Samsung Foundry nabs Nvidia

[benb]

Samsung shares rise after Nvidia's Huang flags tie-up on new AI chips​

By Hyunjoo Jin
March 16, 20265:20 PM CDTUpdated 12 hours ago

A Samsung Electronics logo and a computer motherboard appear in this illustration taken August 25, 2025. REUTERS/Dado Ruvic/Illustration Purchase Licensing Rights, opens new tab

March 17 (Reuters) - Shares of Samsung Electronics (005930.KS), opens new tab rose as much as 5% on Tuesday after Nvidia (NVDA.O), opens new tab CEO Jensen Huang said ‌the South Korean company was producing Nvidia's new artificial intelligence chips.
The news fuelled expectations that Samsung's foundry division, which makes logic chips for customers including Tesla (TSLA.O), opens new tab , Apple (AAPL.O), opens new tab and Samsung's phone division, may be able to turn around as early as next year after posting ⁠billions of dollars in annual losses in recent years, analysts said.

At Nvidia's GTC developer conference in California on Monday, Huang unveiled Nvidia's new AI inference processor based on technology from chip startup Groq.

"I want to thank Samsung who manufactures the Groq LP30 chip for us and they're cranking as hard as they can," he said, adding the chips were in production, and would be shipped in the second half of this ‌year.
Samsung ⁠also showcased the Nvidia chips made using its 4-nanometer manufacturing process at the GTC.
Samsung shares were up 4.3% at 196,800 won as of 0252 GMT, after earlier reaching 198,000 won. The broader market (.KS11), opens new tab was up 2.7%.

Sohn In-joon, an ⁠analyst at Heungkuk Securities, expected Samsung's foundry business would be able to reach breakeven later next year. But he said weak demand from mobile phones stemming ⁠from surging memory chip prices could weigh on foundry earnings.
Advanced Micro Devices' (AMD.O), opens new tab CEO Lisa Su will meet Samsung Electronics Chairman Jay Y. Lee ⁠in South Korea on Wednesday, media reports said, with eyes on whether the two would discuss cooperation in memory chips and logic semiconductors.

Samsung shares rise after Nvidia's Huang flags tie-up on new AI chips

Shares of Samsung Electronics rose as much as 5% on Tuesday after Nvidia CEO Jensen Huang said ‌the South Korean company was producing Nvidia's new artificial intelligence chips.

www.reuters.com

 

[benb]

Strength in HBM4 memory chips corresponding to strength in Samsung Foundry.

 

[siliconbruh999]

Memory Chips are also made in Foundry so the title is correct but i still confused it to a logic foundry deall

 

[Xebec]

It looks like Samsung is making the Groq 3 LPU (Language Processing Unit) inference chip for Nvidia. Certainly SRAM heavy but also definitely a "logic" chip:

Nvidia’s Jensen Huang Thanks Samsung as It Manufactures ‘Groq 3’ AI Inference Chips

Nvidia Corp. is introducing its Groq 3 language processing unit (LPU), optimized for artificial-intelligence inference, along with the Groq 3 LPX rack system. The platform is designed to operate in conjunction with Nvidia’s next-generation AI system, Vera Rubin, with Samsung Electronics handling chi

www.thelec.net

The Groq 3 incorporates SRAM, which offers faster speeds than DRAM. Memory bandwidth ranges from 22 terabytes per second (TB/s) to as high as 150 TB/s—about seven to 45 times faster than high-bandwidth memory (HBM). This significantly reduces latency in AI token generation compared with HBM-based systems. Samsung Foundry is producing the chips. “I would like to thank Samsung Electronics for manufacturing the LPU chips,” Mr. Huang said. “Samsung is currently producing as many as possible.”

 

[blueone]

This Groq 3 initiative by Nvidia is a curiosity to me. So all this time Nvidia knew inference processing was coming up fast (they probably had more intelligence about where AI was going than any organization on the planet), they knew GPUs don't have a great architecture for inference, yet they have do to this weird acqui-hire to get a leadership inference design quickly? Nvidia admittedly does have a stellar track record with acquisitions (their entire networking strategy, now the highest revenue networking product group in the industry (supposedly), is based on two acquisitions), but the LPU is just another processor, albeit one with an avant-garde flexible HW processing flow chip architecture. I can see Intel needing SambaNova after PatG (and others) hollowed out Intel's R&D for anything that wasn't an x86 CPU, but Nvidia? I expected more from Nvidia internally.

