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!
I think many here will be interested in this news too. I haven't seen it posted yet so I share it here.
I am slightly sceptic though, why all of a sudden has TSMC decided to change its mind, Morris Cheng has always said he will never enter DRAM and NAND business. All while Samsung breaks record after record in NAND / DRAM profits.
There may just be someone with a little too much money, has no idea and headache how to spend it, may wants to lend to TSMC for extremely low interest rate in exchange for stable supply and pricing of NAND and DRAM. But if that was really the case, why they haven't done it sooner?
From following TSM for years some possiblities are that their work on Crossbar phase change memory for at least five years that I know of isn't proving viable. That would leave Micron with 3dXpoint, which they have split with Intel, but might have an understanding that MU has buy out rights. Also it looks like closer links between memory and processing are the future game that everyone will have to participate in or suffer the consequences. TSM could become a major customer offering tightly integrated memory and processing with TSM's advanced packaging technologies. With Apple behind TSM and sitting on a cash pile, nothing is out of reach. TSM's last fab at 26 billion I believe is just one example of the financtial muscle at play here and that is after spending almost 17 billion on a fab just a few years ago. This could be just one phase of what may be TSM's drive to dominate the high end semi business.
(a) Competing in commodity memory (ie DRAM or flash) makes little sense to me. TSMC has no special skills in these areas, and there's no great synergy.
(b) How about Optane and Optane wannabe's? To me the argument of (a) still holds, with the additional caveat that, so far anyway, these seem to me much more relevant to enterprise than to personal use cases. (There just isn't a compelling use case that I see for personal uses that justifies the cost/performance tradeoff. Maybe I'm wrong?) Meaning that I can't see TSMC's tier one customers like Apple and QC really caring about this.
(c) What IS interesting is "emerging" memory technologies. On the one hand we have wonderboy Nantero, and who knows how that will turn out. I expect everyone and his dog is talking to them...
On the other hand (and this is the case that really makes sense, IMHO) we have MRAM. This DOES look like it could have various interesting applications for personal devices, either on-SoC (larger LLCs? perhaps the persistence on an MRAM LLC can be used for fast rebooting? or perhaps replacing the DRAM cache of Flash? or provide persistent SoC storage for IoT devices, so no need for any external persistent storage?)
So that would be my guess, to the extent that there's any truth to this claim -- it's a discussion with some company about some form of MRAM, with the goal being to provide it as one more useful block on a TSMC SoC.
Their wafer bonding technology also allows them to set up a modest amount of DRAM fab aimed at chips which offer dense "local" memory at low latency and high bandwidth through 3D. This could be expected to offer a premium and also does not need an extravagant number of wafers churned out. They may have some technology adders relative to the 3D which form a moat that the commodity producers do not find attractive.
The DRAM producers have very sophisticated processes that are applied to huge volumes of identical wafers to keep costs down. A modern server may have 10x the sq cm of DRAM in it than all logic chips put together, so that is a scale, uniformity, and process very different than what TSMC has mastered.
Remember Apple already has some sort of stake/control in a memory manufacturer, via their part in the Bain consortium that bought Toshiba Memory three months ago. It seems unlikely that Apple wants TSMC to go against Toshiba, or sees any value in doing so. That's one reason my bet is on a non-traditional memory, not more DRAM or flash.
Remember Apple already has some sort of stake/control in a memory manufacturer, via their part in the Bain consortium that bought Toshiba Memory three months ago. It seems unlikely that Apple wants TSMC to go against Toshiba, or sees any value in doing so. That's one reason my bet is on a non-traditional memory, not more DRAM or flash.
Nanya, Powerchip, Windbond? TSMC could easily take over Taiwan DRAM production and make it profitable. Would China rather buy DRAM from the US, Korea, or Taiwan? Who would Apple rather buy DRAM from?
Remember Apple already has some sort of stake/control in a memory manufacturer, via their part in the Bain consortium that bought Toshiba Memory three months ago. It seems unlikely that Apple wants TSMC to go against Toshiba, or sees any value in doing so. That's one reason my bet is on a non-traditional memory, not more DRAM or flash.
In 2017, TSMC at one point was interested to bid on the Toshiba's memory division. They didn't continue the process due to 1. Heavy price competition from other bidders. 2. Lack of synergy. 3. Japanese government's objections.
So that's obvious TSMC has some strategical thinking about getting into certain type of memory business or wants to apply/integrate their process technologies with memory.
In TSMC's game plan, I think Apple and Foxconn are two important factors. Actually Foxconn was TSMC's partner during their bid on Toshiba memory division last year. And we know both Foxconn and TSMC collaborate with Apple deeply.
(a) Competing in commodity memory (ie DRAM or flash) makes little sense to me. TSMC has no special skills in these areas, and there's no great synergy.
