3 ms·

I wonder: if my sentiment from https://news.ycombinator.com/item?id=41587866 https://news.ycombinator.com/item?id=41587866 is correct and Intel is essentially g

by ilyagr 2y ago

I wonder: if my sentiment from https://news.ycombinator.com/item?id=41587866 https://news.ycombinator.com/item?id=41587866 is correct and Intel is essentially giving up on innovation in chip fabrication, what, if anything, does this mean for AMD? I'm quite ignorant on how AMD works, e.g. do they even have a chip fabrication plants? If they do, can they compete with Intel's for US govt grants, or has that ship sailed now that Intel got a grant? If not, is there room for meaningful innovation in x86 chip design? How about TSMC? Do they now have a monopoly on state-of-the-art chip fabrication?

unsnap_biceps 2y ago
> is there room for meaningful innovation in x86 chip design? I think there is plenty of room for meaningful innovation. x86 won't ever hold the crown for performance per watt, but could keep building on just raw total performance and extremely wide vectors. Why stop at AVX10? Let's move to supporting 1024/2048 bits or even, eventually, get to full page size vector instructions at 4096 bytes!
alephnerd 2y ago
> I wonder: if my sentiment from https://news.ycombinator.com/item?id=41587866 https://news.ycombinator.com/item?id=41587866 is correct and Intel is essentially giving up on innovation in chip fabrication Ignore HN. Almost every commenter on here is well outside of their depth. Intel's been putting a significant amount of money into adopting High NA EUV in 18A, which is bleeding edge enough for a decade. Furthermore, even though TSMC and Fabrication have become hot words on HN recently, it has been a financial dead end for almost 15 years now. This is why TSMC, UMC, and Samsung cornered the chip fabrication market - the margins are very very low and the upfront costs are very very high. There are 3 core processes to bringing a chip to market - Design: high margins, low capex, this is where American players like Nvidia, AMD, Intel, Qualcomm, etc really shine as well as their Israeli and Indian offices. This is also basically software at this point, as almost all design is done via EDA/ECAD - Fabrication: Fabricating the aforementioned designed chips. low margins, high capex, commodified due to Taiwanese and South Korean industrial policy along with recent advances in China for 28nm processes, recently prioritized in the US due to COVID era chip shocks - Packaging and Testing: packaging fabricated chips. low margins, high capex, commodified due to Malaysian, Thai, Chinese, South Korean, Japanese industrial policy, recently prioritized in the US due to COVID era chips shock. Before the mid-2010s, American companies like AMD, Intel, and IBM Microelectronics used to do all 3, but the low margins of Fabrication and Packaging caused players like IBM Microelectronics and AMD to almost go bankrupt, forcing both of them to divest their fabrication and packaging capacity to a spinoff called GlobalFoundaries. Intel kept their packaging and fabrication capacity, but it has a negative impact on Intel's R&D capacity in the 2010s. Now with CHIPS Act money, Intel is modernizing it's fabrication and packaging offerings, and splitting them off into "Intel Foundary Services" while keeping design under "Intel". There are a lot more intricacies as well (bleeding edge nodes vs lagging nodes, analog chips, memory chips, the entire EDA tooling chain, etc) with entirely distinct market dynamics that deserve a paper unto themselves, but that's a long story and I really really hate HN.
- ilyagr 2y ago
  Thanks, this is interesting to read! If what you're saying is true, it's a bit sad that all these large-capex industries that are so important for society to have expertise and progress in are not considered promising or competitive.
  - alephnerd 2y ago
    Conversely, the money that was redeployed into higher margin industries had a net benefit for innovation. AMD EPYC would not have existed if AMD was stuck in the fabrication and packaging space because they wouldn't have had the money to spend building a dedicated design team, and Nvida's breakthroughs in GPUs would have never happened if Nvidia had to also fabricate and package. A company cannot have every single subsegment of an industry - it's unrealistic and a bad use of capital. And high capex low margins industries are bad for workers from a salary perspective. This is why CS became so popular in comparison to EE/CE - the starting salaries are easily 2x as a SWE compared to working as a Fab Engineer. And it's not a simple question of raising salaries. Becuase margins are very low (within the low double digits at most), a foreign competitor like TSMC, Samsung, and UMC can win by paying peanuts - which is what they did. Engineering salaries at those companies in Taiwan and SK are around $40-70k with 996 hours de facto enforced (and why Samsung employees globally are on strike right now)
magicalhippo 2y ago
> do they even have a chip fabrication plants They had, but divested[1] as they couldn't keep up. GloFo stopped chasing the cutting edge, but does have 12-14 nm capability thanks to deals with ST and Samsung. [1]: https://en.wikipedia.org/wiki/GlobalFoundries https://en.wikipedia.org/wiki/GlobalFoundries

