Colleagues from Tom’s Hardware have decided to answer the question of how much NVIDIA production of chips for the new flagship video card of GeForce RTX 5090 at the TSMC plant. The GB202 graphic processor is huge (761.56 mm²) and high cost. Also, the printed payment with dense installation also has a high cost of production. But specific figures are not known to the consumer. By operating some free access data, we can conclude that the production of one GPU costs about $ 300.
Photography Techpowerup
The standard 300mm plate can accommodate about 72 potential GB202 chips with 31.5 × 24.2 mm. It is known that TSMC can charge up to $ 16,000 per plate, which is produced by 4 or 5 nm technology. And not all crystals from this plate will be working. The production of GB202 uses the TSMC 4NP technological node with its specific NVIDIA settings. According to the TSMC itself, the defect defect N5/N6 was about 0.05-0.06 defects per square centimeter six quarters after the start of mass production.
To date (4.5 years after the mass launch N5), it may be even lower due to the continuous improvement of production technology. But let’s say it is 0.05. Then, according to the calculator of the release of suitable crystals SemiAnalysisone plate will give 47 ideal graphic chips with 24756 CUDA nuclei and 21 more crystal with small defects. And these 21 crystals can also be used by NVIDIA, because the GeForce RTX 5090 has a chip configuration with 21760 active CUDA nuclei. That is, for the video card can be selected any number of chips from partially defective crystals.
If you consider only 47 ideal crystals, the cost of one chip at $ 340. Assuming that NVIDIA will be able to attract up to 55 crystals from one plate, it is already $ 290. Also, the video card cost includes a complex printed board, new GDDR7 memory chips and a cooling system. But at the recommended price of 1999, the NVIDIA does not remain in losing and receives good profit from each GeForce RTX 5090.
It is necessary to make a serious caveat that such calculations do not take into account the cost of development. And to remain a market leader, NVIDIA clearly pours serious money into this direction. On the other hand, we can still see professional solutions based on the GB202 chip, where the price will be even higher.
Source:
Tom’s Hardware
