๐Ÿค–Octo ML - ์Šคํƒ€ํŠธ์—… ํ”ผ์น˜๋ฑ ๋ถ„์„

์Šคํƒ€ํŠธ์—… ํ”ผ์น˜๋ฑ | ํ”ผ์น˜๋ฑ ์˜ˆ์‹œ | ์‹œ๋ฆฌ์ฆˆ C | ML AI ์Šคํƒ€ํŠธ์—… ํ”ผ์น˜๋ฑ | Octo ML
๐Ÿค–Octo ML - ์Šคํƒ€ํŠธ์—… ํ”ผ์น˜๋ฑ ๋ถ„์„
  • ๋‹จ๊ณ„: Series C
  • ํŽ€๋”ฉ๊ทœ๋ชจ: $85,000,000.00
  • ๊ธฐ์—…์„ค๋ช…: OctoML์€ ์—”์ง€๋‹ˆ์–ด๋ง ํŒ€์ด ์–ด๋–ค ํ•˜๋“œ์›จ์–ด์—์„œ๋“  ๋จธ์‹  ๋Ÿฌ๋‹ ๋ชจ๋ธ์„ ๋ฐฐํฌํ•  ์ˆ˜ ์žˆ๋„๋ก ์ง€์›ํ•˜๋Š” ์•ก์…€๋Ÿฌ๋ ˆ์ด์…˜ ํ”Œ๋žซํผ์„ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค.
  • ์ด ์Šฌ๋ผ์ด๋“œ๋Š” OctoML ํšŒ์‚ฌ์— ๋Œ€ํ•œ ์ฃผ์š” ์ •๋ณด๋ฅผ ์š”์•ฝํ•˜์—ฌ ๋ณด์—ฌ์ฃผ๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ์—ฌ๊ธฐ์— ํฌํ•จ๋œ ๋‚ด์šฉ์€ ๋‹ค์Œ๊ณผ ๊ฐ™์Šต๋‹ˆ๋‹ค:
    • ํšŒ์‚ฌ ์„ค๋ฆฝ ์ •๋ณด: OctoML์€ 2019๋…„ ์ค‘๋ฐ˜์— ์›Œ์‹ฑํ„ด์ฃผ ์‹œ์• ํ‹€์—์„œ ์„ค๋ฆฝ๋˜์—ˆ์Šต๋‹ˆ๋‹ค.
    • ์ง์› ์ˆ˜: ํ˜„์žฌ 85๋ช… ์ด์ƒ์˜ ์ง์›์ด ๊ทผ๋ฌดํ•˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค.
    • ์ž๊ธˆ ์กฐ๋‹ฌ: ์ด 4700๋งŒ ๋‹ฌ๋Ÿฌ์˜ ์ž๊ธˆ์„ ์‹œ๋“œ ๋ฐ A/B ๋ผ์šด๋“œ๋ฅผ ํ†ตํ•ด Madrona, Amplify, Addition ๋“ฑ์˜ ํˆฌ์ž์ž๋กœ๋ถ€ํ„ฐ ์œ ์น˜ํ–ˆ์Šต๋‹ˆ๋‹ค.
    • ์ฃผ๋ ฅ ์ œํ’ˆ: OctoML์˜ ํ•ต์‹ฌ ์ œํ’ˆ์€ OctoML SaaS ํ”Œ๋žซํผ์ž…๋‹ˆ๋‹ค.
  • ๋˜ํ•œ, ์Šฌ๋ผ์ด๋“œ์—๋Š” ํšŒ์‚ฌ์˜ ๋‹ค์–‘ํ•œ ํŒ€์›๋“ค์˜ ์‚ฌ์ง„์ด ๋ณด์—ฌ์ ธ, OctoML์˜ ๋‹ค์–‘์„ฑ๊ณผ ํŒ€ ๊ตฌ์„ฑ์„ ๊ฐ•์กฐํ•˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ์ด๋Š” ํšŒ์‚ฌ์˜ ์ปค๋ฎค๋‹ˆํ‹ฐ์™€ ์ง์›๋“ค์˜ ํฌ๊ด„์ ์ธ ๊ตฌ์„ฑ์„ ๋ณด์—ฌ์ฃผ๋Š” ๋™์‹œ์—, ํšŒ์‚ฌ ๋ฌธํ™”์™€ ํŒ€์›Œํฌ์˜ ์ค‘์š”์„ฑ์„ ๊ฐ•์กฐํ•˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค.

์ด ์Šฌ๋ผ์ด๋“œ๋Š” Apache TVM์˜ ์˜คํ”ˆ ์†Œ์Šค ์ƒํƒœ๊ณ„ ์„ฑ์žฅ์— ๋Œ€ํ•ด ์„ค๋ช…ํ•˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ์—ฌ๊ธฐ์„œ ๋‹ค๋ฃจ๋Š” ์ฃผ์š” ๋‚ด์šฉ์€ ๋‹ค์Œ๊ณผ ๊ฐ™์Šต๋‹ˆ๋‹ค:

1. Apache TVM ํ”„๋กœ์ ํŠธ์˜ ๊ธฐ์›:

  • Apache TVM์€ ์›Œ์‹ฑํ„ด ๋Œ€ํ•™๊ต์—์„œ OctoML์˜ ๊ณต๋™ ์ฐฝ๋ฆฝ์ž๋“ค์— ์˜ํ•ด ์‹œ์ž‘๋œ ํ”„๋กœ์ ํŠธ์ž…๋‹ˆ๋‹ค. ์ด๋Š” ํ•™๊ณ„์™€ ์‚ฐ์—…๊ณ„์˜ ์—ฐ๊ฒฐ์„ ๊ฐ•์กฐํ•˜๋Š” ์ข‹์€ ์˜ˆ์ž…๋‹ˆ๋‹ค.

2. ํ”„๋กœ์ ํŠธ์˜ ์„ฑ์žฅ๊ณผ ๊ธฐ์—ฌ์ž ์ˆ˜:

  • ํ˜„์žฌ๊นŒ์ง€ 645๋ช…์˜ ๊ธฐ์—ฌ์ž๊ฐ€ TVM ํ”„๋กœ์ ํŠธ์— ์ฐธ์—ฌํ•˜๊ณ  ์žˆ์œผ๋ฉฐ, ์ด๋Š” ์ง€์†์ ์œผ๋กœ ์ฆ๊ฐ€ํ•˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ์ด ์ปค๋ฎค๋‹ˆํ‹ฐ๋Š” ๋‹ค์–‘ํ•œ ํ•˜๋“œ์›จ์–ด ์ง€์›, ์—ฌ๋Ÿฌ ์ƒ์‚ฐ ๋ฐฐํฌ, ๊ทธ๋ฆฌ๊ณ  ๊ฐ•๋ ฅํ•œ ์ปค๋ฎค๋‹ˆํ‹ฐ ๋ฐ ์ƒํƒœ๊ณ„ ํ†ตํ•ฉ์„ ํ†ตํ•ด ์„ฑ์žฅํ•˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค.

์™ผ์ชฝ ๊ทธ๋ฆผ์€ Apache TVM๊ณผ ๊ด€๋ จ๋œ ๋‹ค์–‘ํ•œ ํ•˜๋“œ์›จ์–ด ํ”Œ๋žซํผ๊ณผ ํ”„๋ ˆ์ž„์›Œํฌ๋ฅผ ๋‚˜ํƒ€๋‚ด๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋ฆผ์—์„œ ์—ฌ๋Ÿฌ ํ•˜๋“œ์›จ์–ด ๋ฐ ๊ธฐ์ˆ  ๊ตฌ์„ฑ ์š”์†Œ๊ฐ€ ์„œ๋กœ ์—ฐ๊ฒฐ๋œ ๋ชจ์Šต์„ ํ†ตํ•ด Apache TVM์ด ์–ด๋–ป๊ฒŒ ๋‹ค์–‘ํ•œ ํ™˜๊ฒฝ์—์„œ ํ™œ์šฉ๋  ์ˆ˜ ์žˆ๋Š”์ง€๋ฅผ ์‹œ๊ฐ์ ์œผ๋กœ ํ‘œํ˜„ํ•˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ๊ตฌ์ฒด์ ์œผ๋กœ ๋‹ค์Œ๊ณผ ๊ฐ™์€ ์š”์†Œ๋“ค์ด ํฌํ•จ๋˜์–ด ์žˆ์Šต๋‹ˆ๋‹ค:

  1. ๋‹ค์–‘ํ•œ ํ•˜๋“œ์›จ์–ด ํ”Œ๋žซํผ:
    • ์—ฌ๋Ÿฌ ํ•˜๋“œ์›จ์–ด ๋””๋ฐ”์ด์Šค์™€ ์นฉ์…‹์ด ๋ณด์ž…๋‹ˆ๋‹ค. ์ด๋Š” TVM์ด ๋‹ค์–‘ํ•œ ์œ ํ˜•์˜ ํ•˜๋“œ์›จ์–ด, ์˜ˆ๋ฅผ ๋“ค์–ด GPU, CPU, ๋ชจ๋ฐ”์ผ ๋””๋ฐ”์ด์Šค ๋ฐ ํŠน์ˆ˜ ํ•˜๋“œ์›จ์–ด ํ”Œ๋žซํผ์— ์ ์šฉ๋  ์ˆ˜ ์žˆ์Œ์„ ๋‚˜ํƒ€๋ƒ…๋‹ˆ๋‹ค.
  1. ํ”„๋ ˆ์ž„์›Œํฌ ๋ฐ ๋„๊ตฌ ์—ฐ๊ฒฐ์„ฑ:
  • Vulkan, TensorRT, LLVM, cuDNN ๋“ฑ์˜ ๊ธฐ์ˆ  ์•„์ด์ฝ˜์€ TVM์ด ์ด๋“ค ๊ธฐ์ˆ ๊ณผ ํ†ตํ•ฉ๋˜์–ด ์ž‘๋™ํ•จ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ํ†ตํ•ฉ์„ ํ†ตํ•ด TVM์€ ๋‹ค์–‘ํ•œ ์ปดํ“จํŒ… ํ™˜๊ฒฝ๊ณผ ํ”„๋กœ๊ทธ๋ž˜๋ฐ ์–ธ์–ด์—์„œ์˜ ์ตœ์ ํ™”์™€ ์‹คํ–‰์„ ์ง€์›ํ•ฉ๋‹ˆ๋‹ค.
  1. DL Pack๊ณผ MLIR:
  • DL Pack๊ณผ MLIR ์•„์ด์ฝ˜์€ TVM ์ƒํƒœ๊ณ„ ๋‚ด์—์„œ ์ค‘์š”ํ•œ ์—ญํ• ์„ ํ•˜๋Š” ๊ตฌ์„ฑ ์š”์†Œ๋กœ ํ‘œํ˜„๋˜์–ด ์žˆ์Šต๋‹ˆ๋‹ค. ์ด๋“ค์€ ๋ชจ๋ธ์˜ ํฌ๋งท๊ณผ ์ตœ์ ํ™”๋ฅผ ๋„์™€, ๋” ๋„“์€ ๋ฒ”์œ„์˜ ํ•˜๋“œ์›จ์–ด ๋ฐ ์†Œํ”„ํŠธ์›จ์–ด ํ”Œ๋žซํผ์—์„œ์˜ ์‹คํ–‰์„ ๊ฐ€๋Šฅํ•˜๊ฒŒ ํ•ฉ๋‹ˆ๋‹ค.

์ด ๊ทธ๋ฆผ์€ TVM์ด ํ•˜๋“œ์›จ์–ด ๋…๋ฆฝ์ ์ธ ๋ฐฉ์‹์œผ๋กœ ๋จธ์‹ ๋Ÿฌ๋‹ ๋ชจ๋ธ์„ ์ตœ์ ํ™”ํ•˜๊ณ  ๋ฐฐํฌํ•  ์ˆ˜ ์žˆ๋„๋ก ์ง€์›ํ•˜๋Š” ๊ฐ•๋ ฅํ•œ ๋„๊ตฌ์ž„์„ ๊ฐ•์กฐํ•ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ, TVM์˜ ์œ ์—ฐ์„ฑ๊ณผ ๊ด‘๋ฒ”์œ„ํ•œ ์ ์šฉ ๊ฐ€๋Šฅ์„ฑ์„ ๋‚˜ํƒ€๋‚ด๋ฉฐ, ์ด๋ฅผ ํ†ตํ•ด ๊ฐœ๋ฐœ์ž๋“ค์ด ๋ณด๋‹ค ์‰ฝ๊ฒŒ ๋‹ค์–‘ํ•œ ๊ธฐ์ˆ  ์Šคํƒ๊ณผ ํ•˜๋“œ์›จ์–ด์—์„œ ์ตœ์ ์˜ ์„ฑ๋Šฅ์„ ๋‹ฌ์„ฑํ•  ์ˆ˜ ์žˆ๋„๋ก ๋•์Šต๋‹ˆ๋‹ค.

