Reports stated that due to Samsung Foundry's massive improvement in its 4nm process yield rate, it has already secured an AI chip foundry order from a major data center client.
The semiconductor sector is recovering slower than expected. Although the second half of 2023 could see the traditional peak season effect, the extent of the demand momentum and the overall visibility of the economic environment right now is still unclear.
Reports from Korean media like ZDNet and Korea Economic Daily showcased the recent contrast between TSMC and Samsung Foundry. TSMC, even during the economic downturn, still secured high-end GPU-related AI chip orders from Nvidia and AMD. The order volumes were so large that, in response, TSMC even had to expand production capacity rapidly.
Korean media once lamented that Samsung was unable to secure a share of these orders. However, reports from industry sources and security firms suggest that Samsung's 4nm yield rate has improved to 75%. This has enhanced Samsung's capability to compete for orders with TSMC. Additionally, sources revealed that Samsung has already secured an AI chip foundry order from a data center client.
Although Korean media didn't specifically report the source of the orders, foreign media directly mentioned AMD as a possible client for the chip order. Furthermore, it's also reported that both Google and AMD have signed agreements with Samsung to manufacture AI chips with its 4nm process.
Despite that, foreign media also pointed out that AMD may not transfer orders from TSMC to Samsung yet. Because its CPU still relies heavily on TSMC's advanced process technology, the possibility of an order transfer is quite low in the short term. Nevertheless, sources believe that because TSMC is currently prioritizing production expansion to fulfill Nvidia's orders, AMD can use an order transfer to Samsung as a way to affect Nvidia.
Whether Samsung Foundry can become a weapon for AMD to hinder Nvidia remains uncertain. Although AMD's high-end AI processor APU and Google's TPU require high-volume HBM, Samsung's HBM capability does not guarantee it will secure the contract to manufacture AI chips for companies like AMD or Google.
The capacity bottleneck right now for high-end AI GPUs and HPC chips isn't the single chip in the front-end process but rather the severe capacity shortage of back-end advanced packaging. The reason why AI HPC clients are competing over TSMC's CoWoS capacity is that its advanced packaging technology is top of the market in yield rate and customer satisfaction, thus making it the preferred choice for customers.
Samsung's 4nm yield rate improvements don't guarantee that it'll secure high-end AI GPU chip manufacturing. Instead, it depends on what kind of 2.5D or 3D packaging solutions Samsung can present, or if Samsung can partner with high-quality OSAT companies. Only after securing advanced packaging capacity will Samsung's 4nm process have a chance to excel. This is likely the deciding factor for AI HPC clients.
