Unifying Primary Cache, Scratch, and Register File Memories in a Throughput Processor
Authors:
- Mark GebhartMGMark Gebhart
- Stephen W. KecklerSKStephen W. Keckler
- Brucek KhailanyBKBrucek Khailany
- Ronny KrashinskyRKRonny Krashinsky
- William J. DallyWDWilliam J. Dally
- 1 December 2012
- conference paper
- Published by Institute of Electrical and Electronics Engineers (IEEE)
- No. 10724451,p. 96-106
Abstract
Modern throughput processors such as GPUs employ thousands of threads to drive high-bandwidth, long-latency memory systems. These threads require substantial on-chip storage for registers, cache, and scratchpad memory. Existing designs hard-partition this local storage, fixing the capacities of these structures at design time. We evaluate modern GPU workloads and find that they have widely varying capacity needs across these different functions. Therefore, we propose a unified local memory which can dynamically change the partitioning among registers, cache, and scratchpad on a per-application basis. The tuning that this flexibility enables improves both performance and energy consumption, and broadens the scope of applications that can be efficiently executed on GPUs. Compared to a hard-partitioned design, we show that unified local memory provides a performance benefit as high as 71% along with an energy reduction up to 33%.Keywords
This publication has 18 references indexed in Scilit:
- A compile-time managed multi-level register file hierarchy2011
- Register packing for cyclic reduction2011
- Understanding the Energy Consumption of Dynamic Random Access Memories2010
- Fast tridiagonal solvers on the GPU2010
- Optimal loop unrolling for GPGPU programs2010
- A closer look at GPUsCommunications of the ACM, 2008
- ALPACM Transactions on Architecture and Code Optimization, 2007
- Resolving register bank conflicts for a network processor2003
- Reconfigurable caches and their application to media processing2000
- Automatically Tuned Linear Algebra Software1998