by ilyagr 2y ago

unsnap_biceps 2y ago

> is there room for meaningful innovation in x86 chip design? I think there is plenty of room for meaningful innovation. x86 won't ever hold the crown for performance per watt, but could keep building on just raw total performance and extremely wide vectors. Why stop at AVX10? Let's move to supporting 1024/2048 bits or even, eventually, get to full page size vector instructions at 4096 bytes!

alephnerd 2y ago

> I wonder: if my sentiment from https://news.ycombinator.com/item?id=41587866 https://news.ycombinator.com/item?id=41587866 is correct and Intel is essentially giving up on innovation in chip fabrication Ignore HN. Almost every commenter on here is well outside of their depth. Intel's been putting a significant amount of money into adopting High NA EUV in 18A, which is bleeding edge enough for a decade. Furthermore, even though TSMC and Fabrication have become hot words on HN recently, it has been a financial dead end for almost 15 years now. This is why TSMC, UMC, and Samsung cornered the chip fabrication market - the margins are very very low and the upfront costs are very very high. There are 3 core processes to bringing a chip to market - Design: high margins, low capex, this is where American players like Nvidia, AMD, Intel, Qualcomm, etc really shine as well as their Israeli and Indian offices. This is also basically software at this point, as almost all design is done via EDA/ECAD - Fabrication: Fabricating the aforementioned designed chips. low margins, high capex, commodified due to Taiwanese and South Korean industrial policy along with recent advances in China for 28nm processes, recently prioritized in the US due to COVID era chip shocks - Packaging and Testing: packaging fabricated chips. low margins, high capex, commodified due to Malaysian, Thai, Chinese, South Korean, Japanese industrial policy, recently prioritized in the US due to COVID era chips shock. Before the mid-2010s, American companies like AMD, Intel, and IBM Microelectronics used to do all 3, but the low margins of Fabrication and Packaging caused players like IBM Microelectronics and AMD to almost go bankrupt, forcing both of them to divest their fabrication and packaging capacity to a spinoff called GlobalFoundaries. Intel kept their packaging and fabrication capacity, but it has a negative impact on Intel's R&D capacity in the 2010s. Now with CHIPS Act money, Intel is modernizing it's fabrication and packaging offerings, and splitting them off into "Intel Foundary Services" while keeping design under "Intel". There are a lot more intricacies as well (bleeding edge nodes vs lagging nodes, analog chips, memory chips, the entire EDA tooling chain, etc) with entirely distinct market dynamics that deserve a paper unto themselves, but that's a long story and I really really hate HN.

ilyagr 2y ago
Thanks, this is interesting to read! If what you're saying is true, it's a bit sad that all these large-capex industries that are so important for society to have expertise and progress in are not considered promising or competitive.
- alephnerd 2y ago
  Conversely, the money that was redeployed into higher margin industries had a net benefit for innovation. AMD EPYC would not have existed if AMD was stuck in the fabrication and packaging space because they wouldn't have had the money to spend building a dedicated design team, and Nvida's breakthroughs in GPUs would have never happened if Nvidia had to also fabricate and package. A company cannot have every single subsegment of an industry - it's unrealistic and a bad use of capital. And high capex low margins industries are bad for workers from a salary perspective. This is why CS became so popular in comparison to EE/CE - the starting salaries are easily 2x as a SWE compared to working as a Fab Engineer. And it's not a simple question of raising salaries. Becuase margins are very low (within the low double digits at most), a foreign competitor like TSMC, Samsung, and UMC can win by paying peanuts - which is what they did. Engineering salaries at those companies in Taiwan and SK are around $40-70k with 996 hours de facto enforced (and why Samsung employees globally are on strike right now)

magicalhippo 2y ago

> do they even have a chip fabrication plants They had, but divested[1] as they couldn't keep up. GloFo stopped chasing the cutting edge, but does have 12-14 nm capability thanks to deals with ST and Samsung. [1]: https://en.wikipedia.org/wiki/GlobalFoundries https://en.wikipedia.org/wiki/GlobalFoundries