Visual walkthrough — Wafer-scale engines (Cerebras-style)

6.5.17 · D2 · Hardware › Advanced & Emerging Architectures › Wafer-scale engines (Cerebras-style)

Hum central result $Y=e^{-DA}$ ko scratch se build karenge, dikhayenge kyun ye ek bade chip ko maar deta hai, phir redundancy fix $M\ge \frac{N}{1-f}$ derive karenge jo use bachata hai.

---

Step 1 — Wafer aur defect actually hote kya hain

WHAT: wafer ko ek flat baking sheet ki tarah imagine karo, aur defects ko uspar randomly bikhere hue raisins ki tarah.

WHY yahan se shuru karein: baad ke har symbol ($D$, $A$, $\lambda$) sirf ek area mein inhi raisins ko count karne ka tarika hai. Agar raisins dikh rahe hain, toh algebra obvious hai.

PICTURE: neeche ki sheet par defects bikre hue hain. Ek chota die (dashed box) sheet par baitha hai — dhyan do ki woh khali hai, toh woh survive karta hai. Ek bada die (solid box) sheet ka zyaadatar hissa cover karta hai aur kaafi defects niga l jaata hai, toh woh mar jaata hai.

---

Step 2 — Defect density $D$: raisins ko ek number mein todna

WHAT: humne "dots ka ek messy scatter" ko ek steepness-jaisi number mein compress kar diya.

WHY yeh tool aur sirf "defects count karo" kyun nahi? Raw count useless hai — ye depend karta hai ki tum kitne bade region ko dekh rahe ho. Density count per unit area hai, toh woh same rehti hai chahe tum ek tiny die dekhein ya poori plate. Yahi woh property hai jo hume chahiye, kyunki aage hum poochhenge "ek area $A$ wale die mein kitne defects land hote hain?" — aur density × area is sawaal ka jawab turant deta hai.

PICTURE: wahi wafer, ab ek $1{\text{mm}}^2$ reference square uspar drawn hai. Uss unit square ke andar dots count karna hi $D$ hai.

---

Step 3 — Ek die mein expected defects: product $\lambda =D\cdot A$

$\underset{\text{die mein avg defects}}{\underbrace{\lambda }}=\underset{{\text{defects per mm}}^2}{\underbrace{D}}\times \underset{{\text{mm}}^2\text{ of the die}}{\underbrace{A}}$

WHAT: density times area ek pure count deta hai — ${\text{mm}}^2$ units cancel ho jaate hain, "defects" bachta hai.

WHY: agar raisins $D$ per unit area ki rate se spread hain, toh $A$ units of area mein tum expect karte ho $D\cdot A$ of them. Yeh wohi logic hai jaise "$3$ cars per km over $10$ km $\Rightarrow$ expect $30$ cars." Hum is average ko $\lambda$ (Greek letter lambda) kehte hain kyunki yeh us distribution ke mean ka standard naam hai jise hum aage milenge.

PICTURE: do dies side by side — ek chhota ($A$ small $\Rightarrow \lambda$ small) aur ek wafer-sized ($A$ huge $\Rightarrow \lambda$ huge). Har ke neeche ka bar uss ki expected defect count area ke saath badhti dikhaata hai.

---

Step 4 — Poisson distribution kyun, aur $P(0)$ kyun

Sahi tool hai Poisson distribution: yeh batata hai ki exactly $k$ rare, independent events milne ki probability kya hai jab tum expect karte ho $\lambda$ of them.

Ab $k=0$ set karo (hume zero defects chahiye):

$Y=P(0)=\frac{{\lambda }^0{e}^{-\lambda }}{0!}=\frac{1\cdot e^{-\lambda }}{1}=e^{-\lambda }=\boxed{e^{-DA}}$

• ${\lambda }^0=1$ (kisi bhi cheez ki power $0$, $1$ hoti hai).
• $0!=1$ (kuch bhi arrange karne ka exactly ek hi tarika hai).
• Toh top wali power aur bottom wala factorial dono $1$ ho jaate hain, aur clean $e^{-DA}$ bachta hai.

