At the GTC 2026 event earlier today, Samsung unveiled its HBM4E chip for the first time. During the same event, Nvidia CEO Jensen Huang announced several new chips and platforms for AI processing. One of them is the Groq 3 chip, which will be manufactured by Samsung Foundry.
This marks a huge shift, as Samsung Foundry had lost Nvidia as a client for foundry. Samsung last produced Nvidia chips (RTX 3000 series GPUs) back in 2020. Later, Nvidia switched to TSMC as the sole manufacturer of its consumer and non-consumer GPUs and other AI chips. Now, Nvidia is back, and this is great news for Samsung Foundry. Nvidia is also using HBM3E and HBM4 chips and SOCAMM2 modules in its AI systems.
Groq 3 LPU will be fabricated on Samsung Foundry's 4nm process node
Groq 3 is designed by Groq, a US-based fabless semiconductor company that specializes in AI inference chips. The new language processing unit (LPU) chip will be manufactured by Samsung Foundry using its 4nm (SF4X) process and is expected to ship in the third quarter of this year. It features 500MB of SRAM and 98 billion transistors, delivering up to 150TB/s of bandwidth. This is significantly higher than the 22TB/s offered by an equivalent amount of HBM4 memory.
The chip delivers nearly seven times more bandwidth than Nvidia’s Rubin platform with HBM. It is particularly optimized for sequential, low-latency tasks, such as the decode phase of large language model (LLM) inference. It is also designed for agentic AI, an emerging area of AI focused on systems that can perform complex, multi-step tasks autonomously.
While Groq 3 is specialized for real-time chats with AI and agentic AI tasks, Nvidia's Rubin chips are general-purpose and more suited for data analytics and AI model training. Groq 3 can be used with Nvidia’s Vera Rubin Ultra system when it launches later this year.
During his keynote, Huang said, “I want to thank Samsung, who manufactures the Groq 3 LPU chip for us. They are working as hard as they can, and I really appreciate it.â€
In December 2025, Nvidia signed a $20 billion deal to license Groq’s AI inference technology. The agreement also included the transfer of several senior Groq executives to Nvidia.
Post a Comment