์ด ์ด๋ฏธ์ง€๋Š” OctoML์˜ ์กด์žฌ ์ด์œ ์™€ ๋ชฉํ‘œ๋ฅผ ์„ค๋ช…ํ•˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ์ œ๊ณต๋œ ์ •๋ณด์— ๋”ฐ๋ฅด๋ฉด OctoML์˜ ์ฃผ์š” ๋ชฉ์ ์€ ๋‹ค์Œ ๋‘ ๊ฐ€์ง€์ž…๋‹ˆ๋‹ค: 1/ ์‚ถ์„ ๊ฐœ์„ ํ•˜๊ธฐ ์œ„ํ•ด ์‹ ์ค‘ํ•˜๊ฒŒ ํ™œ์šฉ๋˜๋Š” ์ง€์† ๊ฐ€๋Šฅํ•˜๊ณ  ์ ‘๊ทผ ๊ฐ€๋Šฅํ•œ AI๋ฅผ ์ œ๊ณตํ•˜๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค. ์ด๋Š” OctoML์ด ์œค๋ฆฌ์ ์ด๊ณ  ๋ชจ๋“  ์‚ฌ๋žŒ๋“ค์ด ์ ‘๊ทผํ•  ์ˆ˜ ์žˆ์œผ๋ฉฐ ์‚ฌ๋žŒ๋“ค์˜ ์‚ถ์— ๊ธ์ •์ ์ธ ์˜ํ–ฅ์„ ๋ฏธ์น˜๋Š” AI ์†”๋ฃจ์…˜ ๊ฐœ๋ฐœ ๋ฐ ์ œ๊ณต์„ ๋ชฉํ‘œ๋กœ ํ•œ๋‹ค๋Š” ๊ฒƒ์„ ์˜๋ฏธํ•ฉ๋‹ˆ๋‹ค.

์ด ์Šฌ๋ผ์ด๋“œ๋Š” OctoML์˜ ์กด์žฌ ์ด์œ ์™€ ๊ทธ ๋ชฉ์ ์„ ๋ช…ํ™•ํžˆ ์„ค๋ช…ํ•˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. OctoML์ด ์ถ”๊ตฌํ•˜๋Š” ์ฃผ๋œ ๋ชฉํ‘œ๋Š” ๋‹ค์Œ๊ณผ ๊ฐ™์Šต๋‹ˆ๋‹ค:

  1. ์ง€์† ๊ฐ€๋Šฅํ•˜๊ณ  ์ ‘๊ทผ ๊ฐ€๋Šฅํ•œ AI ์ œ๊ณต(Offer sustainable and accessible AI used thoughtfully to improve lives):
    • OctoML์€ ์ง€์† ๊ฐ€๋Šฅํ•˜๊ณ  ์‚ฌ์šฉ์ž์—๊ฒŒ ์ ‘๊ทผ ๊ฐ€๋Šฅํ•œ ์ธ๊ณต์ง€๋Šฅ ์†”๋ฃจ์…˜์„ ์ œ๊ณตํ•˜๋Š” ๊ฒƒ์„ ๋ชฉํ‘œ๋กœ ํ•˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ AI ๊ธฐ์ˆ ์€ ์‚ฌ๋ ค ๊นŠ๊ฒŒ ์‚ฌ์šฉ๋˜์–ด ์‚ฌ๋žŒ๋“ค์˜ ์‚ถ์„ ๊ฐœ์„ ํ•˜๋Š” ๋ฐ ๊ธฐ์—ฌํ•˜๋„๋ก ์„ค๊ณ„๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ์ด๋Š” ๊ธฐ์ˆ ์ด ๋‹จ์ˆœํ•œ ๋น„์ฆˆ๋‹ˆ์Šค ๋„๊ตฌ๋ฅผ ๋„˜์–ด ์‚ฌํšŒ์ , ํ™˜๊ฒฝ์  ์˜ํ–ฅ์„ ๊ณ ๋ คํ•˜์—ฌ ๊ฐœ๋ฐœ๋˜์–ด์•ผ ํ•œ๋‹ค๋Š” ์ฒ ํ•™์„ ๋ฐ˜์˜ํ•ฉ๋‹ˆ๋‹ค.
  2. Apache TVM ์ƒํƒœ๊ณ„ ์„ฑ์žฅ ์ด‰์ง„ ๋ฐ ์ตœ์ ํ™” ํ”Œ๋žซํผ ๊ตฌ์ถ•(Catalyze Apache TVMโ€™s ecosystem growth and build a platform to enable anyone to easily deploy ML models on any hardware at peak performance):
    • OctoML์€ Apache TVM ์ƒํƒœ๊ณ„์˜ ์„ฑ์žฅ์„ ์ด‰์ง„ํ•˜๊ณ , ์ด๋ฅผ ํ†ตํ•ด ๋ˆ„๊ตฌ๋‚˜ ์‰ฝ๊ฒŒ ๋จธ์‹ ๋Ÿฌ๋‹ ๋ชจ๋ธ์„ ๋‹ค์–‘ํ•œ ํ•˜๋“œ์›จ์–ด์—์„œ ์ตœ์ ์˜ ์„ฑ๋Šฅ์œผ๋กœ ๋ฐฐํฌํ•  ์ˆ˜ ์žˆ๋Š” ํ”Œ๋žซํผ์„ ๊ตฌ์ถ•ํ•˜๋Š” ๊ฒƒ์„ ๋ชฉํ‘œ๋กœ ํ•ฉ๋‹ˆ๋‹ค. ์ด๋Š” ๊ธฐ์ˆ ์˜ ์ ‘๊ทผ์„ฑ์„ ๋†’์ด๊ณ , ์‚ฌ์šฉ์ž๊ฐ€ ํ•˜๋“œ์›จ์–ด์˜ ์ œ์•ฝ ์—†์ด ์ž์‹ ์˜ ์š”๊ตฌ์— ๋งž๋Š” ์†”๋ฃจ์…˜์„ ์„ ํƒํ•  ์ˆ˜ ์žˆ๊ฒŒ ํ•จ์œผ๋กœ์จ, ๋ณด๋‹ค ๋„“์€ ๋ฒ”์œ„์˜ ํ˜์‹ ๊ณผ ์‘์šฉ์„ ๊ฐ€๋Šฅํ•˜๊ฒŒ ํ•ฉ๋‹ˆ๋‹ค.

์ด ๋‘ ๋ชฉํ‘œ๋Š” OctoML์ด ๊ธฐ์ˆ  ๋ฐœ์ „๊ณผ ์‚ฌํšŒ์  ๊ฐ€์น˜ ์‚ฌ์ด์˜ ๊ท ํ˜•์„ ์ถ”๊ตฌํ•˜๋ฉฐ, ๋ณด๋‹ค ํฌ๊ด„์ ์ด๊ณ  ์ง€์† ๊ฐ€๋Šฅํ•œ ๋ฐฉ์‹์œผ๋กœ ์ธ๊ณต์ง€๋Šฅ ๊ธฐ์ˆ ์„ ํ™œ์šฉํ•˜๋ ค๋Š” ๊ทธ๋“ค์˜ ๋น„์ „์„ ๋ฐ˜์˜ํ•ฉ๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ์ ‘๊ทผ ๋ฐฉ์‹์€ ๊ธฐ์ˆ ์ด ๋‹จ์ˆœํ•œ ๋„๊ตฌ๊ฐ€ ์•„๋‹ˆ๋ผ, ์‹ค์ œ๋กœ ์‚ฌ๋žŒ๋“ค์˜ ์‚ถ์„ ๊ฐœ์„ ํ•˜๊ณ  ํ™•์žฅํ•  ์ˆ˜ ์žˆ๋Š” ์ˆ˜๋‹จ์œผ๋กœ ํ™œ์šฉ๋  ์ˆ˜ ์žˆ์Œ์„ ๊ฐ•์กฐํ•ฉ๋‹ˆ๋‹ค.

์ด ์Šฌ๋ผ์ด๋“œ๋Š” ๊ธฐ์ˆ  ๊ฐœ๋ฐœ์—์„œ ๋‹ฌ์„ฑํ•ด์•ผ ํ•  ์„ธ ๊ฐ€์ง€ ์ฃผ์š” ๋ชฉํ‘œ๋ฅผ ์„ค๋ช…ํ•˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค:

  1. ์„ฑ๋Šฅ(Performance): ๋ชจ๋ธ์ด ๋ฐฐํฌ ํ™˜๊ฒฝ์˜ ํ•˜๋“œ์›จ์–ด๋ฅผ ์ตœ๋Œ€ํ•œ ํ™œ์šฉํ•  ์ˆ˜ ์žˆ๋„๋ก ํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค. ์ด๋Š” ๊ธฐ์ˆ ์ด ํ•ด๋‹น ํ•˜๋“œ์›จ์–ด์—์„œ ์ตœ์ ์˜ ์„ฑ๋Šฅ์„ ๋ฐœํœ˜ํ•  ์ˆ˜ ์žˆ๋„๋ก ์ง€์›ํ•˜๋Š” ๊ฒƒ์„ ์˜๋ฏธํ•ฉ๋‹ˆ๋‹ค.
  2. ์ž๋™ํ™”(Automation): ๋“œ๋ฌผ๊ณ , ๋Š๋ฆฌ๋ฉฐ, ๋น„์šฉ์ด ๋งŽ์ด ๋“œ๋Š” ์—”์ง€๋‹ˆ์–ด๋ง์— ์˜์กดํ•˜์ง€ ์•Š๊ณ ๋„, ๊ธฐ์ˆ ์„ ์ž๋™์œผ๋กœ ๊ตฌํ˜„ ๋ฐ ์ตœ์ ํ™”ํ•  ํ•„์š”๊ฐ€ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด๋Š” ๊ณต์ •์„ ๋” ๋น ๋ฅด๊ณ  ๋น„์šฉ ํšจ์œจ์ ์œผ๋กœ ๋งŒ๋“ค๊ธฐ ์œ„ํ•จ์ž…๋‹ˆ๋‹ค.
  3. ์„ ํƒ๊ถŒ(Choice): ์–ด๋–ค ๋ชจ๋ธ์ด๋“  ์–ด๋–ค ํ•˜๋“œ์›จ์–ด์—์„œ๋„ ์‹คํ–‰ํ•  ์ˆ˜ ์žˆ์–ด์•ผ ํ•ฉ๋‹ˆ๋‹ค. ์ด๋Š” ๋‹ค์–‘ํ•œ ํ•˜๋“œ์›จ์–ด ์˜ต์…˜์— ๋Œ€ํ•œ ์œ ์—ฐ์„ฑ์„ ์ œ๊ณตํ•˜์—ฌ ์‚ฌ์šฉ์ž๊ฐ€ ์ž์‹ ์˜ ํ•„์š”์— ๊ฐ€์žฅ ์ ํ•ฉํ•œ ์†”๋ฃจ์…˜์„ ์„ ํƒํ•  ์ˆ˜ ์žˆ๋„๋ก ํ•ฉ๋‹ˆ๋‹ค.

์ด๋Ÿฌํ•œ ๋ชฉํ‘œ๋Š” ๊ธฐ์ˆ ์˜ ์ ‘๊ทผ์„ฑ๊ณผ ํšจ์œจ์„ฑ์„ ๋†’์ด๋Š” ๋ฐ ์ค‘์ ์„ ๋‘๊ณ  ์žˆ์œผ๋ฉฐ, ์ด๋Š” ๋” ๋„“์€ ๋ฒ”์œ„์˜ ์‚ฌ์šฉ์ž์™€ ํ™˜๊ฒฝ์—์„œ ๊ธฐ์ˆ ์„ ์‰ฝ๊ฒŒ ๋ฐฐํฌํ•˜๊ณ  ์ตœ์ ํ™”ํ•  ์ˆ˜ ์žˆ๋„๋ก ์ง€์›ํ•ฉ๋‹ˆ๋‹ค.

