Network-on-Chip (NoC) topologies: Question bank — Hinglish deep dive

6.3.6 · D5 · Hardware › Interconnects, Buses & SoC › Network-on-Chip (NoC) topologies

Yeh deck un misconceptions pe attack karta hai — woh jagahein jahan intuition tumhe NoC topologies ke baare mein quietly gumraah karta hai. Har answer tumhe reasoning deta hai, sirf verdict nahi. Question padho, answer cover karo, apni reasoning zyaaze bolo, phir reveal karo.

Koi bhi question samajhne se pehle, hamesha use kiye jaane wale symbols ko samajhna zaroori hai. Yeh glossary ek baar padho; neeche diye figures har idea ko visible banaate hain.

Metrics ke liye plain-word anchors:

Hop ::: Ek router se directly-linked neighbour tak ek step; yahaan "distance" hops mein count hoti hai, millimetres mein nahi.
Diameter ::: Kisi bhi do nodes ke beech worst-case shortest hop-distance — woh sabse lamba safar jo tumhe karna pad sakta hai.
Bisection bandwidth ( $B_{b}$ ) ::: Network ko ek cut se do equal halves mein kato; cut se kaate gaye links count karo, $B_{link}$ se multiply karo. Yeh woh sabse tanga gala hai jisse halves ke beech traffic guzarni padti hai.
Degree ::: Ek node se kitne links nikalte hain — yahi woh cheez hai jo ek router ka port count, aur isliye uska area aur power, pay karta hai.

True or false — justify

Recall

T/F — Ek torus hamesha same-size mesh ki latency ko half kar deta hai. False ::: Yeh roughly diameter (worst hops) ko half karta hai, lekin iske wraparound wires physically lambe hote hain, isliye har hop zyada cycles aur zyada power cost kar sakta hai — end-to-end latency simply half nahi hoti. Latency and Throughput Trade-offs yeh precisely batata hai: latency = (hops) × (per-hop cost) + queueing, aur sirf pehla factor shrink hota hai.

Recall

T/F — Hypercube hamesha mesh se better hai kyunki iska diameter $lo g_{2} N$ hai. False ::: Diameter sirf ek axis hai. Hypercube degree $lo g_{2} N$ ke saath grow karta hai (har naya dimension har node mein ek port add karta hai), isliye routers ke zyada ports hote hain (area, power, wiring), aur graph non-planar hai — flat silicon pe lay out karna mushkil hai. "Better diameter" ka matlab "better chip" nahi hai.

Recall

T/F — Ring ki bisection bandwidth hamesha kharaab hoti hai chahe kitne bhi nodes hon. True ::: Ring ko kahin bhi kato aur tum hamesha exactly do links severs karte ho, toh $B_{b} = 2 B_{link}$ chahe $N$ kuch bhi ho — nodes add karne par throat kabhi wide nahi hota, isliye yeh choke karta hai.

Recall

T/F — Topology mein zyada links add karna performance ke liye sirf helpful ho sakta hai. False ::: Har extra link ek extra router port hai jo area, power, aur control complexity cost karta hai; ek 16-port router ek 4-port wale se kaafi bada hota hai. Ek point ke baad, power/area budget ko cache ya cores pe spend karna zyada better hai.

Recall

T/F — Mesh mein, har node identically behave karta hai. False ::: Interior nodes ka degree 4 hota hai, lekin edges ka 3 hota hai aur corners ka sirf 2 — yeh asymmetry matlab hai ki edge nodes ke paas fewer paths hain aur alag congestion hai, isliye performance uniform nahi hai.

Recall

T/F — Torus mesh ki node-asymmetry problem fix karta hai. True ::: Wraparound links har node ko degree 4 dete hain, isliye saare nodes topologically equal hain — woh uniformity evenly-spread traffic ke liye ek real advantage hai.

Recall

T/F — Fat-tree classic tree ke root bottleneck ko remove karta hai. True ::: Ek plain tree sabhi cross-network traffic ko ek thin root link se funnel karta hai. Ek fat-tree root ki taraf links ko wider karta hai taaki har horizontal cut same bandwidth carry kare — "fat" ka poora point yahi hai.

