¿ÀÇ¿§ÁöÅ×Å©³î·ÎÁö(ÁÖ)

NPU Compiler °³¹ßÀڠä¿ë

¸ðÁýºÎ¹® ¹× ÀÚ°Ý¿ä°Ç

¸ðÁýºÎ¹® ´ã´ç¾÷¹« ÀÚ°Ý¿ä°Ç Àοø

NPU Compiler 

°³¹ßÀÚ














[´ã´ç¾÷¹«]

- NPU Compiler °³¹ß (backend optimization)
- Computation graph optimization
- Computation resource allocation & Scheduling
- DRAM, Scratch pad memory management supporting deep learning training & inference
- Quantization, calibration, quantization-aware training µî ÃÖÀûÈ­ °³¹ß


[±Ù¹«ºÎ¼­ ¹× Á÷±Þ/Á÷Ã¥]

    ±Ù¹«ºÎ¼­: NPUÆÀ
    Á÷±Þ/Á÷Ã¥: ÆÀ¿ø

[ÀÚ°Ý¿ä°Ç]

- °æ·Â: °ü·Ã ºÐ¾ß °æ·Â 2³â ÀÌ»ó
- ÇзÂ: ´ëÇб³(4³â)Á¹¾÷ ÀÌ»ó

[Çʿ俪·®]

- Compiler °³¹ß °æÇè

- Computer Architecture ¹× ÃÖÀûÈ­ ¾Ë°í¸®Áò ÀÌÇØ

- Neural network training, inference °æÇè ¹× µ¿ÀÛ ¹æ½Ä ÀÌÇØ


0 ¸í

ȸ»ç¼Ò°³

  • ¿ÀÇ¿§ÁöÅ×Å©³î·ÎÁö´Â ´õ ¸¹Àº »ç¶÷µéÀÌ ÀΰøÁö´ÉÀÇ ±â¼úÀÇ ÇýÅÃÀ» ´©¸± ¼ö ÀÖµµ·Ï 
  • ½Ã½ºÅ۹ݵµÃ¼ Ĩ ¼³°èÀÇ Çٽɱâ´ÉÀ» °³¹ßÇÏ¿© »ç¶÷°ú ±â¼úÀÌ ¸¸³ª´Â Á¢Á¡À» ³ÐÇô°¡´Â ¹ÝµµÃ¼ IP Ç÷§Æû Àü¹® ±â¾÷ ÀÔ´Ï´Ù.
  • ¿ÀÇ¿§ÁöÀÇ AI Platform IP´Â ÀÚÀ²ÁÖÇàÂ÷, 5G ¹«¼± Åë½Å, IoT µî ´Ù¾çÇÑ ºÐ¾ß¿¡¼­ ÀÀ¿ëµÇ¾î Çõ½ÅÀ» ÁöÇâÇÏ´Â °í°´µéÀÇ ¾ÆÀ̵ð¾î¸¦ Çö½Ç·Î
  • ±¸ÇöÇØ¿Ô½À´Ï´Ù. Çѱ¹À» ½ÃÀÛÀ¸·Î ÇöÀç ¹Ì±¹,  ij³ª´Ù ÀÚȸ»ç¿Í ÇÔ²² ½Ã½ºÅ۹ݵµÃ¼ »ýÅ°è ÃÖÀü¹æ¿¡¼­ ¹ÝµµÃ¼ Ĩ ¼³°èÀÇ Çٽɱâ´ÉÀ» °³¹ßÇÏ¸ç °í°´°ú ÇÔ²² ½º¸¶Æ®ÇÑ ¼¼»óÀ» ¿­¾î°¡´Âµ¥ ¾ÕÀå¼­°í ÀÖ½À´Ï´Ù.

