Samsung’s foundry customers reportedly flock to TSMC — three firms move to Taiwanese chipmaker in latest exodus

According to ZDNet, South Korean developers of AI processors that used to make chips at Samsung Foundry are turning to TSMC as they roll out their next-generation processors. The website says that the companies are diversifying their supply. However, these chip designers adopt TSMC’s production nodes instead of Samsung Foundry’s technologies while still producing their previous-generation chips.

The list of companies that preferred TSMC’s production nodes over Samsung Foundry’s manufacturing processes reportedly includes at least three companies: DeepX, FuriosaAI, and Mobilint. ZDNet claims that companies are choosing different foundries for different processors in a bid to optimize performance and reduce risks. Yet, while this may be true in some cases, it does not seem to be the case in others.

DeepX has been using Samsung Foundry’s services for several of its products, including the DX-M1 25 TOPS AI accelerator made on one of SF’s 5nm-class nodes (the chip is also used for its 100/200 TOPS DX-H1 cards) and the 5 TOPS DX-V1 processor made on a 28nm-class node. However, according to the report, the company plans to use TSMC’s 12nm-class fabrication process for its 15 TOPS DX-V3 chip.

FuriosaAI used Samsung’s 14nm-class process for its first-generation chip, Warboy, aimed at computer vision applications. However, its next-generation processor for LLMs, called Renegade (RNGD), offers 512 FP8 TLOPS, 512 INT8 TOPS, and 1024 INT4 TOPS at 150W and is made by TSMC using one of its 5nm-class manufacturing technologies. The solution uses TSMC’s CoWoS packaging with 48 GB of HBM3 memory. FuriosaAI reportedly intends to make its refined RenegadeS product at TSMC as well. Since FuriosaAI’s Warboy and Renegade are aimed at different applications, we may say that the company diversifies its suppliers and optimizes costs by choosing the foundry that offers the best combination of power, performance, and price.

Mobilint uses Samsung Foundry’s 14nm fabrication technology to make its Aries neural processing unit (NPU), which delivers 64 – 80 TOPS at 25W and is equipped with 16 GB or 32 GB of LPDDR4X memory. However, its ultra-low-power Regulus chip, which promises 10 TOPS performance at 3W, will be made by TSMC at its 12nm-class production node (presumably an ultra-low-power one). While we may say that Mobilint is adopting a dual-foundry approach, the company saw more promise with TSMC’s 12nm technology for its low-power chip.

In general, diversification strategies make sense for both big and small companies. The former gains leverage to negotiate better pricing, and the latter gains a failsafe if one of its chips fails to become successful. Still, given the limited resources of small fabless chip designers, diversification may be a too costly endeavor for them.

However, in all three cases, companies adopt TSMC for subsequent generations of products. While TSMC’s nodes may be more competitive for the three aforementioned products, the Taiwan-based foundry may also provide better yields, better support, and better pricing conditions than the Korean contract chipmaker.

Industry experts suggest that Samsung Electronics strengthen its support for smaller fabless companies to compete with TSMC. They argue that Samsung needs to develop its ecosystem of design assets and services to better cater to these clients and close the growing gap in market share.