%22%2F%3E%0A%20%3Cg%20opacity%3D%220.12%22%3E%3Crect%20x%3D%22120%22%20y%3D%22126%22%20width%3D%22420%22%20height%3D%2212%22%20fill%3D%22url(%23g)%22%2F%3E%3Crect%20x%3D%22660%22%20y%3D%22472.5%22%20width%3D%22360%22%20height%3D%2212%22%20fill%3D%22url(%23g)%22%2F%3E%3C%2Fg%3E%0A%20%3Cg%3E%3Cpath%20d%3D%22M%20189%20351%20L%201061%20284%22%20stroke%3D%22url(%23g)%22%20stroke-opacity%3D%220.25%22%20stroke-width%3D%223%22%2F%3E%3Cpath%20d%3D%22M%201061%20284%20L%20661%20316%22%20stroke%3D%22url(%23g)%22%20stroke-opacity%3D%220.25%22%20stroke-width%3D%223%22%2F%3E%3Cpath%20d%3D%22M%20661%20316%20L%20304%20268%22%20stroke%3D%22url(%23g)%22%20stroke-opacity%3D%220.25%22%20stroke-width%3D%223%22%2F%3E%3Cpath%20d%3D%22M%20304%20268%20L%20640%20116%22%20stroke%3D%22url(%23g)%22%20stroke-opacity%3D%220.25%22%20stroke-width%3D%223%22%2F%3E%3Cpath%20d%3D%22M%20640%20116%20L%20313%20117%22%20stroke%3D%22url(%23g)%22%20stroke-opacity%3D%220.25%22%20stroke-width%3D%223%22%2F%3E%3Ccircle%20cx%3D%22189%22%20cy%3D%22351%22%20r%3D%2211%22%20fill%3D%22url(%23g)%22%20opacity%3D%220.6%22%2F%3E%3Ccircle%20cx%3D%221061%22%20cy%3D%22284%22%20r%3D%2215%22%20fill%3D%22url(%23g)%22%20opacity%3D%220.6%22%2F%3E%3Ccircle%20cx%3D%22661%22%20cy%3D%22316%22%20r%3D%2210%22%20fill%3D%22url(%23g)%22%20opacity%3D%220.6%22%2F%3E%3Ccircle%20cx%3D%22304%22%20cy%3D%22268%22%20r%3D%2213%22%20fill%3D%22url(%23g)%22%20opacity%3D%220.6%22%2F%3E%3Ccircle%20cx%3D%22640%22%20cy%3D%22116%22%20r%3D%2215%22%20fill%3D%22url(%23g)%22%20opacity%3D%220.6%22%2F%3E%3Ccircle%20cx%3D%22313%22%20cy%3D%22117%22%20r%3D%2213%22%20fill%3D%22url(%23g)%22%20opacity%3D%220.6%22%2F%3E%3C%2Fg%3E%0A%20%3Cg%3E%0A%20%3Ctext%20x%3D%2250%25%22%20y%3D%2250%25%22%20dominant-baseline%3D%22central%22%20text-anchor%3D%22middle%22%20font-family%3D%22Inter%2Csystem-ui%2CArial%22%20font-size%3D%22108%22%20font-weight%3D%221000%22%20fill%3D%22none%22%20stroke%3D%22%230b1020%22%20stroke-width%3D%2210%22%20stroke-opacity%3D%220.6%22%20style%3D%22paint-order%3A%20stroke%20fill%3B%22%3E%23TURBOQUANT%3C%2Ftext%3E%0A%20%3Ctext%20x%3D%2250%25%22%20y%3D%2250%25%22%20dominant-baseline%3D%22central%22%20text-anchor%3D%22middle%22%20font-family%3D%22Inter%2Csystem-ui%2CArial%22%20font-size%3D%22108%22%20font-weight%3D%221000%22%20fill%3D%22url(%23g)%22%20filter%3D%22url(%23drop)%22%3E%23TURBOQUANT%3C%2Ftext%3E%0A%20%3C%2Fg%3E%0A%20%3Cg%20transform%3D%22rotate(-8%20264%20126)%22%3E%3Ctext%20x%3D%22216%22%20y%3D%22126%22%20font-family%3D%22Inter%2Csystem-ui%2CArial%22%20font-size%3D%2234%22%20font-weight%3D%22700%22%20fill%3D%22url(%23g)%22%20fill-opacity%3D%220.9%22%3EGoogle%3C%2Ftext%3E%3C%2Fg%3E%0A%20%3Cg%20transform%3D%22rotate(10%20936%20516.6)%22%3E%3Ctext%20x%3D%22864%22%20y%3D%22516.6%22%20font-family%3D%22Inter%2Csystem-ui%2CArial%22%20font-size%3D%2234%22%20font-weight%3D%22700%22%20fill%3D%22url(%23g)%22%20fill-opacity%3D%220.9%22%3ELlm%3C%2Ftext%3E%3C%2Fg%3E%0A%20%3C!--%20Code%20card%20background%20for%20contrast%20--%3E%0A%20%3Cg%20opacity%3D%220.18%22%3E%0A%20%3Crect%20x%3D%2260%22%20y%3D%22378%22%20rx%3D%2216%22%20ry%3D%2216%22%20width%3D%22600%22%20height%3D%22176.4%22%20fill%3D%22%230b1020%22%20stroke%3D%22url(%23g)%22%20stroke-opacity%3D%220.35%22%2F%3E%0A%20%3C%2Fg%3E%0A%20%0A%20%3Crect%20width%3D%221200%22%20height%3D%22630%22%20fill%3D%22%23000%22%20opacity%3D%220%22%20filter%3D%22url(%23grain)%22%2F%3E%0A%3C%2Fsvg%3E)
الذكاء الاصطناعي والتعلم الآلي١١ دقيقة
TurboQuant من Google: استهلاك ذاكرة أقل بـ ٦ مرات لاستنتاج LLM (٢٠٢٦)
تقوم أداة TurboQuant من Google بضغط ذاكرة KV caches الخاصة بنماذج LLM إلى ٣ بت دون أي فقدان في الدقة، مما يقلل استهلاك الذاكرة بمقدار ٦ مرات ويسرع حسابات attention في معالجات H100 حتى ٨ مرات مقارنة بـ FP32.