본문 바로가기
KERD
논문KCI 등재

DNN 모델 추론을 위한 Lookup Table 기반 PIM의 가능성

Potential of LUT-based PIM for DNNs

0

조회

0

다운로드

0

인용

0

참고문헌

키워드

주제분류컴퓨터학

초록

Processing-in-memory (PIM)은 데이터가 저장된 메모리 근처에서의 연산을 통해 데이터를 더 가까운 곳에서 연산할 수 있게 하는 가속기이다. 데이터 이동에 특화된 PIM의 특성상 최근 주목받고 있는 생성형 언어모델과 추천 시스템에서는 PIM이 효율적인 가속기로 이용될 수 있다. 본논문에서는 PIM을 사용하기 적합한 어플리케이션인 추천시스템과 생성형 언어모델을 PIM으로 가속화한 최신 연구 동향을 살핀다. 또한, 선행연구들을 기반으로 PIM의 연구 방향성에 대해 토의한다. 마지막으로, PIM 시스템에서의 LUT의 효용성을 실험으로 성능을 확인하고 검증한다.

Abstract

Processing-in-memory (PIM) is an accelerator that enables data to be processed closer to the stored memory. Due to the nature ofPIM specialized in data movement, PIM can be used as an efficient accelerator in transformer-based generative model and recommendationsystem, which have recently garnered attention. In this paper, we examine the latest research trends of PIM that enhance therecommendtation system and generative language model, which are suitable applications for PIM usage. Additionally, we discuss theresearch direction of PIM based on previous studies. Lastly, we verify the effectiveness of LUT in PIM systems through experiments.

참고문헌 (24)

  1. J. Park, "AttAcc! Unleashing the Power of PIM for Batched Transformer-based Generative Model Inference", Proceedings of the 29th ACM International Conference on Architectural Support for Programming Languages and Operating Systems (ASPLOS), 2, pp.103~119 (2024), ACM
  2. G. Heo, "NeuPIMs: NPU-PIM Heterogeneous Acceleration for Batched LLM Inferencing", Proceedings of the 29th ACM International Conference on Architectural Support for Programming Languages and Operating Systems (ASPLOS), 3 (2024), ACM
  3. L. Ke, "Near-Memory Processing in Action: Accelerating Personalized Recommendation With AxDIMM", IEEE Micro, 42(1), pp.116~127 (2022)[DOI]
  4. Z. Wang, "EL-Rec : Efficient Large-Scale Recommendation Model Training via Tensor-Train Embedding Table", SC22 : International Conference for High Performance Computing, Networking, Storage and Analysis, pp.1~14 (2022), IEEE
  5. Y. Kwon, "TensorDIMM : A Practical Near-Memory Processing Architecture for Embeddings and Tensor Operations in Deep Learning", Proceedings of the 52nd Annual IEEE, 2019, 740~753[DOI]
  6. D. Mudigere, "Software-hardware co-design for fast and scalable training of deep learning recommendation models", Proceedings of the 49th Annual International Symposium on Computer Architecture (ISCA), pp.993~1011 (2022)
  7. L. Ke, "RecNMP: Accelerating Personalized Recommendation with Near-Memory Processing", 2020 ACM, 2020, 790~803[DOI]
  8. J. Park, "TRiM : Enhancing Processor-Memory Interfaces with Scalable Tensor Reduction in Memory", MICRO-54:54th Annual IEEE, 2021, 268~281[DOI]
  9. G. -I. Yu, "Orca : A Distributed Serving System for TransformerBased Generative Models", OSDI: 16th USENIX Symposium on Operating Systems Design and Implementation, pp.521~538 (2022)
  10. W. Kwon, "Efficient memory management for large language model serving with pagedattention", Proceedings of the 29th Symposium on Operating Systems Principles (SOSP), pp.611~626 (2023)
  11. Y. Jeon, "BiQGEMM : Matrix Multiplication with Lookup Table for Binary-Coding-Based Quantized DNNs", SC20:International Conference for High Performance Computing, Networking, Storage and Analysis, pp.1~14 (2020), IEEE
  12. D. C. Ganji, "DeepGEMM: Accelerated Ultra LowPrecision Inference on CPU Architectures using Lookup Tables", 2023 IEEE, 2023, 4656~4664[DOI]
  13. D. C. Ganji, "DeepGEMM : Accelerated Ultra Low Precision Inference on CPU Architectures using Lookup Tables", 2023 IEEE, 2023, 4656~4664[DOI]
  14. "T-MAC: CPU Renaissance via Table Lookup for Low-Bit LLM Deployment on Edge" (2024)[URL]
  15. C. Li, "PIM-DL : Expanding the Applicability of Commodity DRAM-PIMs for Deep Learning via Algorithm System Co-Optimization", Proceedings of the 29th ACM International Conference on Architectural Support for Programming Languages and Operating Systems (ASPLOS), 2, pp.879~896 (2024), ACM
  16. C. Lim, "Design and analysis of a processing- indimm join algorithm: A case study with upmem dimms", Proceedings of the ACM on Management of Data, 1(2), pp.1~27 (2023), ACM
  17. "QuIP#: Even Better LLM Quantization with Hadamard Incoherence and Lattice Codebooks" (2024)[URL]
  18. S. Shen, "Q-BERT : Hessian Based Ultra Low Precision Quantization of BERT", AAAI, 34(05), pp.8815~8821 (2020)[DOI]
  19. S. Kim, "I-BERT : Integer-only BERT Quantization", International Conference on Machine Learning (PMLR), pp.5506~5518 (2021)
  20. "UPMEM SDK"[URL]
  21. S. Noh, "PID-Comm: A Fast and Flexible Collective Communication Framework for Commodity Processingin-DIMM Devices", 2024 ACM, 2024, ~

관련 논문AI추천

이 간행물의 다른 논문 (9건)