  • ¡Ø AI Platform IP: ¿ÀÇ¿§ÁöÀÇ ¸Þ¸ð¸®½Ã½ºÅÛ IP ¿Í Àΰø½Å°æ¸Á ÇÁ·Î¼¼¼­ (NPU) °¡ °áÇÕµÈ Ç÷§Æû IP

Æ÷Áö¼Ç ¼Ò°³

  • NPU Compiler °³¹ßÀÚ´Â Neural Processing Unit À» À§ÇÑ ÄÄÆÄÀÏ·¯¸¦ °³¹ßÇÏ°í ÃÖÀûÈ­ÇÕ´Ï´Ù. 
  • Neural Network Computation ÀÇ ÀÌÇØ¿Í Neural network computation¿¡ ´ëÇÑ ÀÌÇظ¦ °®°í Hardware Architect¿Í ¿øÈ°ÇÑ Ä¿¹Â´ÏÄÉÀ̼ÇÀÌ °¡´ÉÇÑ °³¹ßÀÚ¸¦ ä¿ë ÁßÀÔ´Ï´Ù.

´ã´ç¾÷¹«

  • - NPU Compiler °³¹ß (backend optimization)
    - Computation graph optimization
    - Computation resource allocation & Scheduling
    - DRAM, Scratch pad memory management supporting deep learning training & inference
    - Quantization, calibration, quantization-aware training µî ÃÖÀûÈ­ °³¹ß

ÀÚ°Ý¿ä°Ç

  • - ÇзÂ: ´ëÇб³ (4³âÁ¦) Çлç ÀÌ»ó
  • - °æ·Â: °ü·ÃºÐ¾ß °æ·Â 2³â ÀÌ»ó

¿ì´ë»çÇ×

  • - Compiler °³¹ß °æÇè

    - Computer Architecture ¹× ÃÖÀûÈ­ ¾Ë°í¸®Áò ÀÌÇØ

    - Neural network training, inference °æÇè ¹× µ¿ÀÛ ¹æ½Ä ÀÌÇØ

±Ù¹«Á¶°Ç

  • °í¿ëÇüÅÂ: Á¤±ÔÁ÷(¼ö½À±â°£3°³¿ù)
  • ±Ù¹«Áö: ¼­¿ï½Ã °­³²±¸ ¿ª»ï·Î 114 ÇöÁ׺ôµù 13Ãþ (°­³²¿ª)
  • ±Ù¹«ºÎ¼­: NPUÆÀ
  • ±Þ¿©Á¶°Ç: ¿¬ºÀ ÇùÀÇ ÈÄ °áÁ¤

ÀüÇü´Ü°è 

  • ÀüÇü´Ü°è: ¼­·ùÀüÇü > 1Â÷ ±â¼ú ÀÎÅͺä > 2Â÷ ½ÉÈ­ ÀÎÅͺä > ÃÖÁ¾ÇÕ°Ý  

Á¢¼ö¹æ¹ý ¹× Á¦Ãâ¼­·ù

  • Á¢¼ö¹æ¹ý: ÀÎÅ©·çÆ® ä¿ë½Ã½ºÅÛ ¶Ç´Â ÀÚ»ç ȨÆäÀÌÁö(https://openedges.com)
  • Á¢¼ö¼­·ù: Áö¿ø Á÷¹« °ü·Ã º»ÀÎÀÇ ¾÷¹«°æ·ÂÀ» Àß ¼Ò°³ÇÏ´Â À̷¼­ 1ºÎ
  • Á¢¼ö¾ç½Ä: ÀÎÅ©·çÆ® À̷¼­
  • Á¢¼ö±âÇÑ: ä¿ë½Ã ¸¶°¨

±âŸ À¯ÀÇ»çÇ×

  • ÀÔ»çÁö¿ø¼­ ¹× Á¦Ãâ¼­·ù¿¡ ÇãÀ§»ç½ÇÀÌ ÀÖÀ» °æ¿ì ä¿ëÀÌ Ãë¼ÒµÉ ¼ö ÀÖ½À´Ï´Ù.