์ด ๋‹ค์ด์–ด๊ทธ๋žจ์€ OctoML ํ”Œ๋žซํผ์ด MLOps(๊ธฐ๊ณ„ ํ•™์Šต ์šด์˜) ํ๋ฆ„์—์„œ ์–ด๋–ป๊ฒŒ ํ†ตํ•ฉ๋˜๋Š”์ง€ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. MLOps์˜ ์ฃผ์š” ๋‹จ๊ณ„๋Š” ๋‹ค์Œ๊ณผ ๊ฐ™์ด ์ˆœํ™˜์ ์œผ๋กœ ๋ฐฐ์—ด๋ฉ๋‹ˆ๋‹ค:

  1. ๋ฐ์ดํ„ฐ ์ˆ˜์ง‘(Data Collection): ์ด ๋‹จ๊ณ„์—์„œ๋Š” ํ•„์š”ํ•œ ๋ฐ์ดํ„ฐ๋ฅผ ์ˆ˜์ง‘ํ•ฉ๋‹ˆ๋‹ค. ์ˆ˜์ง‘๋œ ๋ฐ์ดํ„ฐ๋Š” ๋‹ค์Œ ๋‹จ๊ณ„๋กœ ๋„˜์–ด๊ฐ€๊ธฐ ์ „์— ์ฒ˜๋ฆฌ๋ฉ๋‹ˆ๋‹ค.
  2. ์ •์ œ ๋ฐ ์ฃผ์„ ์ฒ˜๋ฆฌ(Cleaning & Annotation): ์ˆ˜์ง‘๋œ ๋ฐ์ดํ„ฐ๋ฅผ ์ •์ œํ•˜๊ณ  ํ•„์š”ํ•œ ๊ฒฝ์šฐ ์ฃผ์„์„ ๋‹ฌ์•„ ๋ฐ์ดํ„ฐ์˜ ํ’ˆ์งˆ์„ ๋†’์ด๊ณ  ๋ชจ๋ธ ํ›ˆ๋ จ์— ์ ํ•ฉํ•˜๊ฒŒ ๋งŒ๋“ญ๋‹ˆ๋‹ค.
  3. ๋ชจ๋ธ ์ƒ์„ฑ ๋ฐ ํ›ˆ๋ จ(Model Creation & Training): ์ •์ œ๋œ ๋ฐ์ดํ„ฐ๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ๊ธฐ๊ณ„ ํ•™์Šต ๋ชจ๋ธ์„ ์ƒ์„ฑํ•˜๊ณ  ํ›ˆ๋ จํ•ฉ๋‹ˆ๋‹ค.
  4. ์„ฑ๋Šฅ ์ž๋™ํ™”(Performance Automation): ๋ชจ๋ธ์˜ ์„ฑ๋Šฅ์„ ์ž๋™์œผ๋กœ ์ตœ์ ํ™”ํ•˜์—ฌ ๋ฐฐํฌ๋œ ๋ชจ๋ธ์ด ์ตœ๋Œ€ํ•œ์˜ ํšจ์œจ์„ ๋‚ผ ์ˆ˜ ์žˆ๋„๋ก ํ•ฉ๋‹ˆ๋‹ค.
  5. ํŒจํ‚ค์ง• ๋ฐ ์• ํ”Œ๋ฆฌ์ผ€์ด์…˜ ํ†ตํ•ฉ(Packaging & App Integration): ๋ชจ๋ธ์„ ์• ํ”Œ๋ฆฌ์ผ€์ด์…˜๊ณผ ํ†ตํ•ฉํ•˜๊ณ  ์‚ฌ์šฉ์ž๊ฐ€ ์‚ฌ์šฉํ•  ์ˆ˜ ์žˆ๋„๋ก ํŒจํ‚ค์ง•ํ•ฉ๋‹ˆ๋‹ค.
  6. ๋ฐฐํฌ(Deployment): ํ›ˆ๋ จ๋œ ๋ชจ๋ธ์„ ์‹ค์ œ ์šด์˜ ํ™˜๊ฒฝ์— ๋ฐฐํฌํ•ฉ๋‹ˆ๋‹ค.

์ด ํ๋ฆ„์€ ์ง€์†์ ์ธ ๊ฐœ์„ ๊ณผ ํšจ์œจ์ ์ธ ๋ชจ๋ธ ๊ด€๋ฆฌ๋ฅผ ์œ„ํ•ด ์ˆœํ™˜์ ์œผ๋กœ ์„ค๊ณ„๋˜์—ˆ์Šต๋‹ˆ๋‹ค. OctoML ํ”Œ๋žซํผ์€ 4, 5, 6๋ฒˆ ๋‹จ๊ณ„์—์„œ ์ž๋™ํ™” ๋ฐ ์ตœ์ ํ™”๋ฅผ ์ œ๊ณตํ•˜์—ฌ ์ „์ฒด MLOps ํ”„๋กœ์„ธ์Šค๋ฅผ ๊ฐ„์†Œํ™”ํ•˜๊ณ  ํ–ฅ์ƒ์‹œํ‚ต๋‹ˆ๋‹ค.

  • ML (๋จธ์‹ ๋Ÿฌ๋‹) ํ˜์‹ ์„ ์œ„ํ•ด์„œ๋Š” ํผํฌ๋จผ์Šค๊ฐ€ ํ•ต์‹ฌ์ ์ž…๋‹ˆ๋‹ค.

์ด ์Šฌ๋ผ์ด๋“œ๋Š” ์ธ๊ณต์ง€๋Šฅ๊ณผ ๋จธ์‹ ๋Ÿฌ๋‹(AI/ML)์˜ ํ™˜๊ฒฝ์  ์˜ํ–ฅ์— ๋Œ€ํ•œ ์ค‘์š”ํ•œ ๋ฌธ์ œ๋“ค์„ ์กฐ๋ช…ํ•˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค.

  1. ์ˆ˜์š”๋ฅผ ๋”ฐ๋ผ์žก๊ธฐ ์œ„ํ•ด ๋น ๋ฅด๊ฒŒ ์„ฑ์žฅํ•˜๋Š” ์šฉ๋Ÿ‰(Capacity growing fast to keep up with demand): AI ๋ฐ ML ๊ธฐ์ˆ ์ด ๋ฐœ์ „ํ•˜๋ฉด์„œ ์ด๋ฅผ ์œ„ํ•œ ์ธํ”„๋ผ์™€ ์ฒ˜๋ฆฌ ๋Šฅ๋ ฅ์ด ๊ธ‰์†ํžˆ ์ฆ๊ฐ€ํ•˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ์ด๋Š” ๊ธฐ์ˆ ์ด ๊ณ„์†ํ•ด์„œ ๋ฐœ์ „ํ•˜๊ณ  ์ˆ˜์š”๊ฐ€ ์ฆ๊ฐ€ํ•จ์— ๋”ฐ๋ผ ํ•„์š”ํ•œ ์ž์›๋„ ๋Š˜์–ด๋‚˜๊ณ  ์žˆ์Œ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค.
  2. ๋น„๊ต์  ์—„์ฒญ๋‚œ ํƒ„์†Œ ๋ฐœ์ž๊ตญ(Carbon footprint is humongous by comparison): ๋Œ€๊ทœ๋ชจ์˜ ์ถ”์ฒœ ๋ชจ๋ธ์„ ์šด์˜ํ•  ๊ฒฝ์šฐ ๋ฐœ์ƒํ•˜๋Š” ํƒ„์†Œ ๋ฐœ์ž๊ตญ์€ ์ƒ๋‹นํžˆ ํฝ๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, ์ด๋Š” 170๊ฐ€๊ตฌ์˜ ๋ฏธ๊ตญ ๊ฐ€์ •์ด 1๋…„ ๋™์•ˆ ์‚ฌ์šฉํ•˜๋Š” ์—๋„ˆ์ง€ ์–‘๊ณผ ๋งž๋จน๊ฑฐ๋‚˜, ํƒ„์†Œ ๋ฐฐ์ถœ์„ ์ƒ์‡„ํ•˜๊ธฐ ์œ„ํ•ด 7,500๊ทธ๋ฃจ์˜ ๋‚˜๋ฌด๋ฅผ ์‹ฌ์–ด์•ผ ํ•  ์ •๋„์ž…๋‹ˆ๋‹ค. ์ด๋Š” AI ๊ธฐ์ˆ ์˜ ํ™•์žฅ์ด ํ™˜๊ฒฝ์— ๋ฏธ์น˜๋Š” ์˜ํ–ฅ์„ ๊ฐ•์กฐํ•˜๋Š” ์ค‘์š”ํ•œ ์ง€ํ‘œ์ž…๋‹ˆ๋‹ค.
  3. ๊ฐœ์ธ์ ์ธ ๋ฌธ์ œ(This, to me, is personal): ์Šฌ๋ผ์ด๋“œ์˜ ๋งˆ์ง€๋ง‰ ๋ถ€๋ถ„์—์„œ๋Š” ๋ฐœํ‘œ์ž๊ฐ€ ์ด ๋ฌธ์ œ๋ฅผ ๊ฐœ์ธ์ ์œผ๋กœ ๋ฐ›์•„๋“ค์ด๊ณ  ์žˆ๋‹ค๋Š” ์ ์„ ๊ฐ•์กฐํ•˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ์—ฌ๊ธฐ์— ํ‘œ์‹œ๋œ ์‚ฐ๋ฆผ ์ด๋ฏธ์ง€๋Š” ์ž์—ฐ์„ ๋ณดํ˜ธํ•˜๊ณ  ์ง€์† ๊ฐ€๋Šฅํ•œ ๊ธฐ์ˆ  ๊ฐœ๋ฐœ์„ ์ถ”๊ตฌํ•˜๋Š” ๊ฒƒ์ด ์–ผ๋งˆ๋‚˜ ์ค‘์š”ํ•œ์ง€๋ฅผ ์ƒ์ง•์ ์œผ๋กœ ๋ณด์—ฌ์ฃผ๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค.

์ด ์Šฌ๋ผ์ด๋“œ๋Š” AI์™€ ML์˜ ๋ฐœ์ „์ด ๊ฐ€์ ธ์˜ค๋Š” ํ™˜๊ฒฝ์  ๋„์ „์„ ์ธ์‹ํ•˜๊ณ , ์ด์— ๋Œ€ํ•œ ์ฑ…์ž„ ์žˆ๋Š” ๋Œ€์‘์„ ๊ฐ•์กฐํ•˜๋Š” ๋‚ด์šฉ์„ ๋‹ด๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค.

์ด ์Šฌ๋ผ์ด๋“œ๋Š” ์„ฑ๋Šฅ์˜ ์ค‘์š”์„ฑ๊ณผ ๊ทธ๊ฒƒ์ด ๋น„์šฉ, ์—๋„ˆ์ง€ ์‚ฌ์šฉ, ๊ทธ๋ฆฌ๊ณ  ์ „์ฒด ์ธํ”„๋ผ์— ๋ฏธ์น˜๋Š” ์˜ํ–ฅ์— ๋Œ€ํ•ด ์„ค๋ช…ํ•˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ๊ฐ ํฌ์ธํŠธ๋ฅผ ํ•œ๊ตญ์–ด๋กœ ์„ค๋ช…ํ•˜๋ฉด ๋‹ค์Œ๊ณผ ๊ฐ™์Šต๋‹ˆ๋‹ค:

