Something to consider as we discuss nodes
@oldpuck @Thraktor @Z0m3le
Curiously, the AD103 silicon inside the 4080 16GB is speculated to be about 116-135 USD per chip. And with a yield of 75%. The chip is about 379mm^2. This doesnโt include the price of packaging thoughโฆ.
But in any case, letโs assume that they use a chip that has 8B transistors (this makes it easier), looking at the other nVidia products at the 8nm, 7nm and the 4N process, they have had densities of 43.5-45.1MTr/mm^2, 65.6MTr/mm^2, 121.1-125.5MTr/mm^2 and 98.3MTr/mm^2
I used the products they have on those nodes:
For the 8nm, I used the RTX 3000 series cards, some dies are slightly sparser than other dies. The Datacenter GPU, GA100, is on the 7nm node. The Hopper based GPU which is on the 4N process along with the other currently known Ada Lovelace cards of the RTX 4000 series.
Ok, so if it is on the 8N node, the chip would be about 177.38mm^2-183.9mm^2
If it is on the TSMC 7nm process, the chip would be 121.95mm^2
If itโs like Hopper, the chip would be 81.38mm^2
If it is like Ada Lovelace, the chip would be between 63.75mm^2-66.06mm^2.
We donโt know the exact price of the 10nm process from Samsung, but according to this:
Samsung doesnโt seem like itโs far off from the 7nm process of TSMC, which according to the image in that tweet is about $10k. I donโt see a reason to believe it deviates heavily even in the 10nm which as shown in that tweet was about $6K at TSMC.
For this sakes we will assume it is $6000 as well for a 12inch/300mm wafer. The others Iโll use the $10k 7nm and the $16k 5nm.
Ok, so letโs
assume some sizes below
(For all, scribe is 0.12mm)
8nm:
177.38mm^2 around close to 12.28mm x14.44mm. You can get 225-267 good working dies at a yield rate of 70.89-84% (clocks canโt save this to near 100%) at a defect of 0.1-0.2#/sq.cm
183.9mm^2 around close to 13.33mm x 13.8mm. You can get 213-254 good working dies at a yield rate of 70-83.39% (clocks canโt save this to near 100%) with a defect of 0.1-0.2.
7nm:
121.95mm^2 around close to 10.5mm x 11.61mm. You can get 372-419 good working dies with a yield of 78.8-88.6% with a defect of 0.1-0.2
4N:
81.38mm^2 around close to 8.8mm x 9.24mm. You can get 614-665 good working dies with a yield of 85.2-92.2% with a defect of 0.1-0.2.
66.06mm^2 around close to 7.9mm x 8.36mm. You can get 784-836 good working dues with a yield of 87.8-93.6% with a defect of 0.1-0.2.
63.75mm^2 around close to 7.8mm x 8.17mm. You can get 817-870 good working dies with a yield of 88.2-93.89% and a defect of 0.1-0.2
Ok. We got the yields, how about we apply this?
Letโs again say for the discussion that the goal is to deliver 3.5M chips for a product launch period, how many wafers would each need and what would the cost be for that? I presume this is where you are better suited at this than I am with regards to finance, but just the wafer and Iโm only trying here
. I should note this however, Iโm not sure of what the price would be, as they sell it to Nintendo at that point and do a markup that factors in the IP, the cost of the chip, packaging and substrate, etc. but I digress. Theyโll figure something out I think? Iโll only focus on the wafer and the SoC itself. Though larger chips tend to be sold for higher due to their lower yields from the wafers.
If itโs the slightly bigger 8nm variant, they would need 13,779-16,431 wafers are needed, so nVidia would pay Samsung $82.6M-$98.58M.
If itโs the slightly smaller 8nm variant, they would need 13,106-15,555 wafers are needed, so nVidia would pay Samsung $78.6M-$93.3M.
If itโs 7nm, 8,353 wafers are needed, so Nvidia would pay TSMC $83.5M
If itโs the Hopper 4N variant, 5,263-5,700 wafers are needed, so Nvidia would pay TSMC $84.2M- $91.2M
If it is the Ada less sparse 4N variant, 4,186-4,464 wafers are needed. So nvidia would pay TSMC $66.9M-$71.4M
If it is using the more sparse 4N variant, 4,024- 4,284 wafers are needed. So nVidia would pay TSMC $64.3M-$68.5M
And since manufacturing starts several months before the product comes out, letโs assume 4 months, they wouldnโt need say, 4k-5k in a month, theyโd need like 1.1-1.5k per month leading up to it and having enough ready for the months afterwards.
You tell me if you see anything wrong with this!
Edit: I suppose this is what Thraktor was referring to with it possibly being cheaper to be on the 5nm than the 8nm.