Seminar: Graduate Seminar

ECE Women Community

GAPiM: Genome Analysis acceleration using Processing-in-Memory

Date: March,16,2023 Start Time: 13:00 - 14:00
Location: 861, Meyer Building
Add to:
Lecturer: Naomie Abecassis
Affiliations: The Andrew and Erna Viterbi Faculty of Electrical & Computer Engineering


In the last decade, the advancements in high throughput DNA sequencing enabled collection of an unprecedented and exponentially growing amounts of genomic data. At the same time, it put a tremendous strain on genomic analysis infrastructure, which is traditionally based on conventional von Neumann computers. To prevent losing valuable genomic data and making wrong therapeutical decisions, new computing approaches to enable fast and high-capacity genome analysis are highly required. Arguably the main computational bottleneck of genome analysis is variant calling, finding variations in a newly sequenced genome that can reveal pathologies and diseases. In this research, we focus on implementing the Pair Hidden Markov Model (Pair-HMM), the most computationally intensive part of the traditional variant caller Genome Analysis ToolKit (GATK) on Processing-in-Memory architecture. We use DPU, a novel Processing in DRAM Memory (PiM) architecture developed by UPMEM. PiM is a non von Neumann computer architecture paradigm whose main purpose is to alleviate the memory wall, which severely limits the performance and energy inefficiency of a conventional von Neumann architecture. Considering the amount of genomic data, being able to process the genomic pipeline directly in the memory can significantly reduce the total amount of energy and improve the performance. To efficiently use UPMEM’s DPU, we convert the Pair-HMM floating-point computations into fixed-point and transform the data to log domain to avoid multiplications. Our solution outperforms the original GATK algorithm running on an Intel Core i7-5820K by 10×, and we show that future PiM architectures can outperform state-of-the-art FPGA implementations by 3×, while having a negligible effect on precision and sensitivity due to the fixed-point computation.

M.Sc. student under the supervision of Prof. Ran Ginosar and Dr. Leonid Yavits (EnICS Labs, BIU).

 

All Seminars
Skip to content