주메뉴 바로가기 본문 바로가기

POPUP ZONE

알기쉬운 AI 반도체 팝업
[강연 안내]인공지능 반도체 최고경영자 과정(온라인) 개최
회원 개인정보 추가 작성 요청
KAIST AI-PIM PIM반도체연구센터

IP 검색

Fixed-point Transformer Self-attention Accelerate unit (BERT) (차세대지능형반도체사업단)

· Transformer encoder 모델인 BERT에 대해 high-accuracy inference가 가능한 self-attention module

Feature
· Precision: 16bit INT, fixed-point unit
· 높은 model accuracy 유지 (low accuracy drop of 0.1 or less for GLUE tasks)
· 높은 정확도 유지를 위한 internal mixed precision (upto 32bit)
· self-attention의 matrix multiplication 연산을 위한 2 output stationary systolic arrays (per single head core)
· output stationary systolic arrarys의 streaming input/output을 매 cycle마다 produce/comsume 할 수 있는, pipelined softmax module
· CORDIC algorithm-based exponential unit (using less BRAM than LUT-based)
· 기존의 CORDIC algorithm의 limited output range (-1~1) 문제를 해결하기 위한 extended CORDIC 적용
· Freq : 215MHz
Application
· SoC
Business Area
Accelerator
Category
Arithmetic & Mathematic > Arithmetic & Logic Unit
Tech Specs
  • IP Name :

    Fixed-point Transformer Self-attention Accelerate unit (BERT) (차세대지능형반도체사업단)

  • Provider :

    Hanyang Univeristy

  • FPGA Device :

    Xilinx ZCU111

  • Foundry :

    SAMSUNG

  • Technology :

    12nm,16nm

Deliverables
· SW integration guide
Validation Status
· RTL simulation-based validation, FPGA-based validation, SW modeling & model accuracy evaluation
Availability
Available
Functional Diagram
Benefits
· -
List