On the other hand (and this is the case that really makes sense, IMHO) we have MRAM. This DOES look like it could have various interesting applications for personal devices, either on-SoC (larger LLCs? perhaps the persistence on an MRAM LLC can be used for fast rebooting? or perhaps replacing the DRAM cache of Flash? or provide persistent SoC storage for IoT devices, so no need for any external persistent storage?)
So that would be my guess, to the extent that there's any truth to this claim -- it's a discussion with some company about some form of MRAM, with the goal being to provide it as one more useful block on a TSMC SoC.
Nanya, Powerchip, Windbond? TSMC could easily take over Taiwan DRAM production and make it profitable. Would China rather buy DRAM from the US, Korea, or Taiwan? Who would Apple rather buy DRAM from?
Nanya and Winbond already supply products which could be disrupted by the acquisition. I believe Powerchip is a memory-inclusive foundry, which might be interesting to TSMC, but the scale is nowhere near the likes of Samsung for example. There is also some new technology which would have to be ported over such as buried word line, and high aspect-ratio capacitor.
I think the biggest reason not to go into DRAM or Flash is an extra player makes things worse for everyone (supplier side, that is).
Isn't MRAM at the point where (unlike DRAM or flash) what you're buying when buy an MRAM company is basically the expertise to do some part of the job better? TSMC may "have" MRAM today, but I can well believe that, in a space so new and unoptimized, there are a number of companies out there each with one small optimization of the MRAM concept; and there's plenty of value in aggregating enough of those together in one place. (Along with, of course, acquiring along the way the people that were responsible for those optimizations, and who can consider how best to improve the suite of ideas so far acquired.)
Isn't MRAM at the point where (unlike DRAM or flash) what you're buying when buy an MRAM company is basically the expertise to do some part of the job better? TSMC may "have" MRAM today, but I can well believe that, in a space so new and unoptimized, there are a number of companies out there each with one small optimization of the MRAM concept; and there's plenty of value in aggregating enough of those together in one place. (Along with, of course, acquiring along the way the people that were responsible for those optimizations, and who can consider how best to improve the suite of ideas so far acquired.)
May be, while we saw sharp decline in NAND price, but all the NAND supplier are still making substantial profits at this level. It also has a very clear roadmap of 8x to 16x capacity increase by ~2025 with die shrink and die stacking. DRAM doesn't enjoy this capacity increase, and with more machine learning it seems we have infinite appetite for In Memory computing that current DRAM market enjoys. I look up the price level of DRAM from 2016, we are anywhere from 2x to 3x increase, and if we move further back to its lows, we are close to 4x. These price level is what makes Samsung the most profitable company leaping ahead of Apple in two quarters.
Given the yield and node improvement, it is likely even if DRAM price drop 50% there are still heavy profits to be made. And I doubt the entry of TSMC into memory business will be the reason for this drop. And May be TSMC is just preparing for DRAM business incase China cant buy any memory from Japan US or Korea.
May be, while we saw sharp decline in NAND price, but all the NAND supplier are still making substantial profits at this level. It also has a very clear roadmap of 8x to 16x capacity increase by ~2025 with die shrink and die stacking. DRAM doesn't enjoy this capacity increase, and with more machine learning it seems we have infinite appetite for In Memory computing that current DRAM market enjoys. I look up the price level of DRAM from 2016, we are anywhere from 2x to 3x increase, and if we move further back to its lows, we are close to 4x. These price level is what makes Samsung the most profitable company leaping ahead of Apple in two quarters.
Given the yield and node improvement, it is likely even if DRAM price drop 50% there are still heavy profits to be made. And I doubt the entry of TSMC into memory business will be the reason for this drop. And May be TSMC is just preparing for DRAM business incase China cant buy any memory from Japan US or Korea.
Flash and DRAM are good business currently but it's always mainly three or four players in each. For DRAM 46% Samsung, 29% SK Hynix, 21% Micron, 3-4% being Nanya, Winbond and Powerchip combined. For NAND 37% Samsung, 35% Toshiba/WD, 12% Micron, 10% SK Hynix, 6% Intel. The revenues of the Taiwan DRAM companies are no match for what TSMC makes currently, it would need a plan to knock out Micron or SK Hynix to make it worthwhile for itself. The process technologies (between logic and DRAM or 3D NAND) also do not overlap so there is significant adjustment needed there as well. China seems to be heading toward developing homegrown Flash and DRAM companies. If it considered TSMC its own, the government would be helping it for sure rather than keep it in Taiwan.
The major question is will the relationship between memory and processing structure change the whole game with something like Automata or an even more radical structure? TSM with their wide range of technologies and backing of Apple could change the whole processing/memory game in one bold move and they definitely have the skill, resources, equipment and above all the strongest veil of secrecy needed to pull this off and surprise everyone. Any thoughts or comments on this appreciated.
