Skip to content

recipes_source/amx.rst 번역#1131

Open
hijyun wants to merge 3 commits into
PyTorchKorea:masterfrom
hijyun:translate/amx
Open

recipes_source/amx.rst 번역#1131
hijyun wants to merge 3 commits into
PyTorchKorea:masterfrom
hijyun:translate/amx

Conversation

@hijyun
Copy link
Copy Markdown
Contributor

@hijyun hijyun commented May 16, 2026

라이선스 동의

  • 기여하기 문서를 확인하였으며, 본 PR 내용에 BSD 3항 라이선스가 적용됨에 동의합니다.

관련 이슈 번호

PR 종류

이 PR에 해당되는 종류 앞의 [ ][x]로 변경해주세요.

  • 오탈자를 수정하거나 번역을 개선하는 기여
  • 번역되지 않은 튜토리얼을 번역하는 기여
  • 공식 튜토리얼 내용을 반영하는 기여
  • 위 종류에 포함되지 않는 기여

PR 설명

recipes_source/amx.rst (Leverage Intel® Advanced Matrix Extensions) 문서를 한국어로 번역했습니다.

  • Intel® AMX 소개, PyTorch에서의 AMX 활용 방법, AMX 활용 여부 확인 방법을 다루는 레시피 문서입니다.
  • 번역 가이드의 문체·용어 규칙을 따랐으며, 소스 코드와 oneDNN 상세 출력 로그는 번역하지 않고 원문을 유지했습니다.

문서 URL: https://tutorials.pytorch.kr/recipes/amx.html

@skt0725
Copy link
Copy Markdown

skt0725 commented May 20, 2026

전문성을 어느 정도 요구하는 글로 보이는데도 잘 번역해주셨습니다!
몇몇 부분들(피동 표현 및 어색한 부분)에 대한 제안 드리겠습니다.
수고하셨습니다!

Comment thread recipes_source/amx.rst Outdated

Guidelines of leveraging AMX with workloads
-------------------------------------------
이 연산은 생성된 실행 코드 경로에 따라 oneDNN에서 완전히 처리됩니다. 예를 들어, 지원되는 연산이 AMX를 지원하는 하드웨어 플랫폼에서 oneDNN 구현으로 실행될 때, oneDNN 내부에서 AMX 명령어가 자동으로 호출됩니다.
Copy link
Copy Markdown

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

첫 문장 'oneDNN이 전적으로 처리합니다' 는 어떨까요?
두 번째 문장 피동 표현이 반복되어 줄이면 좋을 것 같아요
지원하는 연산이라고 수정해도 큰 문제는 없어보입니다!

Copy link
Copy Markdown
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

의견 감사합니다!
말씀해주신 대로 첫 문장은 “oneDNN이 전적으로 처리합니다”로 수정했고, 두 번째 문장도 피동 표현이 반복되지 않도록 조금 더 자연스럽게 다듬었습니다. “지원하는 연산” 표현도 반영했습니다.

@jykimai
Copy link
Copy Markdown

jykimai commented May 20, 2026

번역하시느라 고생 많으셨습니다!!
전문 용어가 많은 문서임에도 원문의 기술적 의미를 정확히 살려 번역해주신 것 같습니다 :)

전반적으로 자연스럽게 잘 번역해주셨고, 두 부분에 대해 다듬어보면 어떨까 하여 제안드립니다!

라인: line 9 (Introduction 섹션 첫 단락)

  • 현재 번역:
    "Advanced Matrix Extensions(AMX)는 Intel® Advanced Matrix Extensions(Intel® AMX)라고도 하며, x86 확장 기능입니다."

  • 근거:
    원문 also known as ~, is an x86 extension의 어순을 그대로 옮기다 보니 '~라고도 하며, x86 확장 기능입니다'의 흐름이 끊겨 보입니다. 주어와 술어 사이의 부연 설명을 재배치하면 더 자연스러울 것 같습니다.

  • 제안:
    "Advanced Matrix Extensions(AMX)는 Intel® Advanced Matrix Extensions(Intel® AMX)라고도 부르는 x86 확장 기능입니다."

라인: line 126 (Conclusion 섹션 마지막 단락)

  • 현재 번역:
    "언제나처럼, 문제가 발생하거나 궁금한 점이 있다면 포럼 이나 GitHub 이슈 를 통해 문의할 수 있습니다."

  • 근거:
    언제나처럼'이라는 표현이 원문의 As always` 의미는 잘 담고 있지만, 한국어로 읽었을 때 문장 첫머리에 오니 살짝 어색하게 느껴질 수 있습니다. 조금 더 자연스럽게 다듬으면 좋을 것 같습니다.

  • 제안:
    "문제가 발생하거나 궁금한 점이 있다면 언제든 포럼 이나 GitHub 이슈 를 통해 문의해 주세요."

    또는 첫머리를 생략하여 "문제가 발생하거나 궁금한 점이 있다면, 포럼 이나 GitHub 이슈 를 통해 문의할 수 있습니다."로 자연스럽게 이어가는 방법도 좋아 보입니다!!

다시 한번 번역 작업에 수고 많으셨습니다!

