Backend Glitches: Who Did Samsung Dump Google's 'Dirty Work' On?

If chip design is likened to running a high-end restaurant, front-end design is developing the recipes, while back-end design and packaging are the kitchen's chopping, plating, and serving. No matter how amazing the recipe is, if the kitchen staff is insufficient, orders pile up and can't be served. Recently, Samsung got stuck at this stage.

According to The Elec, Samsung is considering outsourcing the back-end design work for the I/O dies in Google's 2nm TPU chips. With a 5x growth in foundry demand, plus orders from its own Exynos and various clients, Samsung's internal engineering resources are maxed out. Back-end design, often seen as "grunt work," became the first burden to be offloaded.

What's outsourced isn't technology, but capacity gaps

Samsung is looking for teams capable of physical layout, timing analysis, and packaging verification. These jobs don't have the high technical barriers of front-end design but are extremely labor-intensive. The back-end process for a single 2nm chip, from placement and routing to final GDS file delivery, easily requires dozens of engineers working for half a year.

Samsung's calculation is clear. It wants to push all its top-tier PDK development and process tuning teams onto breaking through front-end and advanced nodes. The back-end I/O chips, although essential "peripheral vessels" for Google's TPUs, are not the core computing units. Rather than letting elite engineers consume resources on these standard processes, it's better to outsource to design service companies specializing in back-end and packaging work.

Behind this is a forced reshaping of ecological niches

Samsung used to be a full-stack self-developed foundry. Clients handed over designs, and Samsung wanted to handle everything from design to manufacturing. But this model couldn't hold up at frontier nodes like 2nm. The history of repeated yield issues at 3nm taught Samsung that its manpower is only enough to guarantee full-process service for 2-3 major clients. Clients like Google, who distribute TPU designs across multiple foundries, must learn to cooperate in a "puzzle-piece" manner.

Item	Samsung Self-Developed Packaging	Outsourced Back-End Design
Resource Input	High, occupies core team	Low, only interface management
Delivery Cycle	Controllable, but long scheduling	Flexible, leveraging external capacity
Yield Risk	Internal closed-loop management	Requires stronger supply chain coordination

Data says it all. R&D personnel in Samsung Electronics' Device Solutions (DS) division grew by about 12% in 2023, but during the same period, demand for design services with contract values over 50% grew by nearly 45%. Manpower growth far lags behind order growth.

Google's Silence and Samsung's Bet

Google has not commented on the matter. But one detail is that Google's TPU design team has poached many people from Samsung in recent years. This is almost an admission: I don't fully trust your back-end process.

If this eventually materializes, it means Samsung admits an awkward reality—its foundry status in advanced nodes has shifted from "all-around player" to "core process provider." The dirty and tiring peripheral work must be supplemented by the external ecosystem.

You can count on one hand the companies globally capable of doing back-end design for 2nm I/O chips. If Samsung really packages this out, it essentially exposes its most critical client interfaces to external service providers. This is a rare move in the semiconductor foundry industry. The last similar scale of back-end outsourcing happened with TSMC during the surge in 7nm capacity.

But TSMC outsourced mature nodes back then. Samsung is aiming for 2nm.

Samsung is betting on two things: First, that Google won't doubt the entire TPU's process consistency due to the outsourcing of I/O chips. Second, that the companies capable of taking on 2nm back-end work have confidentiality capabilities and technical understanding that keep up.

The former concerns client trust. The latter concerns industry capacity. Both are hanging by a thread.

Deeper signals are hidden in the numbers

Samsung's foundry business capital expenditure was cut by more than 8% in 2024, reflecting the contraction of the industry cycle. Meanwhile, custom demand for AI chips is soaring. Nvidia, Google, and Amazon are all building their own AI chips. Foundries face a contradiction: orders are increasing, but each order is becoming heavier and more personalized.

Outsourcing the back-end of I/O chips is essentially trading management complexity for engineering complexity. Samsung chose the former because it can no longer bear the latter.

The outcome of this event hasn't settled yet. But in my view, it will eventually become industry norm. As Moore's Law hits physical limits, "combinatorial design" and "fragmented foundry services" for chips will become survival rules. Samsung is just the first player forced onto the stage to show its cards.

Once this card is played, the win or loss is no longer in Samsung's hands, but in the delivery quality of the outsourcing team. I'll save the reasons for next time.

Original link: https://www.ithome.com/0/977/224.htm