Sign In

Sustainability Is Not Linear: Quantifying Performance, Energy, and Privacy Trade-offs in On-Device Intelligence

์ž‘์„ฑ์ž
  • Haebom
์นดํ…Œ๊ณ ๋ฆฌ
Empty

์ €์ž

Eziyo Ehsani, Luca Giamattei, Ivano Malavolta, Roberto Pietrantuono

๐Ÿ’ก ๊ฐœ์š”

๋ณธ ์—ฐ๊ตฌ๋Š” LLM์„ ํด๋ผ์šฐ๋“œ์—์„œ ์—ฃ์ง€ ๋””๋ฐ”์ด์Šค๋กœ ์˜ฎ๊ธธ ๋•Œ ๋ฐœ์ƒํ•˜๋Š” ์„ฑ๋Šฅ, ์—๋„ˆ์ง€, ๊ฐœ์ธ์ •๋ณด ๋ณดํ˜ธ ๊ฐ„์˜ ๋ณต์žกํ•œ ์ƒ์ถฉ ๊ด€๊ณ„๋ฅผ ํƒ๊ตฌํ•ฉ๋‹ˆ๋‹ค. ์žฌํ˜„ ๊ฐ€๋Šฅํ•œ ์‹คํ—˜ ํŒŒ์ดํ”„๋ผ์ธ์„ ๊ตฌ์ถ•ํ•˜์—ฌ ์‹ค์ œ ์‚ฌ์šฉ์ž ํ™˜๊ฒฝ์„ ๋ฐ˜์˜ํ•œ ์•ˆ๋“œ๋กœ์ด๋“œ ๊ธฐ๊ธฐ์—์„œ ๋‹ค์–‘ํ•œ LLM์˜ ์„ฑ๋Šฅ์„ ์ธก์ •ํ•˜๊ณ , ๋ฉ”๋ชจ๋ฆฌ ์ œ์•ฝ, ์—๋„ˆ์ง€ ์†Œ๋น„, ์‘๋‹ต ํ’ˆ์งˆ ๊ฐ„์˜ ๊ท ํ˜•์ ์„ ๋ถ„์„ํ–ˆ์Šต๋‹ˆ๋‹ค.

๐Ÿ”‘ ์‹œ์‚ฌ์  ๋ฐ ํ•œ๊ณ„

โ€ข
์—ฃ์ง€ ๋””๋ฐ”์ด์Šค์—์„œ LLM์˜ ํšจ์œจ์ ์ธ ํ™œ์šฉ์„ ์œ„ํ•ด์„œ๋Š” ๋ชจ๋ธ ์•„ํ‚คํ…์ฒ˜ ์ž์ฒด๊ฐ€ ์—๋„ˆ์ง€ ์†Œ๋น„์— ๋” ํฐ ์˜ํ–ฅ์„ ๋ฏธ์น˜๋ฉฐ, ์–‘์žํ™” ๊ธฐ๋ฒ•์€ ์—๋„ˆ์ง€ ์ ˆ๊ฐ ํšจ๊ณผ๊ฐ€ ๋ฏธ๋ฏธํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
โ€ข
Mixture-of-Experts(MoE) ์•„ํ‚คํ…์ฒ˜๋Š” ํฌ๊ธฐ์— ๋น„ํ•ด ๋‚ฎ์€ ์—๋„ˆ์ง€ ์†Œ๋น„๋ฅผ ๋ณด์—ฌ์ฃผ์–ด ์—ฃ์ง€ ํ™˜๊ฒฝ์— ์œ ๋ฆฌํ•œ ๋Œ€์•ˆ์ด ๋  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
โ€ข
์‘๋‹ต ํ’ˆ์งˆ๊ณผ ์ง€์† ๊ฐ€๋Šฅํ•œ ์—๋„ˆ์ง€ ์†Œ๋น„ ์‚ฌ์ด์˜ ์‹ค์šฉ์ ์ธ ๊ท ํ˜•์ ์€ Qwen2.5-3B์™€ ๊ฐ™์€ ์ค‘์†Œํ˜• ๋ชจ๋ธ์—์„œ ๋ฐœ๊ฒฌ๋˜์—ˆ์œผ๋ฉฐ, ์ด๋Š” ์—ฃ์ง€ LLM ๋ชจ๋ธ ์„ ํƒ์— ๋Œ€ํ•œ ์ค‘์š”ํ•œ ์ง€์นจ์„ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค.
๐Ÿ‘
Made with Slashpage