전자통신동향분석 (Electronics and Telecommunications Trends)

한국전자통신연구원 (Electronics and Telecommunications Research Institute)
권호 표지
DOI QR코드

DOI QR Code

초거대 인공지능 프로세서 반도체 기술 개발 동향

Technical Trends in Hyperscale Artificial Intelligence Processors

  • 전원 (초거대AI반도체연구실) ;
  • 여준기 (초거대AI반도체연구실)
  • W. Jeon ;
  • C.G. Lyuh
  • 발행 : 2023.10.01
PDF 다운로드
⟨ 이전 논문 다음 논문 ⟩

초록

The emergence of generative hyperscale artificial intelligence (AI) has enabled new services, such as image-generating AI and conversational AI based on large language models. Such services likely lead to the influx of numerous users, who cannot be handled using conventional AI models. Furthermore, the exponential increase in training data, computations, and high user demand of AI models has led to intensive hardware resource consumption, highlighting the need to develop domain-specific semiconductors for hyperscale AI. In this technical report, we describe development trends in technologies for hyperscale AI processors pursued by domestic and foreign semiconductor companies, such as NVIDIA, Graphcore, Tesla, Google, Meta, SAPEON, FuriosaAI, and Rebellions.

키워드

과제정보

이 논문은 2023년도 정부(과학기술정보통신부)의 재원으로 정보통신기획평가원의 지원을 받아 수행된 연구임[No.2022-0-00018, 거대 인공지능 학습을 위한 K-인공두뇌 반도체 개발].

참고문헌

  1. Epoch, Parameter, Compute and Data Trends in Machine Learning, 2023. 8. 21., Retrieved from https://epochai.org/mlinputs/visualization
  2. Yahoo!finance, "ChatGPT on track to surpass 100 million users faster than TikTok or Instagram: UBS," 2023. 2. 3.
  3. A. Vaswani et al., "Attention is all you need," in Proc. Int. Conf. Neural Inform. Process. Syst., (Long Beach, CA, USA), Dec. 2017, pp. 6000-6010.
  4. https://chat.openai.com/
  5. A. Dosovitskiy et al., "An image is worth 16×16 words: Transformers for image recognition at scale," arXiv preprint, CoRR, 2020, arXiv: 2010.11929.
  6. OpenAI, "GPT-4 technical report," arXiv preprint, CoRR, 2023, arXiv: 2303.08774.
  7. S. Bubeck et al., "Sparks of artificial general intelligence: Early experiments with gpt-4," arXiv preprint, CoRR, 2023, arXiv: 2303.12712.
  8. Wired, "OpenAI's CEO Says the Age of Giant AI Models Is Already Over," 2023. 4. 17.
  9. https://www.sapeon.com/
  10. Korea IT News, "Sapeon 'Enhancing omputation with next-generation memory artificial brain 'CIM''," 2022. 11. 17.
  11. N. Jouppi et al., "Tpu v4: An optically reconfigurable supercomputer for machine learning with hardware support for embeddings," in Proc, ISCA 2023, (Orlando, FL, USA), June 2023, pp. 1-14.
  12. Graphcore, "Graphcore documents," accessed online at https://docs.graphcore.ai/en/latest/.
  13. E. Talpes et al., "The microarchitecture of dojo, tesla's exa-scale computer," IEEE Micro, vol. 43, no. 3, 2023, pp. 31-39.
  14. FuriosaAI, FuriosaAI WARBOY: High performance inference chip for the most advanced vision applications.
  15. FuriosaAI, FuriosaAI NPU & SDK 0.10.0 Documents, Accessed online at https://furiosa-ai.github.io/docs/latest/en/
  16. https://kakaoicloud.co.kr/service/detail/1-44
  17. NVIDIA, NVIDIA H100 Tensor Core GPU Architecture: EXCEPTIONAL PERFORMANCE, SCALABILITY, AND SECURITY FOR THE DATA CENTER v1.04, 2023.
  18. Head Topics, "Nvidia's AI Chips Are Pulling Ahead in the Cloud," 2023. 8. 18.
  19. Rebellions, "ATOM: 5nm Versatile Inference SoC," 2023.
  20. https://enterprise.kt.com/pd/P_PD_NE_00_316.do
  21. A. Firoozshahian et al., "MTIA: First generation silicon targeting Meta's recommendation systems," in Proc. ISCA 2023, (Orlando, FL, USA), June 2023, pp. 1-13.
  22. CNN, "The big bottleneck for AI: A shortage of powerful chips," 2023. 8. 6.