  1. 2๋ฐฐ ๋น ๋ฆ„(Faster): ์ฝ”๋“œ๊ฐ€ ๋ฐ์ดํ„ฐ ์„ผํ„ฐ์—์„œ ์ „ ์„ธ๊ณ„์ ์œผ๋กœ ์‹คํ–‰๋  ๋•Œ์˜ ์ƒํ™ฉ์„ ์ƒ์ƒํ•ด ๋ณด์„ธ์š”. ์ด๋Š” ํ”„๋กœ์„ธ์Šค๋ฅผ ๊ฐ€์†ํ™”ํ•˜์—ฌ ์ „์ฒด์ ์ธ ์ž‘์—… ์†๋„๋ฅผ ํ–ฅ์ƒ์‹œํ‚ต๋‹ˆ๋‹ค.
  2. ์—๋„ˆ์ง€ ์‚ฌ์šฉ 1/2(Energy Use): ์„ธ๊ณ„ ๊ณณ๊ณณ์— ํฉ์–ด์ ธ ์žˆ๋Š” ์ˆ˜๋งŽ์€ ์žฅ์น˜๋ฅผ ์ž‘๋™์‹œํ‚ค๊ธฐ ์œ„ํ•ด ํ•„์š”ํ•œ ๋ฐฐํ„ฐ๋ฆฌ๋ฅผ ์ƒ์ƒํ•ด ๋ณด์„ธ์š”. ํšจ์œจ์ ์ธ ์„ฑ๋Šฅ์€ ์ด๋Ÿฌํ•œ ์žฅ์น˜๋“ค์ด ์†Œ๋น„ํ•˜๋Š” ์—๋„ˆ์ง€ ์–‘์„ ์ค„์ด๋Š” ๋ฐ ํฐ ๋„์›€์„ ์ค๋‹ˆ๋‹ค.
  3. ์ธํ”„๋ผ 1/2(Infrastructure): ์ง€๊ตฌ์™€ ํ–‰๋ณตํ•œ ๋‚˜๋ฌด๋“ค์„ ์ƒ๊ฐํ•ด ๋ณด์„ธ์š”. ์„ฑ๋Šฅ ์ตœ์ ํ™”๋Š” ํ•„์š”ํ•œ ์ธํ”„๋ผ์˜ ์–‘์„ ์ค„์ด๊ณ , ๊ฒฐ๊ณผ์ ์œผ๋กœ ํ™˜๊ฒฝ์— ๋ฏธ์น˜๋Š” ๋ถ€์ •์ ์ธ ์˜ํ–ฅ์„ ๊ฐ์†Œ์‹œํ‚ฌ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
  4. ๋น„์šฉ 1/2(Cost): ํšจ์œจ์ ์ธ ์„ฑ๋Šฅ์€ ๋น„์šฉ์„ ์ ˆ๋ฐ˜์œผ๋กœ ์ค„์ด๋Š” ๋ฐ ๊ธฐ์—ฌํ•ฉ๋‹ˆ๋‹ค. ์ด๋Š” ๊ธฐ์—…์˜ ์ž์›์„ ์ ˆ์•ฝํ•˜๊ณ  ์žฅ๊ธฐ์ ์œผ๋กœ ๊ฒฝ์ œ์  ์ง€์† ๊ฐ€๋Šฅ์„ฑ์„ ๋‹ฌ์„ฑํ•˜๋Š” ๋ฐ ๋„์›€์ด ๋ฉ๋‹ˆ๋‹ค.

์ „์ฒด์ ์œผ๋กœ ์ด ์Šฌ๋ผ์ด๋“œ๋Š” ์„ฑ๋Šฅ ์ตœ์ ํ™”๊ฐ€ ๊ธฐ์ˆ ์˜ ํ™•์žฅ์„ฑ, ๋น„์šฉ ํšจ์œจ์„ฑ, ๊ทธ๋ฆฌ๊ณ  ํ™˜๊ฒฝ ๋ณดํ˜ธ์— ์–ผ๋งˆ๋‚˜ ์ค‘์š”ํ•œ์ง€ ๊ฐ•์กฐํ•˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ์ด๋Š” "์„ฑ๋Šฅ์ด ์ค‘์š”ํ•˜๋‹ค(Performance is critical)"๋ผ๋Š” ๋ฉ”์‹œ์ง€๋ฅผ ํ†ตํ•ด ๊ฐ•์กฐ๋ฉ๋‹ˆ๋‹ค.

์ด ์Šฌ๋ผ์ด๋“œ๋Š” ๋จธ์‹ ๋Ÿฌ๋‹ ๋ชจ๋ธ์˜ ์„ฑ๋Šฅ ์ตœ์ ํ™”์˜ ์–ด๋ ค์›€์„ ์„ค๋ช…ํ•˜๋ฉฐ, ๋ชจ๋ธ์˜ ํŠน์„ฑ๊ณผ ์ตœ์ ํ™”๋ฅผ ์œ„ํ•ด ํ•„์š”ํ•œ ์ง€์‹์„ ๋Œ€๋น„ํ•˜์—ฌ ์„ค๋ช…ํ•ฉ๋‹ˆ๋‹ค.

์™ผ์ชฝ ๋ถ€๋ถ„: ML ๋ชจ๋ธ์˜ ํŠน์„ฑ๊ณผ ๋„์ „(Left Part: Characteristics and Challenges of ML Models)

  • ML ๋ชจ๋ธ = ์ฝ”๋“œ + ๋ฐ์ดํ„ฐ(ML Model = Code + Data): ๋ชจ๋ธ์€ ์ฝ”๋“œ์™€ ๋ฐ์ดํ„ฐ๋กœ ๊ตฌ์„ฑ๋ฉ๋‹ˆ๋‹ค.
  • ๋ฐฐํฌ ํ•˜๋“œ์›จ์–ด ๋ฐ ์ธํ”„๋ผ์— ๋Œ€ํ•œ ๋ฏผ๊ฐ๋„(Very sensitive to deployment HW and infrastructure): ๋ชจ๋ธ์€ ์‚ฌ์šฉํ•˜๋Š” ํ•˜๋“œ์›จ์–ด์™€ ์ธํ”„๋ผ์— ๋ฏผ๊ฐํ•ฉ๋‹ˆ๋‹ค.
  • ๋ฉ”๋ชจ๋ฆฌ ๋ฐ ๊ณ„์‚ฐ ์ž์›์— ๋Œ€ํ•œ ์ŠคํŠธ๋ ˆ์Šค(Stresses memory, compute/communication resources): ๋ชจ๋ธ์€ ๋งŽ์€ ๋ฉ”๋ชจ๋ฆฌ์™€ ๊ณ„์‚ฐ ์ž์›์„ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค.
  • ์ •ํ™•๋„์™€ ์ž์› ์‚ฌ์šฉ์˜ ์ƒ์ถฉ ๊ด€๊ณ„(It can trade-off accuracy for lower resource usage): ๋•Œ๋กœ๋Š” ๋” ๋‚ฎ์€ ์ž์› ์‚ฌ์šฉ์„ ์œ„ํ•ด ์ •ํ™•๋„๋ฅผ ํฌ์ƒํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.

์˜ค๋ฅธ์ชฝ ๋ถ€๋ถ„: ์„ฑ๋Šฅ ์ตœ์ ํ™”๋ฅผ ์œ„ํ•œ ํ•„์š” ์ง€์‹(Right Part: Required Knowledge for Performance Optimization)

  • ๋จธ์‹ ๋Ÿฌ๋‹ ๋ชจ๋ธ ์„ธ๋ถ€ ์ •๋ณด(Machine learning model details): ์—”์ง€๋‹ˆ์–ด๋Š” ๋ชจ๋ธ์˜ ๊ตฌ์กฐ์™€ ์ž‘๋™ ๋ฐฉ์‹์„ ์•Œ์•„์•ผ ํ•ฉ๋‹ˆ๋‹ค.
  • ์ตœ์ ํ™” ๊ธฐ์ˆ (Optimization techniques): ์ง€์—ญ์„ฑ ์ตœ์ ํ™” (locality optimizations), ์ฝ”๋“œ ์Šค์ผ€์ค„๋ง (code scheduling), ์–‘์žํ™” (quantization)๋“ฑ์˜ ๊ธฐ์ˆ ์„ ์ ์šฉํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค.
  • ์ปดํŒŒ์ผ๋Ÿฌ ๋ฐ ์ปดํ“จํ„ฐ ๊ตฌ์กฐ(Compilers and even computer architecture): ์†Œํ”„ํŠธ์›จ์–ด์™€ ํ•˜๋“œ์›จ์–ด ์ตœ์ ํ™”๋ฅผ ์œ„ํ•œ ๊นŠ์€ ์ดํ•ด๊ฐ€ ํ•„์š”ํ•ฉ๋‹ˆ๋‹ค.

์ด ์„ค๋ช…์€ ์„ฑ๋Šฅ ์ตœ์ ํ™”์˜ ์ค‘์š”์„ฑ์„ ๊ฐ•์กฐํ•˜๊ณ , ML ๋ชจ๋ธ์„ ํšจ๊ณผ์ ์œผ๋กœ ๊ฐœ๋ฐœํ•˜๊ณ  ์šด์˜ํ•˜๋Š” ๋ฐ ํ•„์š”ํ•œ ๊ธฐ์ˆ ์  ๊นŠ์ด์™€ ์ง€์‹์„ ์ œ์‹œํ•ฉ๋‹ˆ๋‹ค.


์ด ์Šฌ๋ผ์ด๋“œ๋Š” ๋จธ์‹ ๋Ÿฌ๋‹์˜ ์ตœ์ ํ™”์™€ ๋ฐฐํฌ ๊ณผ์ •์„ ๋” ๋งŽ์€ ์‚ฌ๋žŒ๋“ค์ด ํ•  ์ˆ˜ ์žˆ๋„๋ก ๋งŒ๋“œ๋Š” ๊ฒƒ์˜ ์ค‘์š”์„ฑ์„ ์„ค๋ช…ํ•˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค.

  • ์ถฉ๋ถ„ํ•œ ์‚ฌ๋žŒ๋“ค์ด ๋ชจ๋ธ ์ตœ์ ํ™”์™€ ๋ฐฐํฌ๋ฅผ ์•Œ์ง€ ๋ชปํ•จ(Not enough people know how to optimize and deploy models): ๋งŽ์€ ์‚ฌ๋žŒ๋“ค์ด ์•„์ง๊นŒ์ง€ ๋จธ์‹ ๋Ÿฌ๋‹ ๋ชจ๋ธ์˜ ์ตœ์ ํ™”์™€ ๋ฐฐํฌ ๋ฐฉ๋ฒ•์— ๋Œ€ํ•œ ์ถฉ๋ถ„ํ•œ ์ง€์‹์ด๋‚˜ ๊ธฐ์ˆ ์„ ๊ฐ–์ถ”์ง€ ๋ชปํ•˜๊ณ  ์žˆ๋‹ค๋Š” ํ˜„์‹ค์„ ์ง€์ ํ•ฉ๋‹ˆ๋‹ค.
    • ์ด ์ƒํ™ฉ์€ ์ง€์†๋  ์ˆ˜ ์—†์Œ(It canโ€™t stay that way or it will be in the hands of too few players): ๋จธ์‹ ๋Ÿฌ๋‹ ๊ธฐ์ˆ ๊ณผ ์ง€์‹์ด ์†Œ์ˆ˜์— ์˜ํ•ด์„œ๋งŒ ํ†ต์ œ๋˜๋Š” ๊ฒƒ์„ ๋ฐฉ์ง€ํ•˜๊ณ , ์ด ๋ถ„์•ผ๋ฅผ ๋” ๋งŽ์€ ์‚ฌ๋žŒ๋“ค์—๊ฒŒ ๊ฐœ๋ฐฉํ•  ํ•„์š”๊ฐ€ ์žˆ๋‹ค๊ณ  ๊ฐ•์กฐํ•ฉ๋‹ˆ๋‹ค.
    • ML ์„ฑ๋Šฅ์„ ๋„“์€ ๋ฒ”์œ„์˜ ์‹ค๋ฌด์ž๋“ค์—๊ฒŒ ์ ‘๊ทผ ๊ฐ€๋Šฅํ•˜๊ฒŒ ํ•ด์•ผ ํ•จ(Need to make performance in ML accessible to a wide range of practitioners): ๋จธ์‹ ๋Ÿฌ๋‹์˜ ์„ฑ๋Šฅ ์ตœ์ ํ™” ๊ณผ์ •์„ ๋” ๋งŽ์€ ์‹ค๋ฌด์ž๋“ค์ด ์ดํ•ดํ•˜๊ณ  ์ ‘๊ทผํ•  ์ˆ˜ ์žˆ๋„๋ก ์ง€์‹๊ณผ ๋„๊ตฌ๋ฅผ ๋ณด๊ธ‰ํ•ด์•ผ ํ•œ๋‹ค๋Š” ์ ์„ ๊ฐ•์กฐํ•ฉ๋‹ˆ๋‹ค.
  • ๋” ๋งŽ์€ ์‚ฌ๋žŒ๋“ค์ด ๋น ๋ฅด๊ฒŒ ML์„ ์ˆ˜ํ–‰ํ•  ์ˆ˜ ์žˆ๊ฒŒ ๋จ(More people able to do fast ML): ์ง€์‹ ํ™•์‚ฐ์„ ํ†ตํ•ด ๋” ๋งŽ์€ ์‚ฌ๋žŒ๋“ค์ด ๋จธ์‹ ๋Ÿฌ๋‹์„ ๋น ๋ฅด๊ณ  ํšจ๊ณผ์ ์œผ๋กœ ์ˆ˜ํ–‰ํ•  ์ˆ˜ ์žˆ๊ฒŒ ๋จ์œผ๋กœ์จ, ๊ธฐ์ˆ ์˜ ์‚ฌ์šฉ ๋ฒ”์œ„์™€ ํšจ์œจ์„ฑ์„ ์ฆ๊ฐ€์‹œํ‚ฌ ์ˆ˜ ์žˆ๋‹ค๋Š” ๊ธฐ๋Œ€๋ฅผ ํ‘œํ•ฉ๋‹ˆ๋‹ค.
  • ML์—์„œ์˜ ํ˜์‹  ์ฆ๊ฐ€(More innovation in ML): ์ง€์‹๊ณผ ๋„๊ตฌ์˜ ํ™•์‚ฐ์ด ๋” ๋‹ค์–‘ํ•œ ๋ฐฐ๊ฒฝ์„ ๊ฐ€์ง„ ์‚ฌ๋žŒ๋“ค๋กœ๋ถ€ํ„ฐ์˜ ์ƒˆ๋กœ์šด ์•„์ด๋””์–ด์™€ ํ˜์‹ ์„ ๊ฐ€๋Šฅํ•˜๊ฒŒ ํ•จ์œผ๋กœ์จ, ML ๋ถ„์•ผ ์ „๋ฐ˜์— ๊ฑธ์ณ ๋ฐœ์ „์„ ์ด‰์ง„ํ•  ์ˆ˜ ์žˆ๋‹ค๊ณ  ์„ค๋ช…ํ•ฉ๋‹ˆ๋‹ค.

