AIHiranyagarbha · Class 11-12
GPU, parallel compute aur rent-vs-own ka trade-off
India ne AI ke liye GPU kiraye par lene ka model chuna — daam doguna hone par yeh model kyun tootne laga, aur khud khareedna kya pura hal hai?
Socho pehle… phir dekho jawab
Kyunki kiraye ka model daam aur supply ke jokhim supplier par daalta hai: daam badhe toh supplier delivery mein deri karte hain. Khud khareedna control deta hai, par kharch aur hardware ke purane hone ka jokhim sarkar par aa jaata hai.
Mechanism: AI model train karna bahut saare ginti-operations (jaise matrix ke guna aur jod) ka dohrana hai. GPU ek saath bahut saare chhote operations parallel mein chalata hai, isliye train karne ke liye zaroori hai. IndiaAI Mission ne March 2024 mein subsidised GPU access ka vaada kiya tha, jisme sarkar private suppliers se GPU hours kiraye par leti hai.
Data: Economic Times ke mutabik GPU, memory, storage aur networking ke daam ek saal mein 'more than doubled' hue. Nvidia ka H200 lagbhag $20,000 se $40,000 se upar gaya; server lagbhag Rs 2 crore se Rs 4 crore hua; rupaya Rs 88 se Rs 96-97 per dollar hua. Mission ke paas lagbhag 30,000 GPU hain jabki vaada 45,000 ka tha. Ek supplier ne lagbhag 1,100 GPU offer kiye aur lagbhag 160 diye.
Trade-off: C-DAC lagbhag 3,000 GPU khud khareedne ki bid laayi hai, aur saath mein lambi avadhi ke contracts bhi. Khud khareedne se control milta hai, par hardware jaldi purana hota hai aur poonji ka jokhim sarkar uthati hai. Kiraye par lene se shuru mein bojh kam rehta hai, par supplier ki delivery aur daam ka jokhim bahar se aata hai. Ek mixed model dono ko santulit karta hai. Dhyan rahe ki card ke mutabik India yeh chips khud nahi banata, isliye dollar aur supply chain ka jokhim bana rehta hai.
Ghar pe try karo
Ek short answer likho (100 words): 'Rent ya khareedna — IndiaAI jaisa mission kaun se 3 sawaal poochkar tay kare?' Phir ek line: parallel compute aur sequential compute mein kya farak hai?
Wow!
Ek hi supplier ne lagbhag 1,100 GPU ka vaada kiya aur lagbhag 160 diye — yahi hai rent model ka supplier-risk!
3 sawaal
- AI train karne ke liye GPU kyun upyogi hai?
Jawab dekho
Wo bahut saare chhote operations parallel mein chalata hai. GPU parallel mein bahut saari chhoti ginti karta hai.
- Rent model mein supplier-risk ka kya matlab hai?
Jawab dekho
Supplier daam badhne par vaada poora nahi kar paata. Daam doguna hone par kuch supplier poore GPU de nahi paaye.
- Khud GPU khareedne ka ek trade-off kya hai?
Jawab dekho
Control milta hai par hardware jaldi purana hota hai. Control ke saath purane hone ka aur poonji ka jokhim aata hai.
NCERT se jodo: NCERT Class 11 Computer Science / Informatics Practices — computer system aur processors; NCERT Class 12 Economics — sarkari budget aur arthvyavastha