WaferTetris: Efficient Freeform Mapping on Wafer-Scale Accelerators
arXiv cs.CL · 14 Nov 2026, 04:41 AEST
WaferTetris introduces a freeform mapping system for wafer-scale accelerators, enhancing compute capacity utilization. This approach improves performance metrics significantly compared to traditional template-based mappings.
Key facts
- Wafer-scale accelerators offer massive compute capacity through hundreds of thousands of processing elements (PEs) interconnected across a single wafer, but realizing this potential requires efficient application mapping.
- Existing approaches rely on template-based mappings to maintain productivity, but these templates restrict the mapping design space.
- In this work, we present WaferTetris, a system that removes these geometric restrictions through freeform mapping.
- For LLM decoding, it achieves up to $4.42 imes$ the system throughput of a template-based automatic mapping baseline on the same fabric size.
Read aloud
WaferTetris introduces a freeform mapping system for wafer-scale accelerators, enhancing compute capacity utilization. This approach improves performance metrics significantly compared to traditional template-based mappings. Wafer-scale accelerators offer massive compute capacity through hundreds of thousands of processing elements PEs interconnected across a single wafer, but realizing this potential requires efficient application mapping. Existing approaches rely on template-based mappings to maintain productivity, but these templates restrict the mapping design space. In this work, we present WaferTetris, a system that removes these geometric restrictions through freeform mapping. For L L M decoding, it achieves up to four dollars and forty two cents imes$ the system throughput of a template-based automatic mapping baseline on the same fabric size.
Sources in this cluster
- WaferTetris: Efficient Freeform Mapping on Wafer-Scale Accelerators
arXiv cs.CL · 14 Nov 2026, 04:41 AEST