On this last day of 2019, grab a cocoa and let’s do some Fermi approximations to figure out how much of TSMC’s 7nm (N7 node) production AMD is using, as of Q3 2019. See what-if.xkcd.com/84/ for an explanation, and xkcd.com/2205/ for an example.
In response to . Assumptions need to be reasonable, but don’t need to be exact. Let’s assume 7nm chips sold in Q4 were mostly fabbed in Q3. In Q4 AMD revenue is forecast to be $2.2 Billion USD. Let’s gather relevant info about the sources of AMDs revenue.
AMD’s 7nm production so far is limited to a number of GPUs, and the CCDs (core chiplets) of it’s Zen2 based CPU products, including newly released Epyc Rome and Ryzen 3600, 3700, and 39xx series processors. No APUs have any 7nm parts yet.
The Vega 20 based 7nm GPUs (MI50, MI60, Radeon VII, Radeon Pro Vega II) have been out in various versions since the end of 2018. It’s the same die for all these GPUs, 330 mm^2 of TSMC N7.
Let’s say each Vega 20 die represents on average $1,000 in revenue, sold to large clients that buy in bulk. Anyway, should be between $500 and $5,000 so that’s close enough for our purposes.
Navi based 5700 and 5500 series GPUs are also TSMC N7. 5700 die is 250mm^2, each die represents around $300 in revenue. 5500 die is 160mm^2, has several variants including mobile, and is mainly sold through OEMs. MSRP for 5500XT is $170. Assume $100 average across the stack.
This is tricky because the huge difference in $ per mm^2 of N7 between the 5500 ($0.6) and the Vega 20 dies ($3). Vega 20 is for HPC, and might or might not be selling really well, while the 5500 is cheap consumer that will have high volume, but is just starting to ramp up.
Fortunately AMD breaks out two main segments for earnings. Computing and graphics (I’ll call “Consumer”) was $1.28B, and enterprise, embedded and semi-custom (call “Enterprise”) was $525M. The high ASP Vega 20 dies are probably all in Enterprise, as are Epyc server CPUs.
Q4 will have $400 million more in revenue. Let’s say $250M for Consumer, $150M for Enterprise. $1.55B Consumer, $700M Enterprise are the numbers we will go with.
A few quarters ago, AMD stated that GPU and CPU revenue were about even in Enterprise. Since then, Epyc has continued to grow. But they are still selling a lot of 12nm Naples, and other older GPUs. Let’s go with $500M in 7nm products for Enterprise.
$150M for GPUs = 50M mm^2 of N7. $350M for CPUs = ? Need to figure out average price per mm^2 of N7 for Epyc Rome. A lot of the die area is a GlobalFoundries 12nm IO die, which we don’t need to consider.
There’s a handy chart at anandtech.com/show/14694/amd… You can count the 7nm CCDs on a Rome SKU by dividing the L3 by 32. From top to bottom, there isn’t a drastic “skew per SKU” in price per CCD.
Plus higher SKUs probably get a better discount to the hyperscalars. Lets use 4 CCD average, for half list price of a 7402, $900. $225 per CCD. Each 7nm CCD is roughly 75mm^2. That brings us to $0.75 revenue per mm^2 of N7.
$350M / 0.75 ~= 466M mm^2 of N7. Adding 50M from Enterprise GPU we get 500M of 7nm silicon in a quarter from the Enterprise segment. This is not particularly accurate so we can round to the closest 100M.
Ok, lets assume that 2/3 of Consumer revenue will be the new 7nm stuff, 1/3 GPU and 2/3 CPU, which are numbers I just made up. Based on earlier in the thread about 5700 and 5500 series GPUs, $350M represents about 450M mm^2 of N7.
Based on MSRPs of ryzen 3000 series chips, lets go with an average of $250 in revenue per CCD. We will ignore threadripper, which would have a higher $ per CCD. $650M in revenue ~= 2.5M CCDs ~= 195M mm^2 of N7.
