Report: NVIDIA's next-gen Groq LPU chip shifts to 3D stacked DRAM, abandons large SRAM approach
AI summary
South Korean media outlet ZDNET Korea reports that NVIDIA's next-generation Groq LPU AI inference accelerator chip is expected to adopt a 3D stacked DRAM approach based on TSV technology, abandoning the current generation's large-capacity SRAM cache design. SRAM suffers from issues such as small capacity and limited density gains from process upgrades, and has inherent limitations in large-model performance and chip iteration. Cerebras, which also follows the large SRAM route, previously announced that CS-6 will feature 3D vertically integrated DRAM.
Body
The English body is being prepared. You can open the original source.