TechInsights: Lies About Nano
Foundries allow customers to fabricate manufacturing lies
Many people occasionally lie about their age or weight, but if a company does, it could be considered false advertising or at least brand dilution. This was the case in the recent semiconductor market when two leading foundries let their customers claim they were using 4nm when in reality they were still using 5nm. This situation has damaged the image of both parties, especially the foundries. Behind this, it also means that the development of transistors is slow.
This problem originally started with Samsung. In a long-running competition with TSMC for the next node, Samsung announced a year after delivering 5-nanometer chips that it will deliver production 4-nanometer chips by the end of 2021. As shown in Figure 1, TSMC plans to take two years between the 5nm and 4nm nodes, delivering the latter in the second quarter of 2022. To avoid giving Samsung a chance to show off, TSMC decided to "snap up" its N4 node progress by two quarters, just in time to catch up with its rivals (see MPR, "Fewer people think 3nm chips are worth the wait"). The first (and so far only) chip to use N4 is MediaTek's Density 9000.
Figure 1: Revised casting process roadmap. TSMC claimed to have delivered a 4-nanometer node target for production in late 2021, only to mislabel it as 5-nanometer technology at the time.
We should have been skeptical that TSMC could suddenly squeeze a full six months out of its often rigorous and time-consuming production certification cycle. However, things turned sour when TechInsights analyzed the Density 9000 and found that the key process dimensions were identical to TSMC's earlier N5 offerings (see MPR, "Size 9000 is now available in Plus size"). The foundry's claims of 4nm products are false, just as MediaTek claims to have 4nm processors.
Meanwhile, in a "distant galaxy"...
Meanwhile, Samsung's smartphone division is gearing up to launch the Galaxy S22, noting that the company's own Exynos 2200 is built with 4nm technology, but Qualcomm's Snapdragon 8 Gen 1 (which appears in some S22 models) is 5nm technical manufacturing. To avoid customer cravings for one processor over the other, the two companies decided to launch the 8 Gen 1 as a 4nm part. Because Samsung manufactures both, it created a new process called 4LPX that, like the N4, serves only one product. After analyzing the Snapdragon chips, TechInsights found that the 4LPX is not physically different from the 5LPE.
Ironically, the only products that use real 4-nanometer technology perform worse than counterfeit 4-nanometer products. Samsung was too aggressive in putting its 4LPE process into production, resulting in the Exynos 2200's high defect rate and inefficiency -- especially in its new graphics engine (see MPR, "Exynos 2200 debuts with AMD GPUs"). Due to the limited supply of Exynos chips, the phone maker has installed Snapdragon processors in most Galaxy S22 models.
These production woes further prompted Qualcomm to give up and instead create a new version of its flagship processor for TSMC's updated N4P, called the Snapdragon 8+ Gen 1, which we believe is true 4nm technology. The new Snapdragon is the first public N4P product, but we expect Apple's A16 to use N4P as well. Both chips are already in production and will be available on mobile phones around September.
TSMC typically gives Apple a quarter or two of exclusive access to its latest manufacturing technology. It's surprising to see Qualcomm on the first customer list, especially given that its Snapdragon-based smartphones compete directly and fully with Apple's iPhones. TSMC is willing to break Apple's exclusivity to take a major design from Samsung. (Qualcomm uses TSMC for many other products, but past several flagship processors have been based on Samsung products.)
The truth is hard to face
Why are foundries suddenly misreporting their transistor progress? It's easier than acknowledging weak progress. The difference between the fake 4nm process and the "real" 4nm process is only 5% optical shrinkage (10% area reduction). Even this small advance shattered Samsung's production model, which took TSMC two years to complete. The two foundries are expected to achieve 3-nanometer volume production next year, and despite their struggles with the 4-nanometer process, they have failed to inspire much confidence.
Given the long interval between nodes, a foundry may increase speed, power, or yield (ultimately to the same effect) while waiting for the next node. For example, while 4LPX shares the same dimensions as 5LPE, the former may offer some additional advantages to justify its new designation. But without any increase in transistor density, the name should be 5LPX to show that the process is still in the 5nm node category. Most people use the shorthand for 5nm or 4nm instead of the full process name, so the numbers still matter.
While forging new nodes can save foundries from short-term woes, it's detrimental to them. Touting a fake 4nm milestone and letting customers hype it up dilutes the value of the foundry brand. If 4nm isn't 4nm, does the node name still make sense? Why incur the extra cost for 4nm if an older process can be renamed?
For some time, foundries have not maintained a strict correlation between node count and transistor density (see MPR, "Editorial: Nanomania"). But at least when they adopted a smaller number of nanometers in the name, it showed that it had made some density gains. Blatantly relabeling a 5-nanometer process as 4-nanometer kills the last bit of meaning in node numbering. Therefore, when a foundry claims that it has reached the next level, please do not believe it until the disassembly is confirmed. The same goes for chip suppliers that boast about implementing new manufacturing techniques. As Moore's Law stumbles to an end, all we're left with are lies, damned lies, and meaningless node names.