์ด๋Ÿฌํ•œ ์„ธ๋ถ€ ์‚ฌํ•ญ๋“ค์€ ๋จธ์‹ ๋Ÿฌ๋‹์˜ ์ตœ์ ํ™” ๋ฐ ๋ฐฐํฌ ๊ณผ์ •์„ ๋” ๋งŽ์€ ์‚ฌ๋žŒ๋“ค์—๊ฒŒ ์•Œ๋ ค, ์ „์ฒด์ ์ธ ํ˜์‹ ๊ณผ ๋ฐœ์ „์„ ๋„๋ชจํ•˜๊ธฐ ์œ„ํ•œ ํ•„์š”์„ฑ์„ ๊ฐ•์กฐํ•˜๋Š” ๋ฐ ์ค‘์ ์„ ๋‘ก๋‹ˆ๋‹ค.

  • ์ž๋™ํ™”๋Š” ML Deployment๋ฅผ Scalingํ•˜๋Š”๋ฐ ์žˆ์–ด ํ•ต์‹ฌ์ž…๋‹ˆ๋‹ค.

์ด ์Šฌ๋ผ์ด๋“œ๋Š” ๋จธ์‹ ๋Ÿฌ๋‹ ๋ชจ๋ธ ๋ฐฐํฌ๋ฅผ ์œ„ํ•œ ์ˆ˜์ž‘์—… ์—”์ง€๋‹ˆ์–ด๋ง์˜ ํ•œ๊ณ„์™€ ์ž๋™ํ™”์˜ ํ•„์š”์„ฑ์„ ๋‹ค๋ฃจ๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ์—ฌ๊ธฐ์—๋Š” ํฌ๊ฒŒ ๋‘ ๋ถ€๋ถ„์˜ ๋‚ด์šฉ์ด ํฌํ•จ๋˜์–ด ์žˆ์Šต๋‹ˆ๋‹ค:

1. ์ˆ˜์ž‘์—… ์—”์ง€๋‹ˆ์–ด๋ง์˜ ๋ฌธ์ œ์ (Hand Engineering for ML Deployment Doesnโ€™t Scale):

  • ๋ชจ๋ธ์„ ์ƒ์‚ฐ์œผ๋กœ ๊ฐ€์ ธ๊ฐ€๋Š” ๋ฐ ๋ช‡ ๋‹ฌ์ด ๊ฑธ๋ฆด ์ˆ˜ ์žˆ์Œ(Getting a model to production can take months!): ๋จธ์‹ ๋Ÿฌ๋‹ ๋ชจ๋ธ์„ ์‹ค์ œ ์ƒ์‚ฐ ํ™˜๊ฒฝ์— ๋ฐฐํฌํ•˜๋Š” ๋ฐ๋Š” ์ˆ˜๊ฐœ์›”์ด ์†Œ์š”๋  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด ๊ณผ์ •์€ ๋ณต์žกํ•˜๊ณ  ์‹œ๊ฐ„์ด ๋งŽ์ด ๊ฑธ๋ฆฝ๋‹ˆ๋‹ค.
  • AI/ML์˜ ์„ฑ์žฅ ์ถ”์„ธ๋ฅผ ๋”ฐ๋ผ์žก์„ ์ˆ˜ ์žˆ๋Š” ์—”์ง€๋‹ˆ์–ด๊ฐ€ ๋ถ€์กฑํ•จ(And given the growth trends in AI/ML, there are not enough engineers to keep up with demand): AI ๋ฐ ML ๋ถ„์•ผ์˜ ๊ธ‰์†ํ•œ ์„ฑ์žฅ์— ๋น„ํ•ด ์ถฉ๋ถ„ํ•œ ์ˆ˜์˜ ์—”์ง€๋‹ˆ์–ด๊ฐ€ ์—†์–ด, ์ˆ˜์š”๋ฅผ ์ถฉ์กฑ์‹œํ‚ค๊ธฐ ์–ด๋ ต์Šต๋‹ˆ๋‹ค.

2. ์ž๋™ํ™”๊ฐ€ ํ•„์š”ํ•œ ๋ถ€๋ถ„(What needs to be automated?):

  • ๋ชจ๋“  ๋ฐฐํฌ ํ•˜๋“œ์›จ์–ด ์˜ต์…˜์—์„œ ๋ชจ๋ธ ๊ตฌํ˜„์„ ์กฐ์ •ํ•ด์•ผ ํ•จ(Tuning model implementation on all deployment hardware options): ๋‹ค์–‘ํ•œ ํ•˜๋“œ์›จ์–ด ์˜ต์…˜์—์„œ ๋ชจ๋ธ์„ ์ตœ์ ํ™”ํ•˜๊ณ  ์กฐ์ •ํ•˜๋Š” ๊ณผ์ •์„ ์ž๋™ํ™”ํ•  ํ•„์š”๊ฐ€ ์žˆ์Šต๋‹ˆ๋‹ค.
  • ๊ทœ๋ชจ์— ๋งž๋Š” ๋ฐฐํฌ๋ฅผ ์œ„ํ•ด ์ตœ์ ์˜ ํ•˜๋“œ์›จ์–ด ์„ ํƒํ•˜๊ธฐ(Choosing what hardware is best for deployment at scale): ๊ฐ€์žฅ ํšจ์œจ์ ์ธ ํ•˜๋“œ์›จ์–ด๋ฅผ ์„ ํƒํ•˜๋Š” ๊ฒƒ์€ ์ค‘์š”ํ•œ ๊ฒฐ์ •์ž…๋‹ˆ๋‹ค. ์—ฌ๊ธฐ์—๋Š” ์ฒ˜๋ฆฌ๋Ÿ‰ ๋Œ€๋น„ ์—๋„ˆ์ง€ ํšจ์œจ(throughput/watt), ์ฒ˜๋ฆฌ๋Ÿ‰ ๋Œ€๋น„ ๋น„์šฉ(throughput/$), ๋˜๋Š” ์‘๋‹ต ์‹œ๊ฐ„(latency) ๋“ฑ์ด ๊ณ ๋ ค๋ฉ๋‹ˆ๋‹ค.
  • ๋ฐฐํฌ๋ฅผ ์‰ฝ๊ฒŒ ํ•˜๊ธฐ ์œ„ํ•œ ๋ชจ๋ธ ํŒจํ‚ค์ง•(Packaging model for easy deployment): ๋ชจ๋ธ์„ ์‰ฝ๊ฒŒ ๋ฐฐํฌํ•  ์ˆ˜ ์žˆ๋„๋ก ํŒจํ‚ค์ง•ํ•˜๋Š” ๊ณผ์ •๋„ ์ž๋™ํ™”๋˜์–ด์•ผ ํ•ฉ๋‹ˆ๋‹ค.

์ด ์Šฌ๋ผ์ด๋“œ๋Š” ๋จธ์‹ ๋Ÿฌ๋‹ ๋ชจ๋ธ ๋ฐฐํฌ ๊ณผ์ •์—์„œ ์ž๋™ํ™”์˜ ์ค‘์š”์„ฑ์„ ๊ฐ•์กฐํ•˜๊ณ  ์žˆ์œผ๋ฉฐ, ์ˆ˜์ž‘์—…์œผ๋กœ ์ฒ˜๋ฆฌํ•˜๊ธฐ ์–ด๋ ค์šด ๋‹ค์–‘ํ•œ ๊ณผ์ œ๋“ค์„ ํ•ด๊ฒฐํ•˜๊ธฐ ์œ„ํ•ด ์ž๋™ํ™” ๊ธฐ์ˆ ์„ ์ ์šฉํ•  ํ•„์š”๊ฐ€ ์žˆ์Œ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ์ด๋Š” ์ƒ์‚ฐ์„ฑ์„ ๋†’์ด๊ณ , ๋” ๋น ๋ฅด๊ณ  ํšจ์œจ์ ์ธ ๋ชจ๋ธ ๋ฐฐํฌ๋ฅผ ๊ฐ€๋Šฅํ•˜๊ฒŒ ํ•ฉ๋‹ˆ๋‹ค.

์ด ์Šฌ๋ผ์ด๋“œ๋Š” Apache TVM์„ ์‚ฌ์šฉํ•œ ๋จธ์‹ ๋Ÿฌ๋‹ ๋ชจ๋ธ์˜ ์„ฑ๋Šฅ ์ž๋™ํ™”์— ๋Œ€ํ•ด ์„ค๋ช…ํ•˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ์ฃผ์š” ๋‚ด์šฉ์„ ์š”์•ฝํ•˜๋ฉด ๋‹ค์Œ๊ณผ ๊ฐ™์Šต๋‹ˆ๋‹ค:

1. ML ๋ชจ๋ธ(ML Model):

  • ๋จธ์‹ ๋Ÿฌ๋‹ ๋ชจ๋ธ์˜ ์ถœ๋ฐœ์ ์„ ๋‚˜ํƒ€๋ƒ…๋‹ˆ๋‹ค. ์ด ๋ชจ๋ธ์€ ๋‹ค์–‘ํ•œ ํ”„๋ ˆ์ž„์›Œํฌ์™€ ํ™˜๊ฒฝ์—์„œ ๊ฐœ๋ฐœ๋  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.

2. TVM ์‚ฌ์šฉ(TVM: ML-based optimizations to obviate need for hand-tuning for target HW):

  • Apache TVM์€ ๋จธ์‹ ๋Ÿฌ๋‹ ๊ธฐ๋ฐ˜์˜ ์ตœ์ ํ™”๋ฅผ ์ œ๊ณตํ•˜์—ฌ, ํŠน์ • ํ•˜๋“œ์›จ์–ด๋ฅผ ์œ„ํ•œ ์ˆ˜์ž‘์—… ํŠœ๋‹์˜ ํ•„์š”์„ฑ์„ ์—†์• ์ค๋‹ˆ๋‹ค. ์ด๋Š” ๊ฐœ๋ฐœ์ž๋“ค์ด ํ•˜๋“œ์›จ์–ด์— ๋Œ€ํ•œ ๊นŠ์€ ์ง€์‹ ์—†์ด๋„ ๋ชจ๋ธ์„ ์ตœ์ ํ™”ํ•  ์ˆ˜ ์žˆ๊ฒŒ ํ•ด์ฃผ๋Š” ๋„๊ตฌ์ž…๋‹ˆ๋‹ค.

