The latest generation GPUs for artificial intelligence (AI) that are being designed by NVIDIA, AMD or Huawei, among other companies, They are a technological prodigy. However, their performance is deeply conditioned by the performance of the memory chips with which they coexist. And the most advanced GPUs are so fast that they often they are forced to wait until the memory gives them the data they need to be able to continue performing calculations.
HBM4e memory chips (High Bandwidth Memory) seek to end this bottleneck in AI hardware once and for all. The three largest designers of this type of integrated circuits (Samsung, SK Hynix and Micron Technology) are working on their HBM4e solutions, and the two South Korean companies will presumably deliver the first samples to their customers during the second half of 2026. The American company Micron will arrive a little later: in 2027.
SK Hynix currently leads this market with a share close to 70%so that the remaining 30% is shared between Samsung and Micron Technology. However, the future of your HBM4e memories is not only in your hands. To sustain its current market share SK Hynix must produce its future HBM4e memories on a cutting-edge lithography node, so, according to DigiTimes Asiahas decided to bet on it safely: it is evaluating the possibility of TSMC being in charge of manufacturing the core of these memories in its 3nm node.
SK Hynix and TSMC alliance is a seismic movement in the AI industry
Traditionally, memory chip designers have also been responsible for manufacturing their own integrated circuits. However, SK Hynix has three good reasons for leaving the production of its HBM4e memories in the hands of TSMC. The first of them is that this Taiwanese company manufactures the GPUs designed by SK Hynix’s main clients, so if it is also responsible for producing the memory, the final assembly of these two components using COWOS advanced packaging (Chip-on-Wafer-on-Substrate) is much simpler.
HBM4e memory must be produced using extremely small and fast transistors
Additionally, HBM4e memory must be produced using extremely small and fast transistors, and SK Hynix management knows that its current integration technologies are not as advanced as TSMC’s more sophisticated lithography. Finally, HBM4e memories will not only be responsible for storing information; They will also be able to carry out basic operations with the data before delivering it to the GPU in order to optimize hardware performance for AI. In some sense these memories will be more like processors than ever, and TSMC has much more experience than SK Hynix in manufacturing these chips.
Be that as it may, this alliance poses a problem: TSMC’s N3 node is absolutely saturated. NVIDIA, Apple and other clients of this company monopolize it, so TSMC is having very serious difficulties to meet the demand. In fact, it has faced this problem since it started manufacturing 3nm chips. Their second generation of this integration technology, known as N3E, refined N3B lithography enough to make its per wafer yield significantly higher.
In fact, N3E eliminates some of the transfer process steps of extreme ultraviolet lithography and reduces transistor density in order to minimize manufacturing costs and improve per-wafer yield. The third generation of lithography TSMC’s 3nm is called N3P. It is characterized by increasing the density of transistors by 4% and their speed by 5% while their consumption is reduced between 5 and 10% at the same clock frequency. It’s not bad at all.
Additionally, N3P lithography is fully compatible with N3E lithography design rules, so NVIDIA, Apple, and TSMC’s other customers can move their designs from N3E to N3P with virtually nothing done. However, despite the improvements that TSMC has made to its 3nm manufacturing nodes, there are not enough wafers for everyone. It is currently unclear how this Taiwanese company is going to resolve the arrival of SK Hynix to this node. Presumably you will have to maximize yield per wafer and increase your production capacity, but doing so is by no means a piece of cake. This is undoubtedly the biggest challenge TSMC faces because it cannot leave its best customers hanging.
Image | Generated by Xataka with Gemini
More information | DigiTimes Asia | SemiAnalysis
GIPHY App Key not set. Please check settings