GPUs have become cabbages.
Chinese semiconductor industry
- Thread starter Hendrik_2000
- Start date
- Status
Assembly in China is vastly different from assembly in Vietnam/India. Once the chip is made, it is still useless before packaging, assembly into a PCB and connection with non-logic electronic components like displays, touchscreens, batteries or power supplies, external memory, power MOSFET, etc. Vietnam/India don't have the capability to package dies into complete packages effectively nor do they have the capability to make components like displays, batteries, power supplies, etc.
Samsung Vietnam still has to fly components in from China for assembly. Samsung no longer has any phone assembly in China, but are expanding their Xi'an fab which produces 40% of its memory chips.
I'm probably going to sound like an idiot, but entertain me here. So, let's say they have 1 line now and 3 lines in a couple of years that can produce 14nm, N+1 and N+2. Let's say with 14nm, they can do 600k wafers a year. n+1, they can do 400k wafers. n+2, they can do 200k wafers.
My impression is that the die size of:
smartphone CPU around 100m^2
desktop CPU around 200m^2
server CPU around 500m^2
AI GPU around 1050m^2
so each wafer can produce maybe 60 AI GPUs, 120 server CPUs, 300 desktop CPUs, 600 smartphone CPUs.
let's say yield is 80% on 28 nm, 60% on 14 nm, 40% on N+1, 25% on N+2 (am I overestimating? DUVi have improved so I assume they can get higher yield than what TSMC had).
Let's say you want to use N+1 process to produce your next generation desktop chip for Phytium and Zhaoxin..... And the market size is 10 million a year (let's say that's the demand for non-foreign CPUs right now)
each wafer on N+1 process would yield 120 CPUs -> you need 83.3k wafers for that (so about 1/5 of your yearly production per fab)
each wafer on 14nm process would yield 180 CPUs -> you need 55.6k wafer for that (so about 1/11 of your yearly production)
so all else equal, you probably need to make more than twice as much money per CPU to justify the move from 14nm to N+1 process.
for your server chips, let's say between all domestic players, you want to produce 2 million a year, 1 million N+1 and 1 million N+2
each wafer on N+1 process would yield 48 CPUs -> you need 20.9k wafers for that (so about 1/20 of your yearly production per fab)
each wafer on N+2 process would yield 30 CPUs -> you need 33.3k wafers for that (so about 1/7 of your yearly production per fab)
so the revenue on the N+2 process CPUs need to be 3 times as much.
for your AI GPUs, we already calculated based on Nvidia numbers that you need upwards of 500k a year.
each wafer on N+1 process would yield 24 CPUs -> you need 20.9k wafers for that (so about 1/20 of your yearly production per fab)
each wafer on N+2 process would yield 15 CPUs -> you need 33.3k wafers for that (so about 1/7 of your yearly production per fab)
for your smartphone CPUs, let's say we try to capture all the lower to medium end phone market of domestic players (let's say 200 million)
each wafer on N+1 process would yield 240 CPUs -> you need 833k wafers for that (so about the full yearly production of 2 fabs!)
each wafer on N+2 process would yield 150 CPUs -> you need 1.3 million wafers for that (so about the full years production of 6 fabs!)
From what I can see, just have 1 fab running can cover all the non-smartphone needs for the next couple of years. There is no reason they cannot produce all the server, desktop and industrial chips that require N+1/2/14 nm processes with even 1 fab that is capable of doing 400k N+1 wafers a year. But if they want to get into the smartphone business, they can only cover a fraction of the domestic demand (even for lower/medium end) until they can raise the yield on their N+2 process to like 40% (maybe that's with EUVs or could be with improved process)
So it seems to me that the new fabs will probably continue to produce 28 nm or 14 nm until the N+2 yield can raise to an acceptable level. But everything else I talked about (next generation desktop, server CPUs and high end GPUs) can all be produced in 1 extremely overworked SMIC fab. I would say this fab probably needs to some gov't subsidies to generate same margin as the more mature process. Or maybe Huawei/Hisilicon will just pay way over the top to get them to produce all the chips they need.
It's not exactly new that the US/West announces some sanctions or export controls on China, China eventually comes out with something similar or even better. But most of the times, it takes some years to get to it.
This time around, immediately after the US announced the bans of Nvidia GPUs A100 and H100 on China, Chinese companies release products that outperform them. More than one product actually. It might be pure coincidence, but it is still felt like a slap on the face.
