5 ms·

> Six months ago it was a global chip shortage that had supply chains backed up around the planet. chip != chip. TSMC manufactures hundreds of products, check

> Six months ago it was a global chip shortage that had supply chains backed up around the planet. chip != chip. TSMC manufactures hundreds of products, check out slide 4 of the last quarterly results [1]. "AI" (on 5nm or below) is currently only a smaller share of total capacity. 7nm was only around 50% utilization, compared to 110% a year ago. The recent chip shortage was mostly from 28nm or above. Think automotive, and the like. The inventory correction happening now is from a strong reduction in smartphone business, and HPC is also down thanks to PC sales declining 10-15% this year. And there is nothing surprising about this, TSMC has been guiding this inventory correction for more than a year now. [1] https://investor.tsmc.com/english/encrypt/files/encrypt_file/reports/2023-07/aa7d1adbe692cf07f10aa1584c832facb43ee84b/2Q23Presentation%28E%29.pdf https://investor.tsmc.com/english/encrypt/files/encrypt_file...

addaon 3y ago
> The recent chip shortage was mostly from 28nm or above. As a reminder to others, 28 nm was the last planar process; there's a discontinuity in cost moving below 28 nm to FinFET processes. For applications which don't need the last bit of density of power efficiency, 28 nm is likely to stay a sweet spot for a long, long time to come.
- ls612 3y ago
  With 3nm being the last of the FinFETS will it end up in a similar situation in a decade? Will it be used for a lot of applications like 28nm is now for a long time?
  - addaon 3y ago
    It comes down to what the cost delta ends up being between FinFETs and GAAFETs, and if that's where the discontinuity is. I suspect that, even though EUV is "solved" now, the solutions are so expensive and sketchy that the DUV/EUV break is more likely to be important, with the final DUV processes being very long-lived. Unclear where this will end up exactly; I'd expect to see mature 7 nm DUV processes, but it's possible that the economics will either favor maturation of DUV + one or two EUV layers, or that a pure DUV 10 nm process with less multi-patterning ends up maturing so much cheaper that that wins out.
- kurthr 3y ago
  I hear there are developments of half-node reductions in 28nm processes to get down to "22nm" mostly with wiring density. But your point stands, it's the last planar node for mixed signal.
  - addaon 3y ago
    Yep, there will be continuing improvements to 28 nm as a long-lived node; 45 nm was a similarly long-lived node (as are 90 nm and 180 nm, still seeing plenty of use), and today's 40 nm is a heck of a lot more polished than 45 nm was at release, while still allowing trivial design porting.
    - govg 3y ago
      Could you mention where these chips from older mature nodes are used and what they look like(in terms of form factor / development environment), for context? I understand some aspects of chips being used for automobiles , but aren't most compute oriented chips based off ARM designs and such which will be on the newer nodes?
      - addaon 3y ago
        Most of the chips I interact with on a daily basis are 45 nm, some are 28 nm. Currently working automotive, same in aerospace. A lot of this is because embedded flash generally trails several years behind general availability of a node; so if you're designing a µC or similar with embedded flash, you're starting a few years behind; and then you're guaranteeing a 10 or 15 year supply chain, so on average a production part is on a ~10 year old node (but development activities are biased towards the front of this, since you often need 7 years of remaining supply chain guarantees after start of production). Another factor is that it simply takes time to validate nodes. Automotive nodes trail general use nodes; SEU data takes a few years to gather. Even for non-embedded-flash parts, this can cause a similar delay. As an example, the TI Hercules is an old but by no means obsolete safety processor. It's one of the best ways to get single-core lockstep capabilities for new designs today. The original TMS570LS parts are still built on a 130 nm process, but the "newer" (not new, but also newest) TMS570LC parts are on 65 nm. These days most of my work hours are focused on the Aurix TC3xx, which is a 40 nm (tweaked 45 nm) part. This is a multi-core safety processor with an obviously higher transistor count than the Hercules (which is why it was selected for application), matching its smaller process. Similarly in FPGA land, the Lattice MachXO is on an 130 nm process, while the newer MachXO2 migrated to 65 nm. I know Lattice does newer parts on 28 nm (and I'm sure there's 45 nm out there), but I haven't run into them. Certainly for more density-heavy applications newer processes (28 nm, often) have come to the forefront, and we're at the point that we're seeing cost cuts going from 90 nm to 28 nm for equivalent functionality. But a lot of these designs are pin-out limited (the silicon area is dominated by getting signals on and off chip, not by the total area of transistors), so cost very much doesn't scale with transistor density; the cost per transistor has gone down, but the cost per unit area has gone up.
merb 3y ago
But in 2020/2021 we had a chip shortage on 8nm. Of course demand stopped in end 2022/early 2023. of course it was because crypto got corrected downwards. You can clearly see that in the slides
- fomine3 3y ago
  8nm, aka Nvidia Ampere. Now it's being superseded by TSMC 4nm Nvidia Ada Lovelace.
gmerc 3y ago
TSMC is diversified, they are basically an index. Everything from cars to medical devices to household electronics to AI uses their chips. This is why Nvidia can soar while TSMC barely moves. We know winter is coming. The signs are all there, the desperate juicing and enshittification- growth is running out and nobody wants to blink first.

