ÀÌ·± ÀÏÀ» ÇÕ´Ï´Ù.
´Ù¾çÇÑ neural network ¾Ë°í¸®ÁòµéÀ» custom deep learning accelerator¿¡ porting ÇÏ°í ÃÖÀûÈÇÕ´Ï´Ù.
Layerº°, accelerator function blockº° performance analysis¸¦ ¼öÇàÇÏ°í ´Ù¾çÇÑ ÃÖÀûÈ ¹æ¹ýÀ» Àû¿ëÇÏ¿© ¼º´É ¹× Àü·Â¼Ò¸ð¸¦ ÃÖÀûÈ ÇÕ´Ï´Ù.
Quantization, pruningÀ» Æ÷ÇÔÇÑ ´Ù¾çÇÑ ÃÖÀûÈ ¹æ¹ýÀ» È°¿ëÇϸç, training °ú inference ¸ðµÎ¿¡ ´ëÇØ ÃÖÀûÈ ÇÕ´Ï´Ù.
À½¼º, ¿µ»ó, ÀÚ¿¬¾î, Ãßõ µî ´Ù¾çÇÑ ÀÀ¿ë¿¡ Àû¿ëÇÏ°í, ¼º´É »Ó ¾Æ´Ï¶ó ¼ºñ½º°¡ ¿ä±¸Çϴ Ư¼º¿¡ ¸ÂÃç ÃÖÀûÈ ÇÕ´Ï´Ù.
ÀÀ¿ë°³¹ß ¹× ¼ºñ½º¸¦ ¼öÇàÇÏ´Â °í°´»ç¿Í ¿øÈ°ÇÑ communication ÀÌ ÇÊ¿äÇÕ´Ï´Ù.
[ÁÖ¿ä ¼öÇà¾÷¹« ¹× ¿ªÇÒ]
- Neural network algorithmÀÇ deep learning accelerator porting
- Deep learning accelerator±â¹Ý neural network ¼º´É ÃÖÀûÈ
- À½¼ºÀνÄ/ÇÕ¼º, ¿µ»óÀνÄ/ÇÕ¼º, ÀÚ¿¬¾îó¸® µî ´Ù¾çÇÑ ÀÀ¿ëÀÇ NN ¹× SW ÃÖÀûÈ
- Quantization, pruning, quantization-aware training µî ÃÖÀûÈ ¹æ¹ý È°¿ë