 

[KevinK]

Nvidia admittedly does have a stellar track record with acquisitions (their entire networking strategy, now the highest revenue networking product group in the industry (supposedly), is based on two acquisitions), but the LPU is just another processor, albeit one with an avant-garde flexible HW processing flow chip architecture.

The issue is rapid evolution and specialization of processor architectures. Two things happened about 2-3 years ago that radically changed what AI chip companies need to build:
* Transformer-based MoE LLMs became huge - big enough to focus hardware optimizations specifically on their general structure.
* Disaggregation - DeepSeek showed the practical performance benefits of splitting prefill and decode phases of transformers. That resulted in a rapid rethink of AI processor clusters and software.

We're now seeing distinct hardware / processors developed/tuned for prefill and decode, linked together via fast networking and KV memory / caches.
* NVIDIA - Rubin CPX for prefill, Groq 3 LPU for decode, lots of special networking and storage for KV cache.
* Intel - NVIDIA Blackwell prefill, Gaudi 3 for decode

Intel's Hybrid AI Server Fuses Gaudi 3 with NVIDIA Blackwell for Enhanced AI Inference

Intel reveals a hybrid AI server strategy, fusing its cost-efficient Gaudi 3 chips with powerful NVIDIA Blackwell B200 GPUs

www.technetbooks.com

* Cerebras/Amazon - Trainium 3 prefill, Cerebras decode.

Cerebras Systems, Amazon strike deal to offer Cerebras AI chips on Amazon's cloud

Amazon.com and Cerebras Systems on Friday said they have reached a deal to combine the two companies' computing chips in a new service aimed at speeding up chatbots, coding tools and other artificial intelligence services.

www.reuters.com

But the real magic is the rack level software that makes it all happen, and does intelligent resource allocation and routing. I think we're going to see a couple more generations of rack level innovation with increasingly better tuned hardware, for the way current frontier models work.

 

[blueone]

When I read about the complexity of AI hardware, software, administration, and datacenter requirements (power, cooling, etc), it still looks to me like future of enterprise AI is in and with the cloud computing companies. Even if big corporations wanted to do their own AI, how can they compete for workers at any of these levels with the tech companies? I don't think they can. That Amazon deal with Cerebras gives me a lot more confidence in Cerebras's future.

 

[Markwrob]

As I was reading the description of the Groq 3 LPU, I couldn't help but think it sounded alot like a Cerebras compute unit. And then Kevin's post showed both doing decode in separate announced deals.

Mr Blue, I get the feeling Jensen chose this path because his other teams were already working balls-out on other projects.

 

[Fred Chen]

It looks like Samsung is making the Groq 3 LPU (Language Processing Unit) inference chip for Nvidia. Certainly SRAM heavy but also definitely a "logic" chip:

Nvidia’s Jensen Huang Thanks Samsung as It Manufactures ‘Groq 3’ AI Inference Chips

Nvidia Corp. is introducing its Groq 3 language processing unit (LPU), optimized for artificial-intelligence inference, along with the Groq 3 LPX rack system. The platform is designed to operate in conjunction with Nvidia’s next-generation AI system, Vera Rubin, with Samsung Electronics handling chi

www.thelec.net

The Groq 3 incorporates SRAM, which offers faster speeds than DRAM. Memory bandwidth ranges from 22 terabytes per second (TB/s) to as high as 150 TB/s—about seven to 45 times faster than high-bandwidth memory (HBM). This significantly reduces latency in AI token generation compared with HBM-based systems. Samsung Foundry is producing the chips. “I would like to thank Samsung Electronics for manufacturing the LPU chips,” Mr. Huang said. “Samsung is currently producing as many as possible.”

Must be a big chip, i. e., hard to yield.

 

[jorgequinonez]

I was reading online that they're using SF4X node, Samsung's fourth gen 4nm node.

 

[Daniel Nenni]

Competition breeds strength. Winning!

 

[blueone]

As I was reading the description of the Groq 3 LPU, I couldn't help but think it sounded alot like a Cerebras compute unit. And then Kevin's post showed both doing decode in separate announced deals.

Have you seen this paper?

https://cdn.sanity.io/files/chol0sk5/production/4e3e6966d98da9ef4bfd9834dbfc00921da58252.pdf

Mr Blue, I get the feeling Jensen chose this path because his other teams were already working balls-out on other projects.

I have no doubt Jensen keeps his staff busy, but we're talking about the richest company on the planet. Money is essentially no object. Hiring is whatever and whomever it takes. I'm skeptical. As unbelievable as it sounds, it feels like they missed a big transition while focusing on training. Nah, can't be.