Chips & computingResearch findings
Sangam: CXL-Integrated PIM DRAM Accelerator for LLM Inference
Original published 2025.11.16 01:39 KSTAdded to Blackdeck 2026.11.19 03:07 KST
What has changed?
Logic and memory functions are decoupled and implemented in heterogeneous technology nodes via chiplets connected through an interposer
Key points
- LLM inference is increasingly memory-bound due to growing model sizes and context length
- Decoding phase is dominated by memory-bound GEMV/GEMM operations, well-suited for PIM approaches
- Sangam achieves significant speedup and energy savings over H100 GPU in LLM inference tasks
Connect it to your work
Overcomes area cost and fabrication limitations of integrating processing elements within DRAM chips
What remains unproven
This research note is based on an abstract; check the original for test conditions and application scope.
Originals used for this note
Links open the original publisher in a new tab.