| haxxiy said: I've seen that the estimated cost of a 300 mm 5 nm wafer from TSMC is about $14,000-$15,000 vs. ~ $10,000 for the 7 nm process wafer. I don't know how much more expensive Nvidia's 'special' node is but let's assume it's 10% more and they all have a similar rate of defects. So there's a huge price increase there, more than 50% per mm2. Using the die yield calculator: $10,000 / 71 viable GA102 per wafer = $140,84 per chip $15,950 / 75 viable AD102 per wafer = $212,66 per chip That's for the big ones. The smallest (AD104) yields 170 chips per wafer in this estimate and would cost $93,82 per chip, around 66% of the GA102 Ampere chips. If the profit margin were the same, the price of the RTX 4080 12 GB should've been $599-$699. On the other hand, the profit margin of the 4090 is lower than the 3090 as everyone noticed eyeballing it. RTG might have had a good idea before Nvidia for once with the chiplet design. Using standard 5 nm and assuming 250 mm2 per chiplet, the cost would be $138,75 for the Navi 31 die vs. $113,63 for Navi 21 (a 22% increase per mm2). But even with 100 mm2 chiplets the cost should increase another 30% by 2026 over AD102, so yeah. |
I do wonder if Nvidia are actually paying more of a premium than the estimated wafer cost from when all companies were outbidding each other to get the wafers ordered years in advance amid the chip shortages and that cost has been passed onto consumers and why Nvidia wanted to cancel some of their wafers and why TSMC wouldn't allow it.
