TL;DR
This paper introduces Half-DRAM, a novel DRAM architecture that activates only half of a row to significantly reduce power consumption without sacrificing bandwidth. By reorganizing the DRAM array and leveraging sub-array level parallelism, Half-DRAM achieves substantial performance gains and power savings with negligible overhead, making it a promising enhancement for future memory systems.
What Problem Does This Paper Address?
Modern computing systems, especially servers and data centers, are heavily burdened by DRAM power consumption, which can account for over 25% of total power. Reducing this power is crucial for system-level efficiency. While fine-grained DRAM activation techniques have been proposed to tackle activation/precharge power (a significant contributor), they often lead to substantial performance degradation or large area overhead. This is primarily because these methods reduce the effective data width, hindering the ability to sustain full bandwidth. Furthermore, power constraints in DRAM limit performance improvements by imposing timing restrictions like tRRD and tFAW, which cap activation frequency. This paper aims to bridge this gap by proposing a fine-grained activation approach that mitigates these issues.
Methods
The core innovation of Half-DRAM lies in its reorganized internal architecture, specifically the "1RD-2HFF" (one row decoder, two helper flip-flops) structure. Unlike traditional DRAM's "1RD-1HFF" where one row decoder controls an entire row and its associated helper flip-flops (HFFs), Half-DRAM splits a MAT (Memory Access Target) into "left" and "right" slabs. A single row decoder now drives both slabs, activating half a row in each. This allows for activating only half of a row's data while still utilizing the full set of HFFs and global bitlines, thus preserving the full data width and bandwidth.
Two variants are proposed:
- Half-DRAM-1Row: Activates one half-row at a time. This design focuses on power reduction by limiting active bitlines. It also relaxes the tFAW constraint, allowing for more activations within a given window.
- Half-DRAM-2Row: Extends Half-DRAM-1Row by enabling two half-rows (one from the "Odd" group and one from the "Even" group) to be active simultaneously. This leverages sub-array level parallelism (SALP) to further boost performance by allowing more concurrent accesses.
The paper also discusses how the column select logic (CSL) is mirrored to correctly select data from the appropriate half-row, and how the "1RD-2HFF" structure can be implemented with minimal area overhead by reusing existing components and metal layers.
Datasets
The paper uses the SPEC2006 benchmark suite and the STREAM benchmark for evaluation. These are standard benchmarks for CPU and memory-intensive workloads, respectively. The authors also mention using Micron's DDR3 DRAM power calculator and datasheet for power modeling. No new benchmarks were introduced.
Key Insights and Ideas
- The "Row Overfetching" Problem: A significant portion of DRAM power comes from activating entire rows even when only a small fraction of data is needed. This "row overfetching" is inefficient, especially in modern multi-core systems where request patterns can be unpredictable, leading to low row buffer hit rates.
- Fine-grained Activation Trade-offs: Previous attempts at fine-grained activation sacrificed bandwidth by reducing the active data width, leading to performance degradation. Half-DRAM's key insight is to achieve fine-grained activation without reducing the effective data width.
- Relaxing Power Constraints = Performance Gains: By reducing activation power, Half-DRAM relaxes the strict power constraints (like tFAW) that limit DRAM activation frequency. This relaxation directly translates to performance improvements, especially for memory-intensive workloads.
- Sub-array Level Parallelism (SALP) Synergy: Half-DRAM's architecture is well-suited to exploit SALP, allowing for even greater performance gains by enabling concurrent accesses to different halves of a row.
- Minimal Overhead: The proposed "1RD-2HFF" structure and mirrored CSL logic are designed to be implemented with negligible area and power overhead, making it practical for adoption.
Key Results
- Performance:
- Half-DRAM-1Row showed an average speedup of 1.3% over the baseline with Relaxed-ClosePage policy, and even higher improvements (up to 11.8%) when using the Restrict-ClosePage policy (by relaxing tFAW).
- Half-DRAM-2Row achieved an average performance improvement of 10.7% with Relaxed-ClosePage, with some benchmarks reaching up to 19% speedup.
- In Wide I/O memory (which has stricter activation constraints), Half-DRAM-2Row demonstrated significant gains, with an average speedup of 38.1% and some cases exceeding 50%.
- Power:
- Half-DRAM-1Row achieved an average power efficiency improvement of 8.4% over the baseline. Specific benchmarks showed reductions of up to 10%.
- Overhead:
- The design incurs negligible area and power overhead, estimated to be less than 0.003%.
Limitations and Future Work
- Row Buffer Hit Rate: The paper acknowledges that Half-DRAM's smaller effective row buffer size (half-row) might induce more activations due to a potentially lower row buffer hit rate. However, they demonstrate through the "CHR ratio" (Crossing Half Row) analysis that most benchmarks exhibit good data locality, minimizing this issue.
- Relaxed-ClosePage Policy: The performance benefits of Half-DRAM-2Row are most pronounced with the Relaxed-ClosePage policy, which allows for two half-row activations if requests target both halves. The performance gains with Restrict-ClosePage are primarily due to tFAW relaxation.
- Future Work: The paper suggests that Half-DRAM can be a promising enhancement for future power-efficient computing. Further exploration of its integration with emerging memory technologies like 3D-stacked DRAM (e.g., Wide I/O) is highlighted as a key area for future research, as demonstrated by the significant gains in the Wide I/O case study. The authors also imply that further optimization of the row/column logic and management policies could unlock even greater potential.
Summary updated