3. ์ตœ์ ํ™”๋œ ์ฝ”๋“œ(Optimized code specific to target HW):

  • TVM์€ ํŠน์ • ๋Œ€์ƒ ํ•˜๋“œ์›จ์–ด์— ๋งž์ถ”์–ด ์ตœ์ ํ™”๋œ ์ฝ”๋“œ๋ฅผ ์ƒ์„ฑํ•ฉ๋‹ˆ๋‹ค. ์ด ๊ณผ์ •์€ NVIDIA, Intel, AMD์™€ ๊ฐ™์€ ๋‹ค์–‘ํ•œ ํ•˜๋“œ์›จ์–ด ํ”Œ๋žซํผ์—์„œ ๋ชจ๋ธ์˜ ์„ฑ๋Šฅ์„ ๊ทน๋Œ€ํ™”ํ•ฉ๋‹ˆ๋‹ค.

4. ๋‹ค์–‘์„ฑ์— ๊ธฐ๋ฐ˜ํ•œ ๊ฐ•์ (Thrives on diversity: models, frameworks, types of optimizations, HW targets):

  • Apache TVM์€ ๋‹ค์–‘ํ•œ ๋ชจ๋ธ, ํ”„๋ ˆ์ž„์›Œํฌ, ์ตœ์ ํ™” ์œ ํ˜• ๋ฐ ํ•˜๋“œ์›จ์–ด ๋Œ€์ƒ์„ ์ง€์›ํ•˜๋Š” ๊ฐ•์ ์„ ๊ฐ€์ง€๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ์ด๋กœ ์ธํ•ด TVM์€ ๊ด‘๋ฒ”์œ„ํ•œ ํ™˜๊ฒฝ์—์„œ์˜ ๋จธ์‹ ๋Ÿฌ๋‹ ๋ชจ๋ธ ์ตœ์ ํ™”์— ๋งค์šฐ ์œ ์šฉํ•ฉ๋‹ˆ๋‹ค.

์ด ์Šฌ๋ผ์ด๋“œ๋Š” Apache TVM์ด ์–ด๋–ป๊ฒŒ ๋จธ์‹ ๋Ÿฌ๋‹ ๋ชจ๋ธ์˜ ์„ฑ๋Šฅ์„ ์ž๋™ํ™”ํ•˜๊ณ , ๋ณต์žกํ•œ ์ˆ˜๋™ ํŠœ๋‹ ์—†์ด๋„ ๋‹ค์–‘ํ•œ ํ•˜๋“œ์›จ์–ด์—์„œ ๋ชจ๋ธ์„ ์ตœ์ ํ™”ํ•  ์ˆ˜ ์žˆ๋Š” ๋ฐฉ๋ฒ•์„ ์ œ๊ณตํ•˜๋Š”์ง€๋ฅผ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ์ด๋Š” ๋จธ์‹ ๋Ÿฌ๋‹ ๋ชจ๋ธ์˜ ๋ฐฐํฌ์™€ ํ™•์žฅ์„ฑ์„ ํฌ๊ฒŒ ํ–ฅ์ƒ์‹œํ‚ฌ ์ˆ˜ ์žˆ๋Š” ์ค‘์š”ํ•œ ๋„๊ตฌ์ž„์„ ๊ฐ•์กฐํ•ฉ๋‹ˆ๋‹ค.

์ด ์Šฌ๋ผ์ด๋“œ๋Š” Apache TVM์„ ์‚ฌ์šฉํ•˜์—ฌ Apple์˜ M1, M1 Pro, ๋ฐ M1 Max ํ•˜๋“œ์›จ์–ด์—์„œ BERT ๋ชจ๋ธ์„ ์ตœ์ ํ™”ํ•˜๋Š” ์‚ฌ๋ก€ ์—ฐ๊ตฌ๋ฅผ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ์ฃผ์š” ๋‚ด์šฉ์€ ๋‹ค์Œ๊ณผ ๊ฐ™์Šต๋‹ˆ๋‹ค:

1. BERT ๋ชจ๋ธ(BERT Model):

  • BERT๋Š” ์ž์—ฐ์–ด ์ฒ˜๋ฆฌ๋ฅผ ์œ„ํ•œ ์ธ๊ธฐ ์žˆ๋Š” ๋จธ์‹ ๋Ÿฌ๋‹ ๋ชจ๋ธ์ž…๋‹ˆ๋‹ค. ์ด ์‚ฌ๋ก€ ์—ฐ๊ตฌ์—์„œ๋Š” BERT ๋ชจ๋ธ์„ ์‹œ์ž‘์ ์œผ๋กœ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค.

2. Apache TVM์˜ ์—ญํ• (Apache TVM):

  • TVM์€ ์ด ๋ชจ๋ธ์„ Apple์˜ ์‹ ํ˜• ํ•˜๋“œ์›จ์–ด์— ๋งž๊ฒŒ ์ตœ์ ํ™”ํ•ฉ๋‹ˆ๋‹ค. TVM์„ ํ†ตํ•ด ๊ฐœ๋ฐœ์ž๋Š” ์ˆ˜์ž‘์—… ํŠœ๋‹์„ ํ•˜์ง€ ์•Š๊ณ ๋„ ํ•˜๋“œ์›จ์–ด์— ํŠนํ™”๋œ ์„ฑ๋Šฅ ์ตœ์ ํ™”๋ฅผ ์ž๋™์œผ๋กœ ์ˆ˜ํ–‰ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.

3. ์ตœ์ ํ™”๋œ ์ฝ”๋“œ(Optimized code for Apple Hardware):

  • TVM์„ ์‚ฌ์šฉํ•˜์—ฌ ์ƒ์„ฑ๋œ ์ฝ”๋“œ๋Š” Apple์˜ M1, M1 Pro, ๋ฐ M1 Max ์นฉ์…‹์„ ์œ„ํ•ด ํŠน๋ณ„ํžˆ ์ตœ์ ํ™”๋˜์–ด, ์ด ํ”Œ๋žซํผ๋“ค์—์„œ BERT ๋ชจ๋ธ์˜ ์„ฑ๋Šฅ์„ ๊ทน๋Œ€ํ™”ํ•ฉ๋‹ˆ๋‹ค.

4. ๊ฒฐ๊ณผ ๋ฐ ์„ฑ๋Šฅ(Enabled M1 and M1 Pro/Max weeks after they were released. Demonstrates generality and automation):

  • TVM์€ Apple ํ•˜๋“œ์›จ์–ด๊ฐ€ ์ถœ์‹œ๋œ ์ง€ ๋ช‡ ์ฃผ ๋งŒ์— ์ด๋“ค์— ๋Œ€ํ•œ ์ง€์›์„ ํ™œ์„ฑํ™”ํ–ˆ์Šต๋‹ˆ๋‹ค. ์ด๋Š” TVM์˜ ๋ฒ”์šฉ์„ฑ๊ณผ ์ž๋™ํ™” ๋Šฅ๋ ฅ์„ ์ž…์ฆํ•˜๋Š” ์˜ˆ๋กœ, ์ƒˆ๋กœ์šด ํ•˜๋“œ์›จ์–ด์— ๋น ๋ฅด๊ฒŒ ์ ์‘ํ•˜๊ณ  ์ตœ์ ํ™”ํ•  ์ˆ˜ ์žˆ๋Š” ๋Šฅ๋ ฅ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค.

์ด ์‚ฌ๋ก€ ์—ฐ๊ตฌ๋Š” TVM์ด ์–ด๋–ป๊ฒŒ ๋‹ค์–‘ํ•œ ํ•˜๋“œ์›จ์–ด ํ”Œ๋žซํผ์— ๊ฑธ์ณ ๋จธ์‹ ๋Ÿฌ๋‹ ๋ชจ๋ธ์˜ ์„ฑ๋Šฅ ์ตœ์ ํ™”๋ฅผ ์ž๋™ํ™”ํ•  ์ˆ˜ ์žˆ๋Š”์ง€๋ฅผ ๋ณด์—ฌ์ฃผ๋Š” ํƒ์›”ํ•œ ์˜ˆ์ž…๋‹ˆ๋‹ค. ํŠนํžˆ, ์ƒˆ๋กœ์šด ํ•˜๋“œ์›จ์–ด ์ถœ์‹œ ํ›„ ์‹ ์†ํ•œ ์ง€์›๊ณผ ์„ฑ๋Šฅ ํ–ฅ์ƒ์„ ๊ฐ€๋Šฅํ•˜๊ฒŒ ํ•จ์œผ๋กœ์จ, TVM์˜ ์œ ์šฉ์„ฑ๊ณผ ํšจ์œจ์„ฑ์„ ๊ฐ•์กฐํ•ฉ๋‹ˆ๋‹ค.

์ด ์Šฌ๋ผ์ด๋“œ๋Š” Apple M1 Max ์นฉ์„ ์‚ฌ์šฉํ•˜๋Š” ํ™˜๊ฒฝ์—์„œ Apache TVM์„ ํ†ตํ•ด TensorFlow๋ฅผ ์ตœ์ ํ™”ํ•˜๋Š” ์„ฑ๋Šฅ ๊ฒฐ๊ณผ๋ฅผ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ๊ตฌ์ฒด์ ์ธ ๋‚ด์šฉ์€ ๋‹ค์Œ๊ณผ ๊ฐ™์Šต๋‹ˆ๋‹ค:

1. CPU ์„ฑ๋Šฅ ํ–ฅ์ƒ(Appleโ€™s Most Powerful Chips with ~3X CPU Performance Speedups with Apache TVM):

  • TensorFlow on MBP16 w/ M1 Max: Apple M1 Max๋ฅผ ์‚ฌ์šฉํ•˜๋Š” MacBook Pro 16์—์„œ TensorFlow๋ฅผ ์‹คํ–‰ํ•  ๋•Œ์˜ CPU ์ง€์—ฐ ์‹œ๊ฐ„์€ 113.63ms์ž…๋‹ˆ๋‹ค.
  • TVM on MBP16 w/ M1 Max: ๊ฐ™์€ ์„ค์ •์—์„œ Apache TVM์„ ์‚ฌ์šฉํ•˜์—ฌ ์ตœ์ ํ™”ํ•œ ๊ฒฝ์šฐ, CPU ์ง€์—ฐ ์‹œ๊ฐ„์ด 36.36ms๋กœ ์ค„์–ด๋“ค์–ด ์•ฝ 3๋ฐฐ ๋นจ๋ผ์กŒ์Šต๋‹ˆ๋‹ค.

2. GPU ์„ฑ๋Šฅ ํ–ฅ์ƒ(Appleโ€™s Most Powerful Chips with ~2X GPU Performance Speedups with Apache TVM):

  • TensorFlow on MBP16 w/ M1 Max (FP16): Apple M1 Max์˜ GPU๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ TensorFlow๋ฅผ ์‹คํ–‰ํ•  ๋•Œ์˜ GPU ์ง€์—ฐ ์‹œ๊ฐ„์€ 19.68ms์ž…๋‹ˆ๋‹ค.
  • TVM on MBP16 w/ M1 Max (FP16): Apache TVM์„ ์‚ฌ์šฉํ•˜์—ฌ ์ตœ์ ํ™”ํ•œ ๊ฒฝ์šฐ, GPU ์ง€์—ฐ ์‹œ๊ฐ„์ด 10.88ms๋กœ ์ค„์–ด๋“ค์–ด ๊ฑฐ์˜ 2๋ฐฐ์˜ ์„ฑ๋Šฅ ํ–ฅ์ƒ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค.

3. ๊ฒฐ๋ก ๊ณผ ์ „๋ง(Still, need to widen to more users, need further automation, and to squeeze more performance):

  • ์ด ๋ฐ์ดํ„ฐ๋Š” Apache TVM์ด Apple์˜ ์ตœ์‹  ์นฉ์—์„œ ๋งค์šฐ ํšจ๊ณผ์ ์œผ๋กœ ์„ฑ๋Šฅ์„ ํ–ฅ์ƒ์‹œํ‚ฌ ์ˆ˜ ์žˆ์Œ์„ ๋ณด์—ฌ์ฃผ์ง€๋งŒ, ์ด ๊ธฐ์ˆ ์„ ๋” ๋งŽ์€ ์‚ฌ์šฉ์ž์—๊ฒŒ ํ™•์‚ฐ์‹œํ‚ค๊ณ , ์ถ”๊ฐ€์ ์ธ ์ž๋™ํ™”๋ฅผ ํ†ตํ•ด ์„ฑ๋Šฅ์„ ๋”์šฑ ๋†’์—ฌ์•ผ ํ•œ๋‹ค๋Š” ํ•„์š”์„ฑ์„ ๊ฐ•์กฐํ•ฉ๋‹ˆ๋‹ค.

