●Fact check on HBM articles
- The content of the article should be viewed as mixed, some of which are consistent with facts and some of which are not facts. In fact, Samsung Electronics is still conducting Nvidia’s HBM3e 12-layer quality test
- No one has received Nvidia’s 12th tier quality yet. SK Hynix has not passed the 12th tier quality yet. Personally, SK Hynix’s 12th tier supply is expected to start within the end of the year
- In September, SK Hynix ordered 30 units of 12-layer TC bonders to ASMPT (Hong Kong company), but in reality, I think it will take three months to receive the quality. Of course, SK Hynix is the fastest out of the three memory companies to supply 12-layer quality, and the other two companies expect the supply to be pushed back to 2025
- Most importantly, in addition to SK Hynix, there were foreign affairs reports that forecast oversupply from 2025 as competitors’ HBM supply began in earnest, and considering the specificity of HBM and the timing of supply will be pushed back, these concerns oversupply are not necessary to think about right now
- We tend to think of HBM as a single product even if it’s the same HBM3, but there are dozens of types for servers. Nvidia and AMD products are different. Samsung Electronics did not receive HBM3e qualification from Nvidia, but received a qualification from AMD
- For the Nvidia B100, the HBM3e 8th tier will be the main product for the time being, and the 12th tier should be the B200A. With mass production of the B100, there is no need to hurry to mass-produce the HBM3e 12th tier product. Since the time to mass-produce the B200A is 25 years and 2Q, there is more time to supply the 12th tier than the market thinks
- With Nvidia currently accounting for more than 94% of global GPUs and AMD accounting for about 4% of the rest, it is right to make a judgment based on Nvidia’s timeline, but for AMD, I think Samsung Electronics’ HBM3e 12th tier will be supplied from the end of the year. The logic that Samsung Electronics cannot do HBM3e in the first place is not correct
- There are dozens of hundreds of different types of HBMs because they are customized differently by customer and application. Nvidia has a harder condition than AMD, so SK Hynix, which responded well first, is advantageous. It is true that SK Hynix’s technology is high, but it is not an overwhelming difference, and GPU companies such as Nvidia and AMD are pushing for customer diversification and prioritizing the transition to a consumer-oriented market
- In September, Samsung Electronics started preparations to replace its previous DRAM production with a new DRAM, targeting a 25-year-old 4Q. Unlike SK Hynix and Micron, Samsung Electronics conducts the manufacturing process of HBM3e in process 1a. Since Samsung Electronics’ HBM yield was problematic in the entire DRAM process, it seems that it will change process 1a and move on to process 1b
- Of course, there was a big problem with the disk design of process 1a, and this is a situation that affects process 1b, but the design change has also recently begun, and we are attempting to change it to the target of process 1d in 2027. Since the disk design of process 1a is now affected by the wrong part, and it usually requires about 3 years of R&D, we are trying to solve the fault even though it is late to the target in 2027. On the other hand, in process 1b, technology migration to solve the problem of the entire process is strongly entering
- The yield of HBM is actually large not only in the post-process but also in the whole process, and the yield may face a different phase due to an unexpected issue. For example, there was a recent issue of seizure and search by DIT, and SK Hynix may not be able to use the laser annealing equipment used from process 1b. If the annealing equipment is not used, it is expected to affect the HBM yield by about 10%. Therefore, the yield changes and nothing absolute should be considered
- In the article, the fact that Samsung Electronics focuses on HBM4 for reversal means a change in major definition, and it is targeting the fourth quarter of 25 and should be considered that the change in HBM4 will be large on the road map
- HBM4 is expected to change its major, and hybrid bonding will be adopted in the post-process, so finding companies that benefit from it will be an important investment point in the semiconductor sector in 2025
- Hybrid bonding requires plasma activation for low-temperature bonding. Cleanliness for intermolecular bonding and the process of flattening the surface are important. Accordingly, the importance of laser full cut, CMP, cleaning, and heat treatment process technologies will increase, and on the other hand, the demand for AFM (atomic microscope) equipment for measuring fine defects or impurities on the wafer surface will increase
https://n.news.naver.com/mnews/article/003/0012840194
Samsung, Nvidia HBM3E Qualification Test…a de facto abortive procedure