In the past two years, thanks to its own design advantages and the blessing of TSMC’s advanced process, AMD’s position in the processor has been rising, and it has gradually gained the confidence to challenge Intel. AMD is now reportedly turning more orders to TSMC for greater success. And this is good for both TSMC and AMD, and only GF may be injured.
First look at TSMC, based on past experience, they will propose a frontier node about every two years. 7nm in 2018; 5nm in 2020; 3nm in 2022, and so on. The cost per wafer drops rapidly as the process matures.
This is great for customers because not everyone can afford advanced processes, and not everyone needs the performance that the latest technology offers. Customers moving to new processes based on economics, for some, must be on the leading edge, while for others it only makes sense when wafer prices fall to a certain level. Adoption has grown over time, and what was once a lead has become the performance mainstream, then the mainstream, and then the old lag process.
TSMC’s investment in the process will not yield an attractive return on investment unless the process runs through the entire life cycle and capacity is nearly exhausted over the life of the process. This becomes a challenge when a large customer like Apple uses mostly leading features and may only use one node of the leading but not much use of older technologies.
For example, Apple will use the latest 5nm process in 2020 models and 7nm in older models, but probably not in any meaningful quantities in older nodes. When TSMC rolls out the 3nm node, many Apple wafer demand will be at 3nm/5nm, and 7nm numbers will plummet.
Once Apple exits the leading process, TSMC will have to find other customers who can use the process. However, given Apple’s sheer volume of sales, that capacity is difficult to fill. TSMC needs some high-volume applications to fill the capacity bubble. Plenty of PCs, CPUs, and GPUs suffice. For example, when Apple frees up its 7nm process, AMD can use that capacity for mid-range or low-end CPUs. When AMD frees up that capacity, Nvidia (can use it for mainstream or low-end GPUs. After Nvidia, IoT device makers can use that capacity. This is a constant cadence for foundries like TSMC).
One thing TSMC also helps is that Apple is likely paying dearly for the capacity bubble it left behind. Not everyone has Apple’s financial muscle. For AMD, if what AMD wants is leading capabilities, it will be challenging for TSMC to find customers on older processes. If AMD can make a profit on the processing volume of the CPUs it has freed up, then TSMC’s economics will improve, and TSMC may also offer AMD better pricing to help improve utilization of the TSMC process.
Note that AMD is not required to use state-of-the-art features for all of its devices. For example, the company’s cheaper CPUs for low-end laptops and Chromebooks don’t require advanced processes because cost is more important to these chips than performance. The same goes for low-end graphics chips. Taking advantage of the capacity bubble, which is made by moving high-end chips to newer nodes, making these low-end chips could be attractive to AMD.
Consolidating more and more designs on TSMC simplifies AMD’s operations and reduces engineering effort. AMD engineers have to use fewer processes, which can also help the company optimize designs.
Although GlobalFoundries has been a capable partner in the past, it has found leading capacity competition to be very expensive and has reduced investment in advanced processes. Over time, this dynamic will force AMD to move more and more designs to TSMC, as GlobalFoundries may no longer be a viable option for new designs – even at the low end.
Also note that even when it comes to high-end HEDT and server CPU solutions, AMD has only chosen to use the TSMC 7nm process for compute chips, not IO chips (see image below).
As you can see, AMD’s I/O chips are fabricated on a 14nm process for two reasons. The first, and more widely known, is that IO chips will not benefit as much from the 7nm process, and it is more cost-effective to keep them in the GlobalFoundries 14nm process (there is speculation that AMD also uses the TSMC 14nm process for I/O chips ); the second is a little known and unproven issue that AMD may need to license special IO/designs that were available when AMD designed Zen 2, only 14nm instead of 7nm.
Over time, the required third-party I/O designs may be fabricated using TSMC’s 7nm technology. If so, it might make sense for AMD to move IO chips to 7nm and use the increased transistor density to add new features.
Making computers and I/O dies at TSMC could improve AMD’s operational efficiency. As TSMC has been looking to grow its packaging business, the move described in this article could also set the stage for TSMC to win some or all of its packaging business from AMD. If so, this could also help reduce AMD’s manufacturing cycle for complex multi-chip devices.
Considering the above factors, moving incremental business from GlobalFoundries to TSMC is a logical progression for AMD’s business and will likely reduce AMD’s operating and product costs over time. Over time, AMD’s operating expenses will fall and profit margins will increase.
It’s also favorable for TSMC, which has a growing market share in foundry and increased revenue in its desirable frontier process business.
The Links: IRKV300-04 1DI50A-050