caycep 4y ago
  Also perhaps there's a role where software/architecture just got better - they didn't have LLVM or other forms of code translation (if I'm phrasing that right) back then which seems to make the cpu architecture less of a critical issue.
- jjtheblunt 4y ago
  Intel was making its own ARM chips called Xscale 20 years ago, no?
  - klelatti 4y ago
    Indeed and XScale was successor to StrongARM which I think was highest performing Arm design for a while. Sold to Marvell in 2006 right before the iPhone appeared. Intel then spent billions trying to make x86 commercially viable on mobile.
- reindeerer 4y ago
  I've lost count how many misguided bids of taking x86 to low-power they made. Also the apparent unshakeable confidence that ARM will never become viable on server and desktop.
  - klelatti 4y ago
    In this context the Gelsinger video clip in the article is fascinating. Sometimes companies fixate on one element of their strategy that has worked in the past. It seems that Intel had ‘always make x86 / backwards compatible’ as their fixation.
aidenn0 4y ago
That is one of several reasons proposed in TFA; in particular TFA mentions TSMC's earlier investment in EUV being due to the guaranteed volume of smartphone chips.
- Throwawayaerlei 4y ago
  Did not read TFA and if it doesn't focus a lot on the 10 nm/Intel 7 botch it's not fine, but I've read TSMC's very roughly equivalent 7 nm node was also non-EUV but less aggressive than Intel's. And it succeeded unlike Intel, which then made it natural to start introducing in production EUV in their 7+ node. Here I suspect gaining real experience in chips delivered to customers trumps whatever volume you'll end up doing.
- sanxiyn 4y ago
  I very much doubt this. When EUV decision was made Intel was at the top of the world and had more volume than TSMC. As I understand Intel's reluctance was mostly about doubt whether ASML could deliver, and while ASML delivered I must say, even in retrospect, Intel's doubt was justified.
adrian_b 4y ago
This would the correct explanation if Intel would have made no mistakes. However, it is beyond reasonable doubt that Intel has made some extremely ugly management mistakes during their failed attempts to transition from 14 nm to 10 nm, which have absolutely nothing to do with the competition from TSMC or others. Maybe Intel would have lost their technological advance regardless how the company had been managed, due to the difference in volume between smartphone chips and PC chips, but now we cannot know which has mattered more, the internal mismanagement or the change of the semiconductor device market favoring their competition. I wonder if it will ever become public knowledge which were exactly the technology problems of Intel and who are those guilty for the decline at the company, because the failures could not have had any real technical causes but only bad management causes. During their worst years Intel has been capable even of presenting to the public fake technical presentations about the alleged great characteristics of their future 10 nm CMOS process, which is something that I would never have expected from a company like Intel. The marketing messages from any company are expected to be full of lies, but the technical presentations are expected to match reality. During many years, unlike TSMC, Intel has behaved as if it were completely unable to predict the precise characteristics of the CMOS process that they will able to manufacture in the following year. How could this happen is very hard to understand, as the manufacturing processes and the future devices made with them can be simulated long in advance of their implementation, and the models can be tuned continuously while the processes evolve, by fabricating and measuring various test structures. Unlike for Intel, for TSMC the performance predictions for future processes have always been reasonably accurate.
- klelatti 4y ago
  There was the infamous demo when they ‘forgot’ to mention a CPU was overclocked. Bit of a red flag for the culture that permits that. https://www.tomshardware.com/news/intel-28-core-cpu-5ghz,37244.html https://www.tomshardware.com/news/intel-28-core-cpu-5ghz,372...
- sanxiyn 4y ago
  > I wonder if it will ever become public knowledge which were exactly the technology problems of Intel and who are those guilty for the decline at the company, because the failures could not have had any real technical causes but only bad management causes. Why do you think this? It was a yield issue, and the best guess is it was caused by multi patterning. (Intel themselves said so.) I found an analysis from SemiWiki titled "Intel 10nm Yield Issues" pretty convincing. See https://semiwiki.com/semiconductor-manufacturers/intel/7433-intel-10nm-yield-issues/ https://semiwiki.com/semiconductor-manufacturers/intel/7433-...
babypuncher 4y ago
Which is probably why Pat Gelsinger wants to turn Intel into a fab for hire just like TSMC and Samsung.

klelatti 4y ago

The CISC vs RISC point Ben makes is key here. Intel could have been a big player in mobile: it put too much focus on x86 though when it could have been making (its own or others) Arm based SoCs. Edit: not that Arm is necessarily RISC - but it certainly takes an approach that is different to x86 and better suited to mobile.

