NVIDIA's qualification tests seem to be giving HBM manufacturers a hard time, as yield rates are significantly low compared to traditional memory products.
The issue of yield rates is common, mainly associated with semiconductor wafers when measured as the number of semiconductor chips produced from a single silicon wafer. Maintaining yield rates at an optimal level has been a huge concern and task for firms like TSMC and Samsung Foundry, but it seems like this issue has also crawled into the HBM industry.
A report from DealSite, a famous Korean media outlet, has revealed that manufacturers such as Micron and SK Hynix are facing each other head-to-head in the race to pass qualification tests for NVIDIA's next-gen AI GPUs, and it seems like low yield rates are in their way.
In the domain of HBM, yield rates are mainly associated with the complexity of stacked architecture, which involves multiple memory layers and intricate through-silicon vias (TSVs) for inter-layer connections. This complexity increases the chances of defects during manufacturing, potentially lowering yield rates compared to simpler memory designs. As reiterated by DealSite, if one of the HBM chips turns out to be defective, the whole stack is discarded, which shows how complex the manufacturing process is.
The source says that the overall yield rate of the HBM memory is currently around 65%, and if firms start to improve this number, a production volume drop will be witnessed, which is why the actual race lies in finding the solution between the two problems. However, it seems like Micron and SK Hynix are ahead in the race, with Micron reportedly initiating HBM3E production for NVIDIA's H200 AI GPUs, since it has passed the certification stages set by Team Green.
Now, while yield
Read more on wccftech.com