Samsung instead of TSMC for cheaper AMD products?
Against expectations, Intel will start manufacturing processors at TSMC. Similar diversification may occur with AMD, who relied almost entirely on TSMC lately. But according to unofficial reports, the company will start manufacturing some of its chips at Korea’s Samsung, much like Nvidia recently did with some of its GPUs. This would reduce AMD’s dependence on a single supplier, but the move likely seeks to cut costs, too.
AMD is reportedly interested in manufacturing chips at Samsung and is even said to have already agreed to and signed on some kind of contract. According to information that has surfaced on Twitter and is quoted by WCCFtech, AMD will reportedly be using Samsung’s 4nm node (presumably the 4LPP process). The Korean conglomerate is likely to offer production at a lower cost. This may lead to savings on the unit price of chips, but AMD may also use this cooperation to strengthen its position in price negotiations with TSMC, and thus save money indirectly.
This collaboration should concern future projects, not 4nm products already on the market or coming soon, such as Ryzen 7040HS (Phoenix APU). While it would be handy for AMD to have a second supplier for these chips, their design is tied to TSMC’s technology and it won’t be possible to just start producing them on Samsung’s different 4nm node unless that was planned and accounted for for long in advance. Even in that case, the tape-out would have had to take place at virtually the same time as the TSMC version of the dies, almost a year ago.
AMD could produce either or both of CPUs or GPUs at Samsung. Especially cheaper dedicated GPUs (for example, an equivalent of the now upcoming 6nm Navi 33) could be a good fit for this, as there’s a lot of price pressure and the technology doesn’t need to be the absolute most advanced. If the products were to be released in 2024 or 2025, they might be less technologically obsolete than Samsung’s 8nm node used for the Ampere GPUs was when they were released in 2020.
In the case of processors, the 4nm node could be used for example for an equivalent of the cheap Mendocino APUs (currently made on TSMC’s 6nm node). Another consideration is probably the production of APUs for consoles, where the cost pressure is particularly strong, and since it is not mobile hardware, one can turn a blind eye to the fact that the power draw will increase a bit against a hypothetical alternative made on the best available process.
If recent claims are to be believed, Samsung has reportedly managed to improve the performance and yield of its 4nm manufacturing node by now, the problems of which were seen when Qualcomm switched from it to TSMC’s 4nm process for Snapdragon 8 Gen 3 and immediately got better clock speeds and lower power draw. So let’s hope that the rumors of improvement are true, although a slightly worse process isn’t necessarily a serious problem when it comes to technology for low-cost products.
It has to be said that Samsung’s technologies are not a completely new thing for AMD. A number of GPUs (Vega, Polaris generation) and processors (14nm and 12nm Ryzens) were manufactured on Samsung’s 14LPP and 12LPP processes, but that was its licensed form in GlobalFoundries fabs. This company, which emerged from AMD’s fabs (and later absorbed Chartered Semiconductor and IBM’s manufacturing plants), had problems developing its own 14nm process and licensed Samsung’s technology as a replacement for the 14nm generation.
However, AMD, as far as is known, has only ever manufactured their chips at GlobalFoundries and wasn’t using Samsung’s lines. In this respect, the cooperation now being discussed would be something new.
We mentioned that for AMD, this move may lower manufacturing costs, as Samsung is expected to offer better pricing. On the other hand, Samsung may benefit from gaining a fairly significant customer. This is because, like Intel, it needs as much production volume and as many external clients as possible to be able to compete with TSMC in the field of the most advanced manufacturing processes.
Sources: WCCFtech, OreXda (Twitter)
English translation and edit by Jozef Dudáš
⠀