Adding it all up. Enterprise == 500M mm^2, Consumer GPU == 450M mm^2, Consumer CPU == 200M mm^2. Total == 1.15B mm^2 of N7 estimated in Q4. At a yield ~90% (probably better than that IRL ) that represents 1.3B mm^2 of N7 capacity.
We should be getting around 58K mm^2 per 300mm wafer. 1.3B mm^2 / 58K = 22,400 wafers. In 3 months. This equals AMD using 7500 wafers per month of TSMC N7 in Q3 2019. Let’s give this a plausible range of 5K to 15K wafers per month, because of the vagueness of assumptions.
I wouldn’t expect too much lower than 5K WPM, but there are various reasons why it could be much higher than 15K WPM. But I doubt it. So what’s TSMC N7 capacity?
We know they now have over 1 Million wafers per month capacity, but most of that is not 7nm. Check this out tsmc.com/english/dedica…
I could do some more reasoning determining what TSMC 7nm capacity is, but let’s just eyeball that chart, and assume that the additions in 2018 and 2019 are basically all 7nm. It’s around 1.5M wafers, or over 100K Wafers Per Month.
So we can see that the juggernaut that is the newly ascendant AMD was using like less than 10% of TSMC N7 capacity. So stop it with the rumors that AMD is using up all of TSMCs capacity. And also, AMD doesn’t need/want it’s own fabs.
90%+ of the R&D and capital required to create 7nm was funded by anticipated sales to other companies like Apple and Huawei. AMD isn’t even paying top dollar because it’s moving onto new nodes after they are already a year old.
Why is this important to understand? Because a 3 month production cycle + 2 month capacity delay is a lot different than “we can never make more chips than we are making now”. Clients know that they can move to AMD and order what they need and they will get a delivery date.
Plus stock analysts can know that AMD can continue to grow and take market share. If AMD was maxing out TSMCs fabs, then that’s it for like 2 years, it’s all downhill from here in terms of revenue. But it isn’t.
Apple is moving to N5. AMD can use that capacity. They aren’t a big TSMC customer now, but they are about to start shipping 7nm laptop APUs, which is a big deal. Epyc Rome, and then Milan will continue to take share.
Next gen GPUs will probably have larger dies and sell more. But the big thing is the console chips. If they are completely on 7nm or 7nm+ (which I think is the case), then AMD could quadruple to 30-40K WPM. See
Hope you enjoyed this long thread. I enjoy doing this type of analysis, its fun to see what you can derive from public info, that (some) professional journalists don’t know. Plus it’s given me a lot of stock gains on $AMD. 🙂
A few people pointed out some discrepancies in the assumptions. That’s expected, and redoing the math doesn’t change the result too much. Some changes make WPM go up, others make WPM go down. But one change I need to address. I screwed up the math for the Epyc processors.
This has a big enough impact it’s worth showing the difference. $225 per CCD == $3/mm^2 instead of $0.75/mm^2. Oops. 🤦♂️ Also Vega 20 GPU revenue goes into Consumer revenue not Enterprise, I’ve been told. That doesn’t end up changing much.
Updated Enterprise N7 used moves from 500 mm^2 to 170 mm^2 due to the Epyc math correction. Total N7 usage goes from 1.3B to 1B mm^2. New estimate is from 7.5K WPM to *less than 6K Wafers Per Month*, reasonable range would be better at 3K to 12K instead of 5K to 15K.
Even if my estimates are really wrong, it’s just hard for AMD to use a significant amount of TSMC supply. Another approach is to estimate how much AMD paid TSMC, and how much 7nm wafers cost. I have done this in the past and got very similar results.
This math is obviously wrong 🤦♂️ It should be $3 per mm^2 of N7, not $0.75. See the adjusted results here:
It actually doesn't change the big picture results too much, showing the strength of this approach, at least for the level of accuracy I'm going for.
It actually doesn't change the big picture results too much, showing the strength of this approach, at least for the level of accuracy I'm going for.