์ด ์Šฌ๋ผ์ด๋“œ๋Š” Apache TVM์ด ์ตœ์‹  ํ•˜๋“œ์›จ์–ด์—์„œ ์–ด๋–ป๊ฒŒ ์‹ค์งˆ์ ์ธ ์„ฑ๋Šฅ ๊ฐœ์„ ์„ ์ œ๊ณตํ•  ์ˆ˜ ์žˆ๋Š”์ง€๋ฅผ ์‹ค์ œ ์ˆ˜์น˜๋กœ ๋ณด์—ฌ์ฃผ๋ฉฐ, ์ด ๊ธฐ์ˆ ์˜ ์ž ์žฌ๋ ฅ๊ณผ ๊ฐœ์„ ์˜ ํ•„์š”์„ฑ์„ ๊ฐ•์กฐํ•ฉ๋‹ˆ๋‹ค.

์ด ์Šฌ๋ผ์ด๋“œ๋Š” "์„ ํƒ๊ถŒ์ด ๊ธฐ์—… ๊ณ ๊ฐ์—๊ฒŒ ์„ ํƒ ๊ฐ€๋Šฅ์„ฑ์„ ์ œ๊ณตํ•œ๋‹ค(Choice Provides Optionality for Enterprise Customers)"๋Š” ๋ฉ”์‹œ์ง€๋ฅผ ์ „๋‹ฌํ•ฉ๋‹ˆ๋‹ค. ์—ฌ๊ธฐ์„œ ๊ฐ•์กฐํ•˜๋Š” ํ•ต์‹ฌ ๋‚ด์šฉ์€ ๋‹ค์Œ๊ณผ ๊ฐ™์Šต๋‹ˆ๋‹ค:

์„ ํƒ๊ถŒ(Choice):

  • ๊ธฐ์—… ๊ณ ๊ฐ๋“ค์—๊ฒŒ ๋‹ค์–‘ํ•œ ๊ธฐ์ˆ ์  ์„ ํƒ์„ ์ œ๊ณตํ•จ์œผ๋กœ์จ, ๊ทธ๋“ค์˜ ํŠน์ • ๋น„์ฆˆ๋‹ˆ์Šค ์š”๊ตฌ์™€ ์š”๊ฑด์— ๊ฐ€์žฅ ์ž˜ ๋งž๋Š” ์†”๋ฃจ์…˜์„ ์„ ํƒํ•  ์ˆ˜ ์žˆ๋Š” ์œ ์—ฐ์„ฑ์„ ๋ถ€์—ฌํ•ฉ๋‹ˆ๋‹ค.

์„ ํƒ ๊ฐ€๋Šฅ์„ฑ(Optionality):

  • ๊ธฐ์—…๋“ค์ด ๋‹ค์–‘ํ•œ ์ƒํ™ฉ์— ๋”ฐ๋ผ ์ตœ์ ์˜ ๊ธฐ์ˆ ์  ๊ฒฝ๋กœ๋ฅผ ์„ ํƒํ•  ์ˆ˜ ์žˆ๋„๋ก ํ•˜์—ฌ, ์ „์ฒด์ ์ธ ๋น„์ฆˆ๋‹ˆ์Šค ์ „๋žต์— ๋งž๊ฒŒ ์กฐ์ •ํ•  ์ˆ˜ ์žˆ๋Š” ์˜ต์…˜์„ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค.

์ด ์Šฌ๋ผ์ด๋“œ๋Š” ๊ธฐ์—… ๊ณ ๊ฐ๋“ค์ด ์ž์‹ ๋“ค์˜ ํ•„์š”์— ๋งž์ถฐ ๊ธฐ์ˆ  ์†”๋ฃจ์…˜์„ ๋งž์ถคํ™”ํ•  ์ˆ˜ ์žˆ๋Š” ์ค‘์š”์„ฑ์„ ๊ฐ•์กฐํ•˜๊ณ , ์ด๋Ÿฌํ•œ ์„ ํƒ๊ถŒ์ด ๊ธฐ์—…์˜ ์ „๋žต์  ์œ ์—ฐ์„ฑ์„ ํ–ฅ์ƒ์‹œํ‚ค๊ณ  ์‹œ์žฅ์—์„œ์˜ ๊ฒฝ์Ÿ๋ ฅ์„ ๊ฐ•ํ™”ํ•  ์ˆ˜ ์žˆ์Œ์„ ์‹œ์‚ฌํ•ฉ๋‹ˆ๋‹ค.

์ด ์Šฌ๋ผ์ด๋“œ๋Š” ์ธ๊ณต์ง€๋Šฅ(AI)๊ณผ ๋จธ์‹ ๋Ÿฌ๋‹(ML)์—์„œ์˜ ์„ฑ๋Šฅ ์ด์‹์„ฑ(performance portability)์— ๋Œ€ํ•œ ์„ ํƒ์˜ ์ค‘์š”์„ฑ์„ ์„ค๋ช…ํ•˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ์ด๋ฅผ ํ†ตํ•ด ํ•˜๋“œ์›จ์–ด์™€ ๋ชจ๋ธ ๊ฐ„์˜ ๊ฒฐํ•ฉ์„ ๋œ์–ด๋‚ด๊ณ , ๋” ๋„“์€ ์‚ฌ์šฉ๊ณผ ํ˜์‹ ์„ ๊ฐ€๋Šฅํ•˜๊ฒŒ ํ•˜๋Š” ๋ฐฉ๋ฒ•์„ ์ œ์‹œํ•ฉ๋‹ˆ๋‹ค. ์ฃผ์š” ๋‚ด์šฉ์€ ๋‹ค์Œ๊ณผ ๊ฐ™์Šต๋‹ˆ๋‹ค:

1. ํ•˜๋“œ์›จ์–ด ์„ ํƒ(Choice of HW for Your Model):

  • ์‚ฌ์šฉ์ž๋Š” ์ž์‹ ์˜ ML ๋ชจ๋ธ์— ๊ฐ€์žฅ ์ ํ•ฉํ•œ ๋‹ค์–‘ํ•œ ํ•˜๋“œ์›จ์–ด ์˜ต์…˜ ์ค‘์—์„œ ์„ ํƒํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด ์Šฌ๋ผ์ด๋“œ์—์„œ๋Š” USB ๋“œ๋ผ์ด๋ธŒ, Intel ํ”„๋กœ์„ธ์„œ, ๋ชจ๋ฐ”์ผ ๊ธฐ๊ธฐ, ๊ทธ๋ฆฌ๊ณ  ๋‹ค์–‘ํ•œ ์นฉ์…‹๊ณผ ๋ณด๋“œ ๋“ฑ ๋‹ค์–‘ํ•œ ํ•˜๋“œ์›จ์–ด๊ฐ€ ๊ทธ ์˜ˆ๋กœ ๋‚˜ํƒ€๋‚˜ ์žˆ์Šต๋‹ˆ๋‹ค.

2. ๋ชจ๋ธ ์„ ํƒ(Choice of Models for Your HW):

  • ๋ฐ˜๋Œ€๋กœ, ํŠน์ • ํ•˜๋“œ์›จ์–ด๋ฅผ ๊ธฐ์ค€์œผ๋กœ ๊ฐ€์žฅ ์ ํ•ฉํ•œ ML ๋ชจ๋ธ์„ ์„ ํƒํ•  ์ˆ˜ ์žˆ๋Š” ์˜ต์…˜๋„ ์ œ๊ณต๋ฉ๋‹ˆ๋‹ค. ์ด๋Š” ํ•˜๋“œ์›จ์–ด์˜ ์„ฑ๋Šฅ์„ ์ตœ๋Œ€ํ™”ํ•˜๊ณ , ํŠน์ • ์• ํ”Œ๋ฆฌ์ผ€์ด์…˜์— ๊ฐ€์žฅ ์ ํ•ฉํ•œ ๋ชจ๋ธ์„ ์‚ฌ์šฉํ•  ์ˆ˜ ์žˆ๊ฒŒ ํ•ฉ๋‹ˆ๋‹ค.

3. ML๊ณผ ํ•˜๋“œ์›จ์–ด์˜ ํƒˆ๊ฒฐํ•ฉ(Breaking ML from the HW box):

  • โ€œML์„ ํ•˜๋“œ์›จ์–ด ๋ฐ•์Šค์—์„œ ํ•ด๋ฐฉ์‹œํ‚จ๋‹คโ€๋Š” ์„ค๋ช…์€, ML ์†Œํ”„ํŠธ์›จ์–ด์™€ ํ•˜๋“œ์›จ์–ด ๊ฐ„์˜ ์œ ์—ฐํ•œ ๊ฒฐํ•ฉ์„ ํ†ตํ•ด ๋” ๋งŽ์€ ํ˜์‹ ๊ณผ ์‚ฐ์—… ์„ฑ์žฅ์„ ๊ฐ€๋Šฅํ•˜๊ฒŒ ํ•  ๊ฒƒ์ด๋ผ๋Š” ์ ์„ ๊ฐ•์กฐํ•ฉ๋‹ˆ๋‹ค. ์ด๋Š” 1980๋…„๋Œ€ ์šด์˜ ์ฒด์ œ์˜ ๋ฐœ์ „๊ณผ ์œ ์‚ฌํ•œ ๋ฐฉ์‹์œผ๋กœ, ๊ธฐ์ˆ ์˜ ๋ฐœ์ „๊ณผ ๋ณด๊ธ‰์„ ๊ฐ€์†ํ™”ํ•  ์ˆ˜ ์žˆ๋‹ค๊ณ  ์„ค๋ช…ํ•˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค.

์ด ์Šฌ๋ผ์ด๋“œ๋Š” ๊ธฐ์—… ๊ณ ๊ฐ๋“ค์ด ๋‹ค์–‘ํ•œ ํ•˜๋“œ์›จ์–ด์™€ ๋ชจ๋ธ ์˜ต์…˜์„ ์„ ํƒํ•จ์œผ๋กœ์จ ์ตœ์ ์˜ ์„ฑ๋Šฅ๊ณผ ํšจ์œจ์„ฑ์„ ๋‹ฌ์„ฑํ•  ์ˆ˜ ์žˆ๋Š” ๋” ๋งŽ์€ ๊ธฐํšŒ๋ฅผ ๊ฐ–๊ฒŒ ๋จ์„ ๊ฐ•์กฐํ•˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ์ด๋Š” ์ „์ฒด ์‚ฐ์—…์— ๊ฑธ์ณ ๋” ๋งŽ์€ ๋งž์ถคํ™”์™€ ํ˜์‹ ์„ ์ด‰์ง„ํ•  ๊ฒƒ์ž„์„ ์‹œ์‚ฌํ•ฉ๋‹ˆ๋‹ค.

์ด ์Šฌ๋ผ์ด๋“œ๋Š” AI ๋ฐ ML์˜ ์„ ํƒ ์˜ต์…˜์„ ๊ตฌ์ถ•ํ•˜๋Š”๋ฐ ์žˆ์–ด์„œ ํ˜‘๋ ฅ์˜ ์ค‘์š”์„ฑ์„ ๊ฐ•์กฐํ•˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. "ํ˜ผ์ž์„œ๋Š” ์„ ํƒ์„ ๊ตฌ์ถ•ํ•  ์ˆ˜ ์—†๋‹ค(Can't Build Choice Alone...)"๋Š” ๋ฉ”์‹œ์ง€๋Š” ๋‹ค์Œ๊ณผ ๊ฐ™์€ ์ฃผ์š” ๋‚ด์šฉ์„ ๋‹ด๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค:

1. TVM ๊ณตํ—Œ์ž(TVM Contributors):

  • TVM ํ”„๋กœ์ ํŠธ์˜ ๊ธฐ์—ฌ์ž ์ปค๋ฎค๋‹ˆํ‹ฐ๋ฅผ ๋‚˜ํƒ€๋‚ด๋Š” ์ด๋ฏธ์ง€๊ฐ€ ํฌํ•จ๋˜์–ด ์žˆ์Šต๋‹ˆ๋‹ค. ์ด ์ด๋ฏธ์ง€๋Š” ๋‹ค์–‘ํ•œ ๋ฐฐ๊ฒฝ๊ณผ ์ „๋ฌธ์„ฑ์„ ๊ฐ€์ง„ ๊ฐœ์ธ๋“ค๋กœ ๊ตฌ์„ฑ๋œ ์ปค๋ฎค๋‹ˆํ‹ฐ์˜ ๋‹ค์–‘์„ฑ์„ ๋ณด์—ฌ์ฃผ๋ฉฐ, ์ด๋“ค์ด ๊ณต๋™์œผ๋กœ ๊ธฐ์ˆ  ๊ฐœ๋ฐœ์— ์ฐธ์—ฌํ•˜๊ณ  ์žˆ์Œ์„ ์‹œ์‚ฌํ•ฉ๋‹ˆ๋‹ค.

