Two More Chinese DRAM Fabs Ready
anandtech.com
anandtech.com
Ouch, that's an aggressive schedule. Two months to qualify the equipment and the process, train technicians, establish PM schedules, ramp up testing and QA departments, and run pilot lots? If anything at all goes wrong -- one bad mask, one misaligned robot that scratches all the wafers, one misconfigured test program -- you lose three weeks and have to start over. More if you have to revise a mask. Even more if you don't discover the problem until after burn-in. If they hit volume production in the third quarter, it's going to be at the very end of September, and it's going to be a very generous definition of "volume."
[1] https://www.reuters.com/article/us-usa-trade-china-chips/chi...
Nothing today can throw any country back to a stone age than loss of access to microelectronics:
1. Loss of powerplants and the grid infrastructure - possible to rebuild, and immediate impact is not catastrophic
2. Loss of food supply stocks - unless, you are blockaded, you can just buy abroad
3. Loss of fertilizer stocks - just as survivable as loss of food supply.
4. Loss of hydrocarbon supply - an even more laughable national emergency.
5. Heavy industry - just rebuild, you can wholly rebuild it with pre-industrial tech level.
6. Loss of microelectronics - you fc*ked - to make microchips, you need more microchips, and to build those microchips, you need a ton of other microchips. You achieve nothing if you manage to replicate the technology of the era of first ICs.
domestic consumption is also at a crazy level, e.g. in Q1 2018, smartphone sales in China had a "hard landing" down 21% YoY, still 91 million phones were sold in Q1 2018 [1].
[1] https://www.canalys.com/newsroom/chinese-smartphone-market-s...
This cycle was why Grove and Moore got Intel out of the memory business in 1985.
I'm referring to the DRAM fixing scandal. All of the DRAM makers colluded to fix prices. We had one DRAM maker a year for like 6 or 7 years plead guilty to collusion and price fixing under the Sherman Anti Trust Act.
Since then prices haven't been very volatile, they just sort of steadily rise year on.
Back about 35 years ago, perhaps for the 256 kbit DRAM generation, IIRC it was Micron that did something like this. I tried some quick Googling and didn't find anything. But it's really hit and miss how much stuff that precedes the WWW can be turned up by a quick search.
The Micron idea was that they would use a (12,8) Hamming type code. A total of 12 bits including 8 data bits. Which, if you recall the theory behind these codes, meant that you could do Single Error Correction, but not Double Error Detection (that would take a 13th bit). Not doing DED was a "feature", since these chips were pin compatible with standard memory. So you had no easy way to signal an error anyway.
I don't think the product was very successful (if it ever made it into volume production) because of the incredible overhead it would take. Instead of 262144 total memory bits on the die, you needed 50% more, i.e. 393216.
On the plus side, you could make each cell a little bit smaller, and you wouldn't really care about a few individual bad bits (those would be hidden by the ECC). So your yield probably went up. But of course the die was so much bigger, so your yield went way down because of that. Tradeoffs, tradeoffs, tradeoffs!
Overall, if you were Micron I suppose you were just happy if the finished chip met some appropriate AQL standard. What it did on die was just an implementation detail. :)
Rather than getting two more fabs, Chinese state funds have just launched two more barely working ones that add to a long long list of their failed attempts to achieve self sufficiency in memory supply.
There were numerous attempts by Xinhua University Group in the past to launch memory businesses. Judging by that those fabs stand dark for years, none of them succeeded, and a lot of taxpayers money just got wasted again.
2. The money involved before were tiny compared to this round.
3. They now have a market to absorb some lower quality components.
4. The market is now making so much money from NAND and DRAM that is unseen in past time. There is not better time to enter the market. Even at half the yield then other major competitor they would still be able to make money.
5. I don't think they will catch up in 3D NAND any time soon. But DDR4 seems to be a easier target.
I regularly meet Unigroup people just below the C-suite level during industry events in China. The do not give off any sense of brilliance to me, nor I can say anything high of their ability from what they did before.
Sure, they are probably the top fund by concentration of Oxford and Ivy league grads in China, and have few returnee cadres with past stellar career in Western top tier investment funds. I can even go as far as to say that if judged individually, some can be said to be genuine talents.
With all that star crew, why don't they fly? Because they are dominated people with stellar credentials, but all of whom are just meh and so so...
Just look what their "genius stratagem" of joining RDA and Spreadtrum did: they took a leader in one market, a leader on another market, joined them and got one without lead in any.
When Qualcomm was still greatly greatly handicapped in Chinese market by the fact that they were limited to high-end as low-to-mid-end market was a shark tank with likes of Spreadtrum and RDA in their prime. But when Unigroup decided to forcefully uproot both of them from their power base - domestic mass market, and mandate them to "abandon it all, focus on high-end, and make a chip to nail Qualcomm in under a year" they both lost their grip on their market niches, which Qualcomm and Mediatek were happy to devour, and of course they never came even close to scoring anything in high end market that Unigroup wanted. Now they have neither low-end lead, nor high-end lead.
As the saying goes: you don't get iron by joining brass and bronze.
Let just say marking DRAM isn't rocket science. The key really is yield. And there is going to be some learning curve involves. UniIC actually posted a picture last year in December about their success on DDR4, but it was quickly deleted because the yield was only in single digit, i.e really bad, the info was shared on Weibo and known internally within the industry but it was deleted.
In Feb when Anandtech posted the news, I thought it was hoax again. Turns out UniIC actually did post a PR and yield is now 30%+. Which to be honest is still crap, but at least in today's market value, they could still break even. The more trial and error they do, the quicker they will be to up its yield. Something the market has allow them to do, if it was any previous years, the lost would be too great no one will be willing to "shoulder" this lost and lots of work will be needed to save faces.
I don't think Unigroup will mess up again, unless they decide to focus on something like HBM or LPDDR4, otherwise I think there is going to be sizeable low end market within China using UniC. I don't expect them to complete in the high end, but their entry I hope will put a stop in the memory pricing. Which is like at a ridiculous level.
P.S - UniGroup also has a change a direction and new objective under Xi, no longer are they acting mindlessly. Which is both a good and bad thing.
https://en.wikipedia.org/wiki/UMC_Green_CPU
Of course, VIA is still around too, as is this:
That is the exact non-sense talked by many people 10-15 years ago regarding BOE and its state backed investment. Look at what BOE achieved in mid/long term.
EDIT: If random quora answers are to be trusted, taxes are about 90% of government funding: https://www.quora.com/How-does-the-Chinese-government-fund-i...