DNN Compiler EngineerSenior Engineer)

  • Tokyo
  • Remote OK - Anywhere in Japan
  • Full-time
  • May 10, 2023
Conditions
location-icon
Apply from Japan Only
(You must live in Japan to apply)
Requirements
language-icon
Language Requirements
Japanese: Fluent
English: Conversational
career-icon
Minimum Experience
Senior or above

What you will do/仕事内容

We, LeapMind Inc., are working on developing a unique accelerator IP for deep learning (DNN).

This accelerator IP consists of various modules, each of which autonomously executes the sequence of instructions given by the host processor. There is limited computational resource and buffer within an accelerator IP, and its capability of computational resource depends on its configuration. 

Having a sophisticated compiler is critical for the IP to efficiently utilize these limited computational resources and buffer to maintain its computing unit being highly utilized while lowering the data transmission rate between the accelerator and DDR memory.

Designing and implementing such a compiler is a challenging task, which requires both high coding skills and deep knowledge of various algorithms.

You will be in charge of developing a compiler that takes deep learning networks as input and generates a sequence of instructions to those modules as an output.


LeapMindでは、深局孊習向けの独自アクセラレヌタIPを開発しおいたす。

この独自アクセラレヌタIPは様々なモゞュヌルから構成されおおり、これらのモゞュヌルは倖郚から䞎えられた呜什列に埓っおそれぞれ自埋的に挔算実行を行いたす。アクセラレヌタIP内の蚈算資源やバッファは限られおおり、たたこれらの蚈算資源量はコンフィギュレヌションによっお増枛したす。これら蚈算資源やバッファを効率的に運甚し、DDRメモリずのデヌタ転送量を抑え぀぀、挔算噚の高い皌働率を維持するには、高床なコンパむラが求められたす。そのようなコンパむラの蚭蚈および実装は、高いコヌディング胜力だけでなく様々な蚈算機アルゎリズムの深い知識が求められる、非垞にチャレンゞングな仕事です。本ポゞションでは、Deep Learningのネットワヌクを入力ずし、これらモゞュヌルぞの呜什列を出力ずするコンパむラの開発を担圓しおいただきたす。今回募集の開発チヌムは、PG BATTLE 2021 䌁業の郚で3䜍の結果を誇るチヌムで、非垞に優秀でコヌディング胜力の高いメンバヌず䞀緒に働くこずができたす。カゞュアル面談のご垌望やご応募をお埅ちしおおりたす


What you will need/必須スキル

  • BS degree in Computer Science or equivalent practical experience
  • Deep and broad knowledge of various algorithms
  • Deep and broad knowledge of various data structure
  • Proficient in programming to write complex algorithms and data structure
  • More than 3-year of development experience with C++11 and later
  • コンピュヌタサむ゚ンスの孊士号もしくは同等の実務経隓
  • 様々な蚈算機アルゎリズムの深い知識
  • 様々なデヌタ構造の深い知識
  • 耇雑なアルゎリズムやデヌタ構造を蚘述できる高いコヌディング胜力
  • 3幎以䞊のC++ (C++11以降)による開発経隓


Additional qualifications that are nice to have/歓迎スキル

  • MS degree in Computer Science or equivalent practical experience
  • Development experience of compilers, especially backend and optimization
  • Especially, knowledge of register-allocation, spill/fill, and scheduling
  • Knowledge and experiences of computer architecture and low-level programming
  • Understanding of operations used in machine learning and deep learning
  • Good records in programming contests such as Codeforces and AtCoder
  • コンピュヌタサむ゚ンスの修士号もしくは同等の実務経隓
  • コンパむラ、特にバック゚ンドや最適化の開発経隓
  • 特に register allocation, spill/fill, scheduling に関する知識
  • コンピュヌタアヌキテクチャや䜎レベルプログラミングに関する知識、経隓
  • 機械孊習・深局孊習で䜿甚される挔算の理解
  • Codeforces, AtCoderなどのプログラミングコンテストでの高い成瞟
  • 日垞䌚話レベルの英語によるコミュニケヌション胜力

LeapMind Inc. is developing its business with the company mission, “to create innovative devices with machine learning and make them available everywhere” and our original weight reduction technology for deep learning models, the dedicated circuit design, and leveraging the knowledge gained from the collaboration with more than 150 companies.

Our core product "Efficiera" is an ultra-low power AI inference accelerator that can be implemented on an FPGA device or ASIC/ASSP device, and is specialized for CNN inference operations, taking full advantage of our original deep learning model weight reduction method "extremely low bit quantization". It will enable advanced data processing by deep learning in environments where AI could not be used before.

Ultra low power AI inference accelerator IP EFFICIERA

"機械孊習を䜿った今たでにないデバむスをあたねく䞖に広める"を䌁業理念に、独自のディヌプラヌニングモデルの軜量化技術や専甚回路蚭蚈技術ず150を超える䌁業ずの共創で埗た知芋を掻かし、事業を展開しおいたす。

私たちが開発したコア補品である”Efficiera"は、独自のディヌプラヌニングモデル軜量化手法「極小量子化技術」を最倧限に掻かす、FPGAデバむス䞊もしくはASICデバむス䞊の回路ずしお動䜜するCNNの掚論挔算凊理に特化した超䜎消費電力AI掚論アクセラレヌタIPで、今たでAIが䜿えなかったような環境でもdeep learningによる高床な情報凊理を可胜にしたす。

超䜎消費電力AIアクセラレヌタIP EFFICIERA

View LeapMind's company page

Search jobs from LeapMind in Japan
DNN Compiler EngineerSenior Engineer) at LeapMind
APPLY NOW  âžœðŸ‡¯ðŸ‡µ Residents Only