@hijyun hijyun force-pushed the translate/amx branch 2 times, most recently from 43db81e to 350338d Compare May 24, 2026 13:57
@hijyun
Copy link
Copy Markdown
Contributor Author

hijyun commented May 24, 2026

@jykimai
리뷰 감사합니다! 말씀해주신 line 9와 line 126 모두 반영했습니다

Copy link
Copy Markdown
Member

@hyoyoung hyoyoung left a comment

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

긴 문서 번역하느라 수고하셨습니다.
사소한 몇가지 수정사항 제안하니 확인 부탁드립니다

Comment thread recipes_source/amx.rst Outdated
AMX is designed to work on matrices to accelerate deep-learning training and inference on the CPU and is ideal for workloads like natural-language processing, recommendation systems and image recognition.
Advanced Matrix Extensions(AMX)Intel® Advanced Matrix Extensions(Intel® AMX)라고도 부르는 x86 확장 기능입니다.
이 확장 기능은 두 가지 새로운 구성 요소를 도입합니다. 하나는 ‘tiles’라고 불리는 2차원 레지스터 파일이고, 다른 하나는 이러한 tiles에서 동작할 수 있는 Tile Matrix Multiplication(TMUL) 가속기입니다.
AMX는 행렬에서 동작하도록 설계되어 CPU에서 딥러닝 학습과 추론을 가속하며, 자연어 처리, 추천 시스템, 이미지 인식과 같은 워크로드에 이상적입니다.
Copy link
Copy Markdown
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

행렬 연산에 최적화되어 정도로 의역해보면 어떨까요

Copy link
Copy Markdown
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

좋은 제안 감사합니다! 반영하였습니다.

Comment thread recipes_source/amx.rst Outdated
AMX is designed to work on matrices to accelerate deep-learning training and inference on the CPU and is ideal for workloads like natural-language processing, recommendation systems and image recognition.
Advanced Matrix Extensions(AMX)Intel® Advanced Matrix Extensions(Intel® AMX)라고도 부르는 x86 확장 기능입니다.
이 확장 기능은 두 가지 새로운 구성 요소를 도입합니다. 하나는 ‘tiles’라고 불리는 2차원 레지스터 파일이고, 다른 하나는 이러한 tiles에서 동작할 수 있는 Tile Matrix Multiplication(TMUL) 가속기입니다.
AMX는 행렬에서 동작하도록 설계되어 CPU에서 딥러닝 학습과 추론을 가속하며, 자연어 처리, 추천 시스템, 이미지 인식과 같은 워크로드에 이상적입니다.
Copy link
Copy Markdown
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

워크로드는 작업으로 의역해도 좋을 것 같습니다

Copy link
Copy Markdown
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

수정하였습니다!

Comment thread recipes_source/amx.rst Outdated

Intel은 4세대 Intel® Xeon® Scalable 프로세서와 Intel® AMX를 통해 AI 기능을 발전시켜, 이전 세대 대비 3배에서 10배 더 높은 추론 및 학습 성능을 제공합니다. `Accelerate AI Workloads with Intel® AMX`_ 를 참고하세요.
Intel® Advanced Vector Extensions 512 Neural Network Instructions(Intel® AVX-512 VNNI)를 실행하는 3세대 Intel Xeon Scalable 프로세서와 비교했을 때,
Intel AMX를 실행하는 4세대 Intel Xeon Scalable 프로세서는 한 사이클당 256개의 INT8 연산이 아니라 2,048개의 INT8 연산을 수행할 수 있습니다. 또한 한 사이클당 64개의 FP32 연산과 비교해, 한 사이클당 1,024개의 BF16 연산도 수행할 수 있습니다. `Accelerate AI Workloads with Intel® AMX`_ 의 4페이지를 참고하세요. AMX에 대한 더 자세한 정보는 `Intel® AMX Overview`_ 를 참고하세요.
Copy link
Copy Markdown
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

ISA를 지원한다는 의미에서 AMX를 실행한다는 것도 좋지만, 약간 어색해보입니다
Intel AMX를 지원하는 4세대 Intel Xeon 정도가 어떨까요

Copy link
Copy Markdown
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

수정하였습니다!

Comment thread recipes_source/amx.rst Outdated

The operation is fully handled by oneDNN according to the execution code path generated. For example, when a supported operation gets executed into oneDNN implementation on a hardware platform with AMX support, AMX instructions will be invoked automatically inside oneDNN.
Since oneDNN is the default acceleration library for PyTorch CPU, no manual operations are required to enable the AMX support.
PyTorch는 백엔드인 oneDNN을 통해 BFloat16 기반의 연산 집약적 연산자와 INT8 기반의 양자화에 AMX를 활용하여,
Copy link
Copy Markdown
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

INT8 양자화가 더 나을거 같습니다

Copy link
Copy Markdown
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

수정하였습니다. 감사합니다.

Comment thread recipes_source/amx.rst Outdated
Comment thread recipes_source/amx.rst Outdated
- torch.compile:

- When the generated graph model runs into oneDNN implementations with the supported operators, AMX accelerations will be activated.
- 생성된 그래프 모델이 지원되는 연산자를 사용하여 oneDNN 구현으로 실행될 때, AMX 가속이 활성화됩니다.
Copy link
Copy Markdown
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

생성된 그래프 모델이 oneDNN의 지원 연산으로 실행되면 정도로 의역하면 어떨까요?

Copy link
Copy Markdown
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

수정하였습니다!

@hijyun hijyun requested a review from hyoyoung May 29, 2026 05:54
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

4 participants