Google doubles down on TSMC for future Pixel Tensor chips

[Daniel Nenni]

Pixel 9 Pro laying on a gray chair

• - Google plans to switch from Samsung to TSMC's foundry for its Tensor chips, starting this year.
• - The partnership with TSMC will continue for the next 3–5 years, until at least the Pixel 14.
• - TSMC's superior process node will enable Tensor chips to run cooler and consume less power.

Tensor chips have long been both the strength and weakness of Pixel devices, allowing Google to provide unrivaled AI features. However, their performance and thermal efficiency leave much to be desired, lagging a generation or two behind others. This is largely due to Google using Samsung's inferior foundry to fabricate its Tensor chips. Thankfully, the Pixel 10 will seemingly fix this by switching to TSMC's more advanced foundry. A new report claims this won't be a one-off move, with Google sticking to TSMC for the foreseeable future.

A DigiTimes report claims Google will continue its partnership with TSMC for at least 3–5 years, until at least the Pixel 14 in 2029. Given TSMC's superior fabrication node and technology, this should bode well for the thermal efficiency of the Tensor chips, ensuring they don't overheat or run hot (via @Jukanlosreve). The move will also put Google in the same league as Apple and Qualcomm, which rely on TSMC to fabricate their mobile chips.

A cooler chip should also help with battery life, as evident from the Pixel 9 lineup. Fabricated on Samsung's cutting-edge 3nm node, the phone's Tensor G4 SoC runs significantly cooler and consumes less power than previous Tensors. The new Exynos 5400 modem also helps, as it is more power-efficient than Samsung's previous Exynos modem.

Rumors indicate Google is fabricating the Pixel 10's Tensor G5 chip on TSMC's 3nm node, the same foundry technology used by Apple for the iPhone 16's A18 SoC.

Superior process node won't automatically boost performance​

Google Pixel 9a laying on top of a Google Pixel 9 Pro XL

TSMC's cutting-edge process nodes are typically in high demand, with companies often signing a multi-year contract and booking capacity years in advance. Plus, chipmakers design mobile SoCs based on the process node they plan to use for fabrication, so it makes strategic sense for Google to lock in TSMC's foundries for upcoming Tensor chips over the next few years.

A newer process node improves power efficiency and enables denser chip designs. For Google, this means it can achieve the same performance with lower power consumption or reduce chip size for potential cost savings.

However, performance gains depend more on factors like higher clock speeds, updated CPU cores, and architectural optimizations — not just the manufacturing process. So, even with TSMC's foundries, there's no guarantee future Tensor chips will match the performance of Qualcomm's flagship Snapdragon processors.

Google doubles down on TSMC for future Pixel Tensor chips

No more thermal efficiency troubles

www.yahoo.com

 

[Daniel Nenni]

Not a rumor, it is true. Everyone is using TSMC N3 as there were no alternatives. Plus, if you are competing against a company that is closely partnered with TSMC you are at a distinctive disadvantage. Google is also using TSMC for datacenter chips so this makes complete sense, higher volume, cheaper wafers.

 

[siliconbruh999]

I hope they fix the tensor performance with this one. The tensor performance is severely lacking due to Samsung Foundry they need to nail this design like Xiaomi.

 

[Xebec]

Not a rumor, it is true. Everyone is using TSMC N3 as there were no alternatives. Plus, if you are competing against a company that it closely partnered with TSMC you are at a distinctive disadvantage. Google is also using TSMC for datacenter chips so this makes complete sense, higher volume, cheaper wafers.

Hats off to TSMC, but it's also interesting to watch all of these companies willingly enter into a contract that could create a monopoly on all future nodes.

These are good short term deals, but I'm not so sure this is ideal long term for the fabless companies (or consumers).

ex: I think Nvidia will be worse off if they have only one possible supplier for a future 2035 node, than if they had multiple choices.

 

[siliconbruh999]

Well everything 5 years from now depends on how good 14A turns out to be in implementation and reference designs and IP availability.

A14 is H2 28 and 14A is H2 27. This is a genuine lead if they don't mess it up.

TSMC ecosystem is strong and Google is using mostly off the shelf IP from what was leaked.

 

[Daniel Nenni]

Hats off to TSMC, but it's also interesting to watch all of these companies willingly enter into a contract that could create a monopoly on all future nodes.
These are good short term deals, but I'm not so sure this is ideal long term for the fabless companies (or consumers).
ex: I think Nvidia will be worse off if they have only one possible supplier for a future 2035 node, than if they had multiple choices.

Agreed, which is why it is critical that Intel Foundry succeeds. Even TSMC agrees, they do not want the M word coming into play, but there is only so much they can do. Customers must start using Intel 18A and 14A otherwise there will be no choice other than TSMC.

I understand the dominance of TSMC N3 since there was nothing else available. TSMC N2 I am a bit surprised that it is exceeding N3 in design starts. Intel 18A and Samsung 2nm were late for the first round of designs but round 2, 3, and 4 should be starting now and Intel Foundry is certainly ready. Samsung I do not know, their SAFE conference is next week. Let's see what they say.

 

[osnium]

Agreed, which is why it is critical that Intel Foundry succeeds. Even TSMC agrees, they do not want the M word coming into play, but there is only so much they can do. Customers must start using Intel 18A and 14A otherwise there will be no choice other than TSMC.

I understand the dominance of TSMC N3 since there was nothing else available. TSMC N2 I am a bit surprised that it is exceeding N3 in design starts. Intel 18A and Samsung 2nm were late for the first round of designs but round 2, 3, and 4 should be starting now and Intel Foundry is certainly ready. Samsung I do not know, their SAFE conference is next week. Let's see what they say.

Let’s assume that Intel and Samsung fail and there is a monopoly. It’s not like the US government can do anything about it so why doesn’t TSMC embrace monopoly status and work to accelerate it by undermining any other foundries via pricing and capacity?

 

[kevin01]

Hats off to TSMC, but it's also interesting to watch all of these companies willingly enter into a contract that could create a monopoly on all future nodes.

These are good short term deals, but I'm not so sure this is ideal long term for the fabless companies (or consumers).

ex: I think Nvidia will be worse off if they have only one possible supplier for a future 2035 node, than if they had multiple choices.

The Prisoner's Dilemma. Do you know how to break this? A mafia boss (e.g., the U.S.G.) would break it. At times, the government's hand is simply needed for the sake of this country's competitiveness and the long-term health of industry players.

The Prisoner’s Dilemma and Strict Dominance – Game Theory 101

gametheory101.com

 

[siliconbruh999]

Let’s assume that Intel and Samsung fail and there is a monopoly. It’s not like the US government can do anything about it so why doesn’t TSMC embrace monopoly status and work to accelerate it by undermining any other foundries via pricing and capacity?

They are not doing anything about it even they will just see company go bankrupt like WolfSpeed.