WHAT: humne $k=0$ ko general rule mein plug kiya aur $e^{-\lambda }$ ke siwa sab kuch collapse ho gaya.

WHY: "$k=0$" hi woh engineering sawaal hai "kya yeh die har defect se bach gayi?" — yield $Y$.

PICTURE: ek fixed $\lambda$ ke liye Poisson bar chart. $k=0$ par single bar highlight hai — woh akela bar hi yield hai.

---

Step 5 — Yield curve, aur bade die ki maut

Chalte hain curve ko teen regimes mein — all-cases requirement:

• Tiny die ($DA\to 0$): $e^{-\text{(small)}}\approx 1$. Almost har die perfect hai. Iseelie hum normally wafer ko chhote chips mein dice karte hain.
• Medium die ($DA\approx 1$): $Y\approx e^{-1}\approx 0.37$. Lagbhag ek-tihaii survive karte hain — normal chip economics.
• Whole-wafer die ($DA\gg 1$, jaise $DA=46$ ek plate-sized die ke liye realistic $D$ par): $Y=e^{-46}\approx 10^{-20}$. Tumhe ek perfect wafer paane ke liye ek kamre mein atoms se zyada wafers chahiye honge.

WHAT: exponential ne ek badi area ko ek vanishingly small survival chance mein badal diya.

WHY iska matter karta hai: yeh parent note ka "yield killer" hai. Ek monolithic, defect-free full wafer impossible hai. Agar yahi ek option hota, toh wafer-scale engines exist nahi kar sakte the.

PICTURE: smooth decay curve $Y=e^{-DA}$ jisme teen regimes dots se mark hain. Dekho curve kaise cliff se girta hai jab hum right ki taraf wafer territory mein jaate hain.

---

Step 6 — Sawaal badlo: dead cores ko tolerate karo

Nayi quantities define karo:

• $M$ = wafer par physically built cores ki number.
• $f$ = cores ka woh fraction jo dead hone ki expect kiya jaata hai (defects se). Toh $f$ $0$ aur $1$ ke beech hai.
• $N$ = achhe cores ki number jo hume actually product ke liye chahiye.

Surviving cores ki number built cores minus dead cores hai:

$\text{good cores}=\underset{\text{built}}{\underbrace{M}}-\underset{\text{dead}}{\underbrace{fM}}=M(1-f)$

Hume sirf itne survivors chahiye:

$\underset{\text{good cores}}{\underbrace{M(1-f)}}\ge \underset{\text{cores needed}}{\underbrace{N}}$

$M$ ke liye solve karo dono sides ko $(1-f)$ se divide karke (jo positive hai, toh inequality direction safe hai):

$\boxed{M\ge \frac{N}{1-f}}$

• $\frac{1}{1-f}$ = over-provisioning factor: strictly zaroorat se itna zyada cores banao.

WHAT: humne "wafer perfect hona chahiye" ko "wafer ke paas $\ge N$ good cores hone chahiye" se replace kiya, aur ek build recipe mili.

WHY yeh tool (simple ratio, exponential nahi): yield $e^{-DA}$ ne jawab diya "kya poora die perfect hai?" — galat sawaal. Sahi sawaal survivors ke baare mein ek counting sawaal hai, aur counting ko sirf "kept fraction × total," yaani simple proportions chahiye.

PICTURE: ek mesh grid jisme kuch cores crossed out hain (dead). Ek path unke aas-paas thread karta hai ek side se doosri side tak — reroute. Ek counter dikhata hai good cores abhi bhi $\ge N$ hain.