addaon 3y ago

> The recent chip shortage was mostly from 28nm or above. As a reminder to others, 28 nm was the last planar process; there's a discontinuity in cost moving below 28 nm to FinFET processes. For applications which don't need the last bit of density of power efficiency, 28 nm is likely to stay a sweet spot for a long, long time to come.

ls612 3y ago
With 3nm being the last of the FinFETS will it end up in a similar situation in a decade? Will it be used for a lot of applications like 28nm is now for a long time?
- addaon 3y ago
  It comes down to what the cost delta ends up being between FinFETs and GAAFETs, and if that's where the discontinuity is. I suspect that, even though EUV is "solved" now, the solutions are so expensive and sketchy that the DUV/EUV break is more likely to be important, with the final DUV processes being very long-lived. Unclear where this will end up exactly; I'd expect to see mature 7 nm DUV processes, but it's possible that the economics will either favor maturation of DUV + one or two EUV layers, or that a pure DUV 10 nm process with less multi-patterning ends up maturing so much cheaper that that wins out.
kurthr 3y ago
I hear there are developments of half-node reductions in 28nm processes to get down to "22nm" mostly with wiring density. But your point stands, it's the last planar node for mixed signal.
- addaon 3y ago
  Yep, there will be continuing improvements to 28 nm as a long-lived node; 45 nm was a similarly long-lived node (as are 90 nm and 180 nm, still seeing plenty of use), and today's 40 nm is a heck of a lot more polished than 45 nm was at release, while still allowing trivial design porting.
  - govg 3y ago
    Could you mention where these chips from older mature nodes are used and what they look like(in terms of form factor / development environment), for context? I understand some aspects of chips being used for automobiles , but aren't most compute oriented chips based off ARM designs and such which will be on the newer nodes?
    - addaon 3y ago
      Most of the chips I interact with on a daily basis are 45 nm, some are 28 nm. Currently working automotive, same in aerospace. A lot of this is because embedded flash generally trails several years behind general availability of a node; so if you're designing a µC or similar with embedded flash, you're starting a few years behind; and then you're guaranteeing a 10 or 15 year supply chain, so on average a production part is on a ~10 year old node (but development activities are biased towards the front of this, since you often need 7 years of remaining supply chain guarantees after start of production). Another factor is that it simply takes time to validate nodes. Automotive nodes trail general use nodes; SEU data takes a few years to gather. Even for non-embedded-flash parts, this can cause a similar delay. As an example, the TI Hercules is an old but by no means obsolete safety processor. It's one of the best ways to get single-core lockstep capabilities for new designs today. The original TMS570LS parts are still built on a 130 nm process, but the "newer" (not new, but also newest) TMS570LC parts are on 65 nm. These days most of my work hours are focused on the Aurix TC3xx, which is a 40 nm (tweaked 45 nm) part. This is a multi-core safety processor with an obviously higher transistor count than the Hercules (which is why it was selected for application), matching its smaller process. Similarly in FPGA land, the Lattice MachXO is on an 130 nm process, while the newer MachXO2 migrated to 65 nm. I know Lattice does newer parts on 28 nm (and I'm sure there's 45 nm out there), but I haven't run into them. Certainly for more density-heavy applications newer processes (28 nm, often) have come to the forefront, and we're at the point that we're seeing cost cuts going from 90 nm to 28 nm for equivalent functionality. But a lot of these designs are pin-out limited (the silicon area is dominated by getting signals on and off chip, not by the total area of transistors), so cost very much doesn't scale with transistor density; the cost per transistor has gone down, but the cost per unit area has gone up.

merb 3y ago

But in 2020/2021 we had a chip shortage on 8nm. Of course demand stopped in end 2022/early 2023. of course it was because crypto got corrected downwards. You can clearly see that in the slides

fomine3 3y ago
8nm, aka Nvidia Ampere. Now it's being superseded by TSMC 4nm Nvidia Ada Lovelace.

gmerc 3y ago

TSMC is diversified, they are basically an index. Everything from cars to medical devices to household electronics to AI uses their chips. This is why Nvidia can soar while TSMC barely moves. We know winter is coming. The signs are all there, the desperate juicing and enshittification- growth is running out and nobody wants to blink first.