Nvidia's next-gen Groq lpu chip shifts to 3d stacked dram, abandoning large sram approach

On October 21, 2026, IT Home reported that ZDNet Korea revealed NVIDIA is planning to use a 3D stacked DRAM solution based on TSV technology for its next-generation Groq LPU AI inference accelerator chip. This marks a shift from the current design, which relies on large SRAM for fast decoding output. SRAM, however, has limitations in capacity and density gains with process upgrades, constraining performance and iteration capabilities for large models. Cerebras, another company using large SRAM technology, has also announced plans to integrate DRAM vertically at the wafer level with its upcoming CS-6 chip. This move highlights a broader industry trend toward 3D stacking to overcome the limitations of traditional memory architectures.