marcosdumay 4y ago
Intel tried to make Arm SoCs. But their customer management sucked, so in the end they only manufactured the SoCs that they designed, and didn't sell enough of them to be sustainable.
caycep 4y ago
Also perhaps there's a role where software/architecture just got better - they didn't have LLVM or other forms of code translation (if I'm phrasing that right) back then which seems to make the cpu architecture less of a critical issue.
jjtheblunt 4y ago
Intel was making its own ARM chips called Xscale 20 years ago, no?
- klelatti 4y ago
  Indeed and XScale was successor to StrongARM which I think was highest performing Arm design for a while. Sold to Marvell in 2006 right before the iPhone appeared. Intel then spent billions trying to make x86 commercially viable on mobile.
reindeerer 4y ago
I've lost count how many misguided bids of taking x86 to low-power they made. Also the apparent unshakeable confidence that ARM will never become viable on server and desktop.
- klelatti 4y ago
  In this context the Gelsinger video clip in the article is fascinating. Sometimes companies fixate on one element of their strategy that has worked in the past. It seems that Intel had ‘always make x86 / backwards compatible’ as their fixation.

aidenn0 4y ago

That is one of several reasons proposed in TFA; in particular TFA mentions TSMC's earlier investment in EUV being due to the guaranteed volume of smartphone chips.

Throwawayaerlei 4y ago
Did not read TFA and if it doesn't focus a lot on the 10 nm/Intel 7 botch it's not fine, but I've read TSMC's very roughly equivalent 7 nm node was also non-EUV but less aggressive than Intel's. And it succeeded unlike Intel, which then made it natural to start introducing in production EUV in their 7+ node. Here I suspect gaining real experience in chips delivered to customers trumps whatever volume you'll end up doing.
sanxiyn 4y ago
I very much doubt this. When EUV decision was made Intel was at the top of the world and had more volume than TSMC. As I understand Intel's reluctance was mostly about doubt whether ASML could deliver, and while ASML delivered I must say, even in retrospect, Intel's doubt was justified.

adrian_b 4y ago

This would the correct explanation if Intel would have made no mistakes. However, it is beyond reasonable doubt that Intel has made some extremely ugly management mistakes during their failed attempts to transition from 14 nm to 10 nm, which have absolutely nothing to do with the competition from TSMC or others. Maybe Intel would have lost their technological advance regardless how the company had been managed, due to the difference in volume between smartphone chips and PC chips, but now we cannot know which has mattered more, the internal mismanagement or the change of the semiconductor device market favoring their competition. I wonder if it will ever become public knowledge which were exactly the technology problems of Intel and who are those guilty for the decline at the company, because the failures could not have had any real technical causes but only bad management causes. During their worst years Intel has been capable even of presenting to the public fake technical presentations about the alleged great characteristics of their future 10 nm CMOS process, which is something that I would never have expected from a company like Intel. The marketing messages from any company are expected to be full of lies, but the technical presentations are expected to match reality. During many years, unlike TSMC, Intel has behaved as if it were completely unable to predict the precise characteristics of the CMOS process that they will able to manufacture in the following year. How could this happen is very hard to understand, as the manufacturing processes and the future devices made with them can be simulated long in advance of their implementation, and the models can be tuned continuously while the processes evolve, by fabricating and measuring various test structures. Unlike for Intel, for TSMC the performance predictions for future processes have always been reasonably accurate.

klelatti 4y ago
There was the infamous demo when they ‘forgot’ to mention a CPU was overclocked. Bit of a red flag for the culture that permits that. https://www.tomshardware.com/news/intel-28-core-cpu-5ghz,37244.html https://www.tomshardware.com/news/intel-28-core-cpu-5ghz,372...
sanxiyn 4y ago
> I wonder if it will ever become public knowledge which were exactly the technology problems of Intel and who are those guilty for the decline at the company, because the failures could not have had any real technical causes but only bad management causes. Why do you think this? It was a yield issue, and the best guess is it was caused by multi patterning. (Intel themselves said so.) I found an analysis from SemiWiki titled "Intel 10nm Yield Issues" pretty convincing. See https://semiwiki.com/semiconductor-manufacturers/intel/7433-intel-10nm-yield-issues/ https://semiwiki.com/semiconductor-manufacturers/intel/7433-...

babypuncher 4y ago

Which is probably why Pat Gelsinger wants to turn Intel into a fab for hire just like TSMC and Samsung.