Foundations — Domain-specific accelerators

6.5.6 · D1 · Hardware › Advanced & Emerging Architectures › Domain-specific accelerators

Is page ke liye assume kiya gaya hai ki aapko kuch nahi aata siwaaye basic arithmetic aur is baat ke ki ek computer chip roughly kya hoti hai. Har woh symbol jo parent note mein use hota hai, yahan ground up se build kiya gaya hai. Upar se neeche padho — har block agli block ko earn karta hai.

0. Chip ki picture (taaki baaki sab sense kare)

Ek chip silicon ka flat square hota hai. Us par billions of tiny switches hote hain jinhein transistors kehte hain. Transistors ke groups blocks banate hain — kuch arithmetic karte hain, kuch instructions fetch karte hain, kuch data store karte hain. Neeche sab kuch is baare mein hai ki kaun se blocks banane layak hain aur kyun.

Figure s01 — chip floor-plan. Chhe labelled boxes dekho: amber box woh ALU hai jahan actual math hoti hai; har white box (fetch, decode, branch prediction, reorder, cache) overhead hai — guessing aur shuffling mein spend hota kaam, computing nahi. Notice karo kitne kam boxes amber hain. Yahi imbalance hai jis ki wajah se DSAs exist karte hain: ek fixed kaam ke liye white boxes delete karo aur almost saari energy real math mein jaati hai.

1. Energy aur power — woh units jinse sab kuch measure hota hai

Isse pehle ki hum keh sakein "DSAs jeet jaate hain", humein yeh kehna hoga ki kismein jeet. Jawab hai energy per operation, isliye humein do words ko clearly samajhna hoga jo log confuse karte hain.

2. Power density — woh wall jisne sab shuru kiya

Ab woh historical rule jo pehle humein bachata tha:

Figure s02 — kyun shrinking free rehna band ho gayi. Pehle cyan line follow karo: Dennard era ke dauran, generation after generation, power density flat rehti hai — free lunch. Dashed line (~2006) par amber line aage aati hai aur steeply chadhti hai: har nayi generation ab same square millimetre mein zyada heat pour karti hai. Amber arrow consequence ki taraf point karta hai — tum power-density ceiling se takraate ho aur transistors ko dark rakhna padta hai.

3. Ek single core se diminishing returns — do named rules

Chahe hum ek bade core ko power de bhi sakein, ek core ko bada banana poorly pay off karta hai.

4. Accelerator ke building blocks

Ab ek DSA kis cheez se bana hota hai ka vocabulary.

Figure s03 — ek baar load karo, stream karo, reuse karo. White arrows weights ko grid mein neeche drop karte hain — har weight ek baar load hoti hai aur waheen rehti hai. Amber arrows activations ko right taraf push karte hain, ek column of cells se agle tak. Ek single weight ko ek cell mein sit karte hue follow karo: har activation jo past flow karti hai woh ek MAC trigger karti hai, toh woh ek loaded number row mein $N$ baar reuse hoti hai. Yahi reuse — fetch kiye gaye har byte par bahut saari operations — wahi hai jo agla section arithmetic intensity kehta hai.

5. Precision — number par kam bits spend karna

6. Arithmetic intensity aur Roofline — decision tool

Parent note mein yeh sabse zyada symbol-heavy idea hai, toh hum har letter build karte hain.

Figure s04 — roofline padhna. Do ceilings "roof" banate hain: sloped amber line (slope $=B$, memory limit — bytes kitni tezi se pahunchte hain) aur flat white line ($P_{\text{peak}}$, compute limit). Tumhara kernel intensity axis par ek vertical position hai. Amber dot $I=50$ par sloped part par land karta hai, $P_{\text{peak}}$ se kaafi neeche: tum memory-bound ho, 15 TOPS par capped. Dotted "ridge" $I=300$ par woh jagah hai jahan do ceilings milti hain — sirf uss ke right wale kernels compute-bound hain. Dot ko right mein move karna (zyada reuse) yahan chadne ka ek hi tarika hai.

Prerequisite map

Equipment checklist

Right side cover karo aur zor se jawab do. Agar koi bhi stumps kare, uska section upar dobara padho.