Rumor
Nvidia may integrate Groq LPU units into Feynman GPUs via 3D die stacking by 2028
Wednesday, December 24, 2025 at 09:04 PM
NVIDIA is rumored to be exploring the integration of Groq LPU technology into its 2028 Feynman GPU architecture, potentially using a 3D chiplet stacking method comparable to AMD's V-Cache technology.
Context
Nvidia is reportedly exploring the integration of Groq’s specialized Language Processing Unit (LPU) technology into its upcoming Feynman GPU architecture, slated for release in 2028. This strategic move would likely utilize advanced 3D die-stacking techniques, mirroring the multi-die approach pioneered by AMD with its X3D series. By layering LPU chiplets directly onto the GPU silicon, Nvidia aims to drastically reduce latency and accelerate inference speeds for large language models, addressing a critical performance bottleneck in current AI hardware.
The potential collaboration signals a major shift toward heterogeneous computing as Nvidia prepares to defend its market dominance against evolving competitors. With the Feynman generation following the 2026 launch of the Rubin platform, the inclusion of dedicated inference hardware could redefine performance benchmarks for global data centers. Investors should monitor this development as a sign of Nvidia’s pivot toward specialized, stacked architectures to maintain its technical lead through the end of the decade.
Related Companies
Nvidia
NVDA