Who knows, this may increasingly become the norm.
Actually I've tried to find these results on the ML Commons website but haven't found anything. Someone able to provide a link?
And the US shook reaction is so satisfying....lol In this boxing fight halfway thru the match at the 5th Round the American punch is getting weaker as the Chinese are able withstand the fight and is now counter attacking, the more it prolong the final outcome is in Chinese favor.
Last edited:
They will respond with tighter sanctions.
Good, if they can manage 60% yield on N+2 process and 70% yield on N+1 process, then they might be able to take more of the smartphone market once they have 3 fabs going. I don't think these things will be loss leaders at those yields. The operating cost in China is really low given the low energy/labour/tax cost that these fabs will encounter. The only thing that cost a lot are the initial Capex on the machines.
I was doing some calculations today. According to this, Phytium have sold over 5 million chips so far for a total sales of 3.5 billion RMB, which works out to be about $100 a CPU.
Now, if they are selling server chips at 4 times the cost of desktop CPUs (and they sold 5 desktop chip per server chip), then each desktop might generate $70 and server chip might generate $300. Let's say Phytium pays SMIC $25 per desktop chip. If you can fit 300 die per wafer for desktop CPU and yield 210 good chips for N+1 process, then that would be $25 x 210 x 400k = $2.1 billion. That's about half of the revenue TSMC generates in its high tech 5nm fabs iirc. I cannot imagine SMIC losing money on that type of revenue generation. That's why I think Chinese gov't might want to find ways to subsidize SMIC or Hisilicon for the initial Capex. After that initial investment, the lower Chinese cost just take over.
If they can get into smartphone business and produce like even 350 chips per wafer on their N+2 process, but make like $35 a chip. Then that would be $35 x 350 x 200k = $2.45 billion. I don't know, seems like if SMIC can get yield on N+2 process up and charge a premium for the chips (since they face no competition here), then they'd be profitable.
btw, took a longer look at Phytium, they definitely seem to have the largest market share among domestic CPU players. Based on their news feed, there seems to be a major push to get domestic software companies to develop to Phytium/Kylin computers and to use their computers to the education system.
I also some major players that adopted their chips including some gov'ts (I found Heilongjiang, Henan and CAAC), large public cloud providers (Baidu, JDCloud, China Telecom, China Mobile) and some large SOEs also.
Apparently, both S5000 server chip and D3000 are still in development. Hopefully, they are the ones that will get mass produced with SMIC's N+1 process.
I was doing some calculations today. According to this, Phytium have sold over 5 million chips so far for a total sales of 3.5 billion RMB, which works out to be about $100 a CPU.
Now, if they are selling server chips at 4 times the cost of desktop CPUs (and they sold 5 desktop chip per server chip), then each desktop might generate $70 and server chip might generate $300. Let's say Phytium pays SMIC $25 per desktop chip. If you can fit 300 die per wafer for desktop CPU and yield 210 good chips for N+1 process, then that would be $25 x 210 x 400k = $2.1 billion. That's about half of the revenue TSMC generates in its high tech 5nm fabs iirc. I cannot imagine SMIC losing money on that type of revenue generation. That's why I think Chinese gov't might want to find ways to subsidize SMIC or Hisilicon for the initial Capex. After that initial investment, the lower Chinese cost just take over.
If they can get into smartphone business and produce like even 350 chips per wafer on their N+2 process, but make like $35 a chip. Then that would be $35 x 350 x 200k = $2.45 billion. I don't know, seems like if SMIC can get yield on N+2 process up and charge a premium for the chips (since they face no competition here), then they'd be profitable.
btw, took a longer look at Phytium, they definitely seem to have the largest market share among domestic CPU players. Based on their news feed, there seems to be a major push to get domestic software companies to develop to Phytium/Kylin computers and to use their computers to the education system.
I also some major players that adopted their chips including some gov'ts (I found Heilongjiang, Henan and CAAC), large public cloud providers (Baidu, JDCloud, China Telecom, China Mobile) and some large SOEs also.
Apparently, both S5000 server chip and D3000 are still in development. Hopefully, they are the ones that will get mass produced with SMIC's N+1 process.
What shock reaction? I think the US politicians are too slow to be shocked by a new development like this. It will probably be a good thing if they don't notice anything until next year. By then, Biren and Kunlun should hopefully be able order and stock enough chips from TSMC that they won't have to be concerned about getting put on some entity list for "national security" reasons.
- Status