Nvidia's GB202 graphics processing unit has a die size of 761.56 mm^2, which makes it one of the largest GPUs for client PCs ever produced. The graphics card model it powers — the GeForce RTX 5090 — also appears among the most expensive add-in boards ever. Perhaps this is because the GB202 chip costs a fortune to produce. We spitballed some figures for what it might cost Nvidia to punch out these massive dies for its flagship GPUs. However, outside of TSMC and Nvidia, the details of actual yields are closely guarded secrets, so take our calculations with a grain of salt. Let's analyze the possibilities.
A 300-mm wafer can fit roughly 72 GB202 candidates, assuming that one die measures roughly 31.5 mm × 24.2 mm. This is not a lot, considering the fact that TSMC may charge as much as $16,000 per 300-mm wafer produced using its 4nm-class or 5nm-class fabrication technologies. Considering defect density and yields, Nvidia may have to spend around $290 to make a GeForce RTX 5090 graphics processor, though it could also increase to $340 if only the perfect dies were sellable. These are very rough napkin-math estimates, though, so take them with a grain of salt. Other factors, such as parametric yields, should also be considered, so calculating the end results involves more than a bit of fuzzy math.
TSMC makes GB202 using the so-called 4NP process technology, which is a custom version of the company’s N4P production node with some customizations for Nvidia. TSMC’s N4 and N4P belong to the N5 family of manufacturing technologies and have been in production for years. Based on TSMC’s defect density performance graph, the defect density of N5/N6 was around 0.05–0.06 defects per square centimeter six quarters after mass production began. By now (4.5 years after N5 entered high-volume manufacturing), it might be even lower due to continuous process improvements, but let us assume that it is still at 0.05 these days.
If this is the case, then each 300-mm wafer carries 47 perfectly fine GB202 graphics processors with 24,576 CUDA cores, four partial dies, as well as 21 dies with some kind of defect, according to the SemiAnalysis Die Yield Calculator (Murphy’s Model). This does not mean that those 21 die go to waste. First of all, Nvidia’s large GPUs tend to include plenty of redundancies, so a minor defect may not even damage the functionality of the die. Second, Nvidia does not need a perfectly fine GB202 die for the GeForce RTX 5090: it needs a GPU with 21,760 functional CUDA cores. Third, even GB202 dies with severe defects can eventually be used for cheaper graphics cards, such as the GeForce RTX 5080. In some cases, even partial dies can be sold.
Since we do not know for sure how many of the dies with defects are broken, how many are sellable as RTX 5090, and how many are sellable as something lower-end, let us assume that each 300-mm wafer carrying GB202 processors carries 55 GPUs that can be sold as GeForce RTX 5090 products. In this case, each processor could cost around $290 without dicing, testing, packaging, and binning, assuming that TSMC charges Nvidia $16,000 per N4/N5 wafer. If only absolutely perfect die were sellable, that would jump to $340 per die.
Even if a fully packaged and binned GeForce RTX 5090 silicon costs Nvidia $350, the company will still be able to make money on its $1999 graphics board. However, not all of that is pure profit, as other additives, such as the VRAM and board assembly, add considerable cost as well. That's also not to mention developing drives, support, supply chains, and a myriad of other costs that go into the final product.
It should be remembered that Nvidia is one of the companies that knows how to build large processors with great yields by adding redundancies and selling partly defective dies. As a result, Nvidia can possibly sell everything it produces on a wafer, albeit at different prices.
What is no less important is that Nvidia may have as many as 47 fully functional GB202 dies per wafer, and these can be sold as the RTX 6000 ‘Blackwell Generation’ professional graphics card for CAD and DCC applications or as the L50s board for datacenter-grade inference AI workloads. These solutions tend to cost thousands of dollars, so Nvidia can make a lot of money not only on the $2,000 GeForce RTX 5090 but also on the $6,000 ProViz and AI products powered by the GB202.
