Refresh reduction in dynamic memories

An effective approach to reduce the static energy consumption of large on-chip memories is to use a low-leakage technology such as embedded DRAM (eDRAM). Unfortunately, eDRAM, being a dynamic memory, requires periodic refresh, which ends up consuming substantial energy for large last-level caches. In upcoming architectures that stack a processor die and multiple DRAM dies, DRAM dies experience higher temperatures. Elevated temperatures increase the periodic refresh requirement of DRAM, also a dynamic memory, which increases its energy and hurts processor performance. In this thesis, we propose and evaluate techniques for refresh reduction in dynamic memories. We examine the opportunity of refreshing only the data that will be used in the near future and only if the data has not been recently accessed (and automatically refreshed). We present Refrint, a simple approach to perform fine-grained refresh of on-chip eDRAM multiprocessor cache hierarchies. We show that an eDRAM-based memory hierarchy with Refrint consumes only 30% of the energy of a conventional SRAM-based memory hierarchy, and induces a slowdown of only 6%. In contrast, an eDRAM-based memory hierarchy without Refrint consumes 56% of the energy of the conventional memory hierarchy, inducing a slowdown of 25%. While it is well known that different eDRAM cells exhibit very different charge-retention properties, current systems pessimistically assume worst-case retention times, and refresh all the cells at a conservatively-high rate. We use known facts about the factors that determine the retention properties of cells to build a new model of eDRAM retention times. The model is called Mosaic. We show that the retention times of cells in large eDRAM modules exhibit substantial spatial correlation. We propose a mechanism to exploit such correlation to save refresh energy. With simple techniques, we reduce the refresh energy of large eDRAM modules by 20x. The result is that refresh energy is all but eliminated. Finally, we focus on temperature and refresh reduction in 3D processor-memory stacks. We propose the Xylem Thermal Through Silicon Via (TTSV) placement schemes, to reduce the temperature of the DRAM dies and the processor die in the stack. The TTSV placement should respect the DRAM array structure and handle unknown multicore hotspots. The resulting temperature distribution still has a significant spatial variation. Therefore, we propose new DRAM refresh schemes that take advantage of this spatial variation in temperature. Our best Xylem TTSV placement scheme reduces the peak temperature of the DRAM stack by an average of 8.7 C. Combined with the best DRAM refresh scheme we reduce the number of refreshes by an average of 85%.

Name thesis:degree_name

Ph.D.

Level thesis:degree_level

Dissertation

Discipline thesis:degree_discipline

Electrical & Computer Engr

Grantor

University of Illinois at Urbana-Champaign

Year dc:date

2015

Author dc:creator

• Agrawal, Aditya

Contributors dc:contributor

• Torrellas, Josep
• Patel, Sanjay J.
• Shanbhag, Naresh
• Somasekhar, Dinesh

• dynamic random-access memory (DRAM)
• embedded dynamic random-access memory (eDRAM)
• Cache
• dynamic memory
• refresh
• three-dimensional (3D)
• stacking
• retention time
• Through Silicon Vias (TSV)
• Thermal Through Silicon Vias (TTSV)
• temperature
• leakage
• variation
• spatial

Statement dc:rights

• Copyright 2014 Aditya Agrawal

Language dc:language

en

Handle dc:identifier

http://hdl.handle.net/2142/72970

OAI identifier oai:identifier

oai:www.ideals.illinois.edu:2142/72970