---

Step 7 — Over-provisioning factor ko saare cases mein padhna

$M\ge \frac{N}{1-f}$

Chalte hain $f$ ke har regime ko cover karte hain taaki koi scenario surprise na kare:

• $f=0$ (koi defect nahi): $M\ge \frac{N}{1}=N$. Exactly utna banao jitna chahiye — ideal, kabhi real nahi.
• $f=0.10$ (10% mar jaate hain): $M\ge \frac{N}{0.9}\approx 1.11N$. Lagbhag 11% extra cores spares ke roop mein banao.
• $f=0.5$ (aadhe mar jaate hain): $M\ge 2N$. Tumhe double banana hoga.
• $f\to 1$ (almost sab mar jaate hain): $\frac{1}{1-f}\to \infty$. Factor blow up hota hai — wafer-scale hopeless ho jaata hai. Yeh degenerate limit hai: redundancy tabhi kaam karta hai jab $f$ chhota ho.

WHAT: ratio ne humein exactly bataya ki kitne spares bake in karne hain.

WHY iska matter karta hai: isliye parent ka core count woh hai jo hai — woh number round marketing nahi hai; yeh $N/(1-f)$ rounded up hai.

PICTURE: over-provisioning factor $\frac{1}{1-f}$ ko $f$ ke against plot kiya. $f=0$ ke paas flat aur sasta, $f\to 1$ ke paas infinity tak shoot karta — marked points $f=0,0.1,0.5$ dikhate hain "cheap zone" jisme WSEs rehte hain.

---

Ek-picture summary

Yeh final figure poori journey compress karta hai: left par exponential yield wall (perfection impossible hai), beech mein mesh-with-reroute (sawaal badlo), aur right par gentle over-provisioning curve (ek chhota spare fraction sab theek kar deta hai).

Recall Feynman: poora walkthrough plain words mein dobara sunao

Raisins (defects) ko randomly ek baking sheet (wafer) par bikherdo. Dekho woh per unit area kitne crowded hain — woh hai $D$. Apne cookie ke size ($A$) se multiply karo aur tum paate ho iske andar kitne raisins expect hote hain, jo $\lambda$ kehlaata hai. Lekin tum sirf woh cookies chahte ho jinmein zero raisins hon, aur us chance ki probability — independent sprinkles ke liye — hai $e^{-\lambda }=e^{-DA}$. Ek chhote cookie ke liye woh nearly $1$ hai (easy). Poori sheet ke size ke cookie ke liye, $\lambda$ bahut bada hai aur chance basically zero hai — ek perfect giant cookie kabhi nahi hoga. Trick yeh hai: ek giant cookie mat banao, balki tiny connected cells ki ek giant sheet banao. Kuch cells ko raisins milte hain aur woh mar jaate hain, lekin cells ke beech ke raaste tumhe dead wale skip karne dete hain. Ab tumhe sirf "enough living cells" chahiye, toh tum thode extra bake karo: $M=N/(1-f)$ cells banao, aur agar sirf ek chhota fraction $f$ mare toh tumhare paas apne $N$ good wale hain. Physics ne kaha "impossible"; sawaal ko "enough good cores" mein badalna ne ise easy bana diya.

Recall Khud check karo

Poisson factor $e^{-\lambda }$ kisi doosre function ki jagah kyun aata hai? ::: Kyunki defects independently aur memorylessly land karte hain; independent random arrivals hamesha factor $e^{-\lambda }$ accumulate karte hain, jo "har tiny region independently hit ho sakti hai" ka fingerprint hai. WSE ke liye $Y=e^{-DA}$ galat metric kyun hai? ::: Yeh ek defect-free whole die measure karta hai; WSEs ko kabhi perfection nahi chahiye, sirf enough good cores, toh sahi metric survivor count hai. Agar $f=0.2$ ho, toh $N$ good cores paane ke liye kitne cores banane honge? ::: $M\ge N/(1-0.2)=1.25N$, yaani 25% extra.

Related: Yield & Defect Density Models · Dennard Scaling & the Memory Wall · Network-on-Chip (NoC) & Mesh Topologies · Chiplets & 2.5D/3D Integration · Systolic Arrays & TPUs · Sparsity in Neural Networks · Liquid Cooling & Power Delivery Networks