In-DRAM Bulk Bitwise Execution Engine

Abstract: Many applications heavily use bitwise operations on large bitvectors as part of their computation. In existing systems, performing such bulk bitwise operations requires the processor to transfer a large amount of data on the memory channel, thereby consuming high latency, memory bandwidth, and energy. In this paper, we describe Ambit, a recently-proposed mechanism to perform bulk bitwise operations completely inside main memory. Ambit exploits the internal organization and analog operation of DRAM-based memory to achieve low cost, high performance, and low energy. Ambit exposes a new bulk bitwise execution model to the host processor. Evaluations show that Ambit significantly improves the performance of several applications that use bulk bitwise operations, including databases.

• The paper introduces Ambit, a novel in-DRAM execution engine that uses triple-row activation and dual-contact cells to perform bulk bitwise AND, OR, and NOT operations directly within memory.
• Ambit demonstrates significant performance gains, achieving up to 44.9x higher throughput and substantial energy savings for bitwise tasks compared to CPU or GPU execution.
• This in-DRAM approach minimizes data movement, dramatically improving the efficiency of bitwise operations crucial for applications like databases, DNA alignment, and machine learning.

Overview of "In-DRAM Bulk Bitwise Execution Engine"

The paper "In-DRAM Bulk Bitwise Execution Engine" addresses the challenges associated with performing bulk bitwise operations in modern computing systems. Traditional approaches involve transferring large data across memory channels to processors, leading to high latency and energy consumption. This paper introduces "Ambit", an innovative mechanism for executing these operations entirely within DRAM, thus leveraging DRAM's inherent structure and operation to significantly reduce the cost, latency, and energy associated with these tasks.

The primary motivation for Ambit is the wide array of applications—such as databases, DNA sequence alignment, encryption, and machine learning—that can benefit from enhanced performance and efficiency in performing bitwise operations. By minimizing data movement through Processing using Memory (PuM), Ambit offers a substantial performance improvement for operations that are conventionally resource-intensive.

Key Concepts and Methodology

Ambit’s core concept is based on two mechanisms: Ambit-AND-OR and Ambit-NOT. Ambit-AND-OR exploits the design of DRAM by using triple-row activation (TRA) to perform bitwise AND and OR operations, leveraging the majority function of bitlines during charge sharing. Ambit-NOT, on the other hand, uses dual-contact cells (DCC) to perform bitwise NOT operations by utilizing the inverters in the DRAM sense amplifiers to flip the stored bits efficiently.

The implementation details of Ambit highlight several innovations:

• Triple-row Activation (TRA): This allows the simultaneous activation of three rows, exploiting the analog properties of DRAM to evaluate the majority function over the bitlines. This mechanism effectively computes the bitwise AND/OR operations in one step across a row-wide scale.
• Dual-Contact Cells (DCCs): These are modified DRAM cells used for bitwise NOT operations, wherein data can be accessed through either of two transistors connecting to distinct wordlines.
• RowClone Integration: The RowClone mechanism is employed to enhance the efficiency of row copying and initialization through two modes—Fast Parallel Mode (FPM) for same-subarray operations and Pipelined Serial Mode (PSM) for inter-bank operations.

Results and Implications

Ambit delivers remarkable improvements in throughput and energy efficiency for bitwise operations. The evaluation uses SPICE simulations to mimic process variations and showcase Ambit's reliability. Benchmark comparisons on platforms like Intel Skylake CPUs, NVIDIA GPUs, and 3D-stacked DRAM architectures reveal that Ambit offers 44.9x–9.7x higher throughput and significant energy savings.

Real-world applications such as database bitmap indexing and BitWeaving are cited, where Ambit provides 6x and up to 11.8x speedups respectively. Such findings exemplify the practical benefits of Ambit, making it a compelling choice for applications requiring intensive

bitwise operations.

Future Directions

The paper suggests promising future research directions, including expanding Ambit to other operations like counting and extending support beyond bitwise operations to arithmetic shifts and more complex logical operations. Additionally, exploring new applications that can be redesigned to fully exploit Ambit's capabilities, such as graph processing and machine learning algorithms, will further leverage Ambit's potential. Consideration of approximate computing with Ambit is also proposed to accommodate applications tolerant to minor errors.

In conclusion, Ambit presents a sophisticated, low-cost extension to DRAM technology that can significantly enhance performance for a wide range of applications by executing bulk bitwise operations entirely within memory. This not only improves computational efficiency but also paves the way for future innovations in DRAM-based processing architectures.