2. ์ฃผ์š” ๊ธฐ์ˆ  ํŒŒํŠธ๋„ˆ(Major Tech Partners):

  • Microsoft Azure, Amazon AWS, Qualcomm ๋“ฑ ์ฃผ์š” ๊ธฐ์ˆ  ํšŒ์‚ฌ๋“ค์˜ ๋กœ๊ณ ๊ฐ€ ๋‚˜ํƒ€๋‚˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด๋Š” OctoML์ด ์ด๋Ÿฌํ•œ ๋Œ€ํ˜• ํ”Œ๋žซํผ๊ณผ ํ˜‘๋ ฅํ•˜์—ฌ ๊ทธ๋“ค์˜ ํ•˜๋“œ์›จ์–ด์™€ ์„œ๋น„์Šค์—์„œ AI/ML ๋ชจ๋ธ์˜ ์ตœ์ ํ™”์™€ ๋ฐฐํฌ๋ฅผ ์ง€์›ํ•˜๊ณ  ์žˆ์Œ์„ ๋‚˜ํƒ€๋ƒ…๋‹ˆ๋‹ค.

์ด ์Šฌ๋ผ์ด๋“œ์˜ ํ•ต์‹ฌ ๋ฉ”์‹œ์ง€๋Š”, ์„ ํƒ๊ณผ ์ตœ์ ํ™”๋ฅผ ์œ„ํ•œ ๋‹ค์–‘ํ•œ ์†”๋ฃจ์…˜์„ ๊ฐœ๋ฐœํ•˜๊ณ  ๊ตฌํ˜„ํ•˜๋Š” ๊ฒƒ์€ ํ˜‘๋ ฅ์„ ํ†ตํ•ด์„œ๋งŒ ๊ฐ€๋Šฅํ•˜๋‹ค๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค. ๋Œ€๊ทœ๋ชจ ํ…Œํฌ ๊ธฐ์—…๋“ค๊ณผ์˜ ํŒŒํŠธ๋„ˆ์‹ญ์€ ๊ธฐ์ˆ  ํ˜์‹ ์„ ์ด‰์ง„ํ•˜๊ณ , ์‚ฌ์šฉ์ž์—๊ฒŒ ๋” ๋งŽ์€ ์„ ํƒ๊ถŒ๊ณผ ์œ ์—ฐ์„ฑ์„ ์ œ๊ณตํ•˜๋Š” ๋ฐ ํ•„์ˆ˜์ ์ž…๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ํ˜‘๋ ฅ์€ AI์™€ ML ๋ถ„์•ผ์—์„œ ๋” ๋„“์€ ์ ์šฉ ๊ฐ€๋Šฅ์„ฑ๊ณผ ์„ฑ์žฅ์„ ๊ฐ€๋Šฅํ•˜๊ฒŒ ํ•ฉ๋‹ˆ๋‹ค.


์ด ์Šฌ๋ผ์ด๋“œ๋Š” ์„ฑ๋Šฅ, ์ž๋™ํ™”, ์„ ํƒ์ด๋ผ๋Š” ์„ธ ๊ฐ€์ง€ ์ฃผ์š” ์š”์†Œ๊ฐ€ AI/ML์˜ ์ง€์† ๊ฐ€๋Šฅ์„ฑ๊ณผ ํ˜์‹ ์— ์–ด๋–ป๊ฒŒ ๊ธฐ์—ฌํ•˜๋Š”์ง€ ์„ค๋ช…ํ•˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ๊ฐ ์š”์†Œ๋“ค์ด ๊ฒฐํ•ฉ๋  ๋•Œ ๋ฐœ์ƒํ•˜๋Š” ํšจ๊ณผ๋ฅผ ์ž์„ธํžˆ ์‚ดํŽด๋ณด๊ฒ ์Šต๋‹ˆ๋‹ค:

1. ์„ฑ๋Šฅ๊ณผ ์ž๋™ํ™”์˜ ๊ฒฐํ•ฉ (Performance + Automation = Sustainability)

  • ์„ฑ๋Šฅ (Performance): ์ด ์š”์†Œ๋Š” AI/ML ์‹œ์Šคํ…œ์ด ์ตœ๋Œ€ํ•œ ํšจ์œจ์ ์œผ๋กœ ์ž‘๋™ํ•˜๋„๋ก ํ•˜๋Š”๋ฐ ์ค‘์ ์„ ๋‘ก๋‹ˆ๋‹ค. ๋” ๋น ๋ฅด๊ณ  ์ •ํ™•ํ•œ ์ฒ˜๋ฆฌ๋ฅผ ๊ฐ€๋Šฅํ•˜๊ฒŒ ํ•˜์—ฌ, ๊ธฐ์ˆ ์  ํ•œ๊ณ„๋ฅผ ๋„˜์–ด์„œ๋Š” ๊ฒฐ๊ณผ๋ฅผ ๋‹ฌ์„ฑํ•˜๋ ค๊ณ  ํ•ฉ๋‹ˆ๋‹ค.
  • ์ž๋™ํ™” (Automation): ๋ฐ˜๋ณต์ ์ด๊ฑฐ๋‚˜ ๋ณต์žกํ•œ ํ”„๋กœ์„ธ์Šค๋ฅผ ์ž๋™ํ™”ํ•จ์œผ๋กœ์จ, ์ธ๊ฐ„์˜ ๊ฐœ์ž…์„ ์ตœ์†Œํ™”ํ•˜๊ณ  ์˜ค๋ฅ˜๋ฅผ ์ค„์ด๋ฉฐ, ์†๋„์™€ ํšจ์œจ์„ฑ์„ ํ–ฅ์ƒ์‹œํ‚ต๋‹ˆ๋‹ค.
  • ์ง€์† ๊ฐ€๋Šฅ์„ฑ (Sustainability): ์„ฑ๋Šฅ๊ณผ ์ž๋™ํ™”์˜ ๊ฒฐํ•ฉ์€ ์—๋„ˆ์ง€ ์†Œ๋น„๋ฅผ ์ตœ์ ํ™”ํ•˜๊ณ , ์ž์› ์‚ฌ์šฉ์„ ํšจ์œจํ™”ํ•˜์—ฌ ์ „๋ฐ˜์ ์ธ ํ™˜๊ฒฝ ์˜ํ–ฅ์„ ๊ฐ์†Œ์‹œํ‚ต๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, ๋” ๋น ๋ฅธ ์•Œ๊ณ ๋ฆฌ์ฆ˜์€ ์ ์€ ์—๋„ˆ์ง€๋กœ ๋” ๋งŽ์€ ์ž‘์—…์„ ์ˆ˜ํ–‰ํ•  ์ˆ˜ ์žˆ๊ณ , ์ž๋™ํ™”๋Š” ํ•„์š”ํ•œ ์—ฐ์‚ฐ ์ž์›์˜ ๋‚ญ๋น„๋ฅผ ์ค„์—ฌ์ค๋‹ˆ๋‹ค.

2. ์ง€์† ๊ฐ€๋Šฅ์„ฑ๊ณผ ์„ ํƒ์˜ ๊ฒฐํ•ฉ (Sustainability + Choice = More innovation, more ways to use AI/ML thoughtfully)

  • ์ง€์† ๊ฐ€๋Šฅ์„ฑ (Sustainability): ์ด๋Š” ์ž์›์„ ์ฑ…์ž„๊ฐ ์žˆ๊ฒŒ ์‚ฌ์šฉํ•˜๊ณ , ์žฅ๊ธฐ์ ์œผ๋กœ ํ™˜๊ฒฝ์ , ์‚ฌํšŒ์ , ๊ฒฝ์ œ์  ์˜ํ–ฅ์„ ๊ณ ๋ คํ•˜๋Š” ๊ฒƒ์„ ํฌํ•จํ•ฉ๋‹ˆ๋‹ค.
  • ์„ ํƒ (Choice): ์‚ฌ์šฉ์ž๊ฐ€ ๋‹ค์–‘ํ•œ ๊ธฐ์ˆ ์  ์˜ต์…˜ ์ค‘์—์„œ ์„ ํƒํ•  ์ˆ˜ ์žˆ๋„๋ก ํ•จ์œผ๋กœ์จ, ๊ทธ๋“ค์˜ ํŠน์ • ์š”๊ตฌ์— ๊ฐ€์žฅ ์ ํ•ฉํ•œ ์†”๋ฃจ์…˜์„ ์„ ํƒํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด๋Š” ํ•˜๋“œ์›จ์–ด๋‚˜ ์•Œ๊ณ ๋ฆฌ์ฆ˜ ์„ ํƒ์— ์žˆ์–ด์„œ์˜ ์œ ์—ฐ์„ฑ์„ ์˜๋ฏธํ•˜๋ฉฐ, ๊ฐ ์‚ฌ์šฉ์ž์˜ ๋…ํŠนํ•œ ํ•„์š”๋ฅผ ์ถฉ์กฑ์‹œํ‚ฌ ์ˆ˜ ์žˆ๋Š” ๋‹ค์–‘ํ•œ ๊ฐ€๋Šฅ์„ฑ์„ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค.
  • ํ˜์‹  ๋ฐ ์‹ ์ค‘ํ•œ AI/ML ์‚ฌ์šฉ (More innovation, more ways to use AI/ML thoughtfully): ์ง€์† ๊ฐ€๋Šฅํ•˜๊ณ  ๋‹ค์–‘ํ•œ ์„ ํƒ์ง€๋ฅผ ์ œ๊ณตํ•จ์œผ๋กœ์จ, ๊ธฐ์—…๊ณผ ๊ฐœ๋ฐœ์ž๋Š” AI์™€ ML์„ ๋” ์‹ ์ค‘ํ•˜๊ณ  ํ˜์‹ ์ ์ธ ๋ฐฉ์‹์œผ๋กœ ํ™œ์šฉํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด๋Š” ์ƒˆ๋กœ์šด ์‚ฌ์šฉ ์‚ฌ๋ก€ ๋ฐœ๊ฒฌ, ๋น„์šฉ ํšจ์œจ์„ฑ ์ฆ๊ฐ€, ๊ทธ๋ฆฌ๊ณ  ์‚ฌํšŒ์ , ํ™˜๊ฒฝ์  ์ฑ…์ž„์˜ ์‹คํ˜„์œผ๋กœ ์ด์–ด์ง‘๋‹ˆ๋‹ค.

์ด ์Šฌ๋ผ์ด๋“œ์—์„œ๋Š” ์ด๋Ÿฌํ•œ ์š”์†Œ๋“ค์ด ์–ด๋–ป๊ฒŒ ์ƒํ˜ธ ์ž‘์šฉํ•˜์—ฌ AI์™€ ML์˜ ์žฅ๊ธฐ์ ์ธ ์„ฑ๊ณต๊ณผ ๋ฐœ์ „์— ๊ธฐ์—ฌํ•˜๋Š”์ง€๋ฅผ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ์„ฑ๋Šฅ๊ณผ ์ž๋™ํ™”๋Š” ํšจ์œจ์„ฑ๊ณผ ๋น ๋ฅธ ๊ฒฐ๊ณผ๋ฅผ ์ œ๊ณตํ•˜๋Š” ๋™์‹œ์—, ์„ ํƒ์€ ์‚ฌ์šฉ์ž์—๊ฒŒ ๋งž์ถคํ˜• ์†”๋ฃจ์…˜์„ ์ œ๊ณตํ•˜์—ฌ ๋” ๋„“์€ ์ ์šฉ ๋ฒ”์œ„์™€ ํ˜์‹ ์„ ๊ฐ€๋Šฅํ•˜๊ฒŒ ํ•ฉ๋‹ˆ๋‹ค. ์ด๋Ÿฐ ์ ‘๊ทผ ๋ฐฉ์‹์€ ๊ธฐ์ˆ ์ด ์ง€์† ๊ฐ€๋Šฅํ•˜๊ณ  ์ฑ…์ž„๊ฐ ์žˆ๊ฒŒ ๋ฐœ์ „ํ•  ์ˆ˜ ์žˆ๋„๋ก ๋•์Šต๋‹ˆ๋‹ค.