Visual walkthrough — Energy efficiency (performance per watt)

6.4.10 · D2 · Hardware › Power, Thermal & Reliability › Energy efficiency (performance per watt)

Hum yeh cheezein, isi order mein, earn karne wale hain:

1. "Capacitor" kya hota hai aur ek logic gate us jaisa kyun hota hai.
2. Use charge karne mein energy $\frac{1}{2}C{V}^2$ kyun lagti hai.
3. Yeh $f$ times per second karne se power $\alpha C{V}^{2}f$ kyun banta hai.
4. Faster jaana voltage ko secretly upar kyun khींchta hai.
5. Yeh power ko cube mein kyun badal deta hai aur efficiency ko $1/f^2$ mein.

Neeche kuch bhi assume nahi kiya gaya ki aapne yeh pehle dekha hai. Chalo paani ki ek bucket se shuru karte hain.

Step 1 — Ek logic gate ek chhoti bucket hai (ek capacitor)

Yeh picture kyun? Ek capacitor invisible aur abstract hota hai, lekin ek paani ki bucket har agle step ko physical bana deti hai: "voltage" paani ki height ban jaata hai, "charge" paani ki miqdar ban jaati hai, aur "energy" paani ko upar uthane ki koshish ban jaayegi. Laal bucket dekho — woh laalpanat is page par har figure mein "jis cheez ke liye hum pay kar rahe hain" ke saath follow karegi.

Step 2 — Bucket bharne mein energy $\frac{1}{2}C{V}^2$ lagti hai

$\frac{1}{2}$ kyun aaya? Figure mein laal triangle dekho. Energy push-vs-charge line ke neeche ka area hai. Kyunki push linearly $0$ se $V$ tak badhti hai jab charge $0$ se $Q$ tak badhta hai, woh area ek triangle hai: $\frac{1}{2}\times \text{base}\times \text{height}=\frac{1}{2}QV=\frac{1}{2}C{V}^2$. Triangle aadha rectangle hota hai — literally yahi se $\frac{1}{2}$ aata hai.

Step 3 — Yeh $f$ times per second karne se power $\alpha C{V}^{2}f$ milta hai

$f$ yahan sirf linear kyun hai? Figure mein pulses ki row dekho — $f$ double karna sirf ek second mein twice as many identical pulses pack kar deta hai. Har pulse ka cost same $C{V}^2$ hai. Toh energy-per-second double hoti hai: ek seedha, linear relationship. Isko $V^2$ se compare karo, jahan har pulse khud mehengi ho jaati hai. Yeh asymmetry yaad rakho — Step 5 ise detonate karta hai.

Step 4 — Hidden coupling: faster jaana secretly zyada voltage maangta hai

Higher $V$ ka matlab faster switching kyun hai? Figure mein, laal curve (high $V$) switching threshold line tak black curve (low $V$) se pehle pahunchti hai. Zyada tall starting push gate ko jaldi charge karta hai, toh bit pehle ready hoti hai — aapko chip ko higher clock karne deta hai. $V$ cut karna fill slow kar deta hai, toh aapko lower clock karna padta hai. Isliye aap cheat nahi kar sakte: speed aur voltage saath badhte aur girte hain.

Step 5 — Substitute karo, aur power $f^3$ mein explode ho jaati hai

Ab efficiency. Performance (throughput) sirf linearly badhti hai $f$ ke saath — double clock, double kaam: $\text{Performance}\propto f$

Performance ko power se divide karo toh performance per watt milta hai (jise parent note ne prove kiya tha ki secretly operations per joule hai): $\frac{\text{Performance}}{\text{Power}}\propto \frac{f}{f^3}=\boxed{\frac{1}{f^2}}$

Yeh kyun poori kahani hai. Figure mein, kaali performance line dheere aur seedhi chadhti hai, jabki laal power curve $f^3$ ki tarah rocket ki tarah upar jaati hai. Unke beech ka gap barbad efficiency hai. $f$ upar push karo aur aap thoda zyada kaam khareedoge bahut zyada heat ke liye — efficiency $1/f^2$ ki tarah girती hai. Yahi ek divergence hai jiske wajah se industry ne clock speed chase karna band kiya aur Multicore and Parallelism ki taraf mudi, aur isliye DVFS - Dynamic Voltage and Frequency Scaling is curve par neeche move karke itna bachata hai.

Step 6 — Degenerate cases (ambush mat ho jaana)

Real chips extremes par nahi rehte, toh chalte hain check karte hain wahan kya hota hai.

Teeno kyun dikhao? Clean $1/f^2$ law sirf curve ka middle part hai. Low-$f$ end par leakage aapko floor kar deta hai; high-$f$ end par cube aapko ceiling kar deta hai. Figure ka laal U-shaped total curve real efficiency landscape dikhata hai — high efficiency ek valley mein rehti hai, aur Thermal Design Power (TDP) woh right wall fix karta hai jise cross karna allowed nahi hai.

Ek-picture summary

Yeh saare chhe steps compress karta hai: ek bucket ($C$) height $V$ tak bhari jaati hai jisme $\frac{1}{2}C{V}^2$ lagta hai (triangle); use ek second mein $f$ times pay karo $\alpha C{V}^{2}f$ watts ke liye; lekin $f$, $V$ ko upar khींchta hai, toh power $f^3$ ban jaati hai (laal rocket) jabki performance linear rehti hai (kaali line) — efficiency $1/f^2$ ki tarah girti rehti hai, tab tak valid jab tak leakage ya voltage floor ride khatam na kar de.