Recall

T/F — Bisection bandwidth aur diameter same cheez measure karte hain. False ::: Diameter ek latency quantity hai (ek packet ke liye worst-case hop count); bisection bandwidth $B_{b}$ ek throughput quantity hai (ek cut ke paas aggregate data). Ring ki diameter chhoti-si hoti hai lekin bisection dreadful — yeh dono sharply disagree kar sakte hain.

Recall

T/F — Hypercube ki $N /2$ links ki bisection bandwidth ise wire-cheap banati hai. False ::: High bisection ka matlab hai ki bahut saari wires midline cross karni chahiye — yeh bandwidth-rich hai lekin wire- aur power-expensive hai, cheap ka ulta. High bisection ek cost hai, discount nahi.

Spot the error

Recall

"Ek $8 \times 8$ mesh ka diameter $⌊ 8/2 ⌋ + ⌊ 8/2 ⌋ = 8$ hai." Error ::: Yeh torus formula hai. Mesh mein koi wraparound nahi hota, isliye iska diameter $(N - 1) + (M - 1) = (8 - 1) + (8 - 1) = 14$ hai; $⌊ N /2 ⌋$ shortcut tabhi exist karta hai jab tum wrap around kar sako.

Recall

"8 nodes wale ring ka diameter 8 hai kyunki 8 nodes hain." Error ::: Tum ring ke dono directions mein travel kar sakte ho, isliye worst case sirf half-way hai: $⌊ 8/2 ⌋ = 4$ hops. Do-directions option bhoolna distance ko double-count kar deta hai.

Recall

"Hypercube mein, node 3 se node 12 tak routing mein utne hi hops lagte hain jitna numeric difference $12 - 3 = 9$ hai." Error ::: Distance differing bits ki count hai (Hamming distance), numeric subtraction nahi. $0011 \oplus 1100 = 1111$ → 4 bits differ → 4 hops, 9 nahi. Jab kai bits differ kar rahe hon toh kaun sa bit flip karna hai — yahi exactly Routing Algorithms decide karte hain.

Recall

"Fat-tree ke leaves aur root ko same link width use karni chahiye taaki simple rahe." Error ::: Equal widths root bottleneck recreate kar deti hain jo fat-tree remove karne ke liye exist karta hai. Link bandwidth ko root ki taraf grow karna chahiye (binary fat-tree mein har level par doubling) taaki har cut equally wide rahe.

Recall

" $4 \times 4$ torus ki bisection bandwidth mesh ke barabar hai, kyunki dono 4 columns cut karte hain." Error ::: Torus wraparound links bhi cut karta hai, isliye ek bisection $2 \times min (N, M)$ links severs karti hai — mesh ka double, jo $B_{b} = 2 min (N, M) B_{link}$ deta hai. Torus bisection mesh ke barabar nahi, double hai.

Recall

"Kyunki hypercube ko $2^{k}$ nodes chahiye, ek 12-node system nearest hypercube use karta hai." Error ::: Dimension $k$ ka hypercube exactly $N = 2^{k}$ nodes rakhta hai; 12 kisi power of two nahi hai. Ya tum 16 tak pad karo (ports waste karo) ya ek aise topology chuno jaise mesh/torus jo arbitrary $N$ accept kare.

Why questions

Recall

Shared buses roughly 8–16 cores ke baad collapse kyun ho jaate hain jabki NoCs scale karte rehte hain? Koi magic number nahi hai — ceiling woh jagah hai jahan saare cores ki bandwidth demand ka sum ek bus ke fixed $B_{bus} = f \cdot w$ se zyada ho jaata hai. 2 GHz pe 128-bit bus (256 Gb/s) cores mein split karke, ~8–16 cores mein se har ek ko sirf single-digit Gb/s milta hai aur contention pe stall ho jaate hain; NoC ke parallel links capacity per core add karte hain.

Recall

Hypercube ka diameter logarithmic kyun hai jabki mesh ka $N$ ke saath grow karta hai? Har hypercube hop ek address bit flip karta hai, "remaining difference" ko half karta hai, isliye $k = lo g_{2} N$ hops kaafi hain. Mesh sirf ek physical axis mein ek grid-step per hop move karta hai, isliye distance side length ke saath grow karta hai — $N^{2}$ total nodes ke liye side $N^{2} = N$ hai, matlab $total$ .

Recall

Fat-tree har level upar jaate hue link width double kyun karta hai jabki mesh links uniform rakhta hai? Tree mein, traffic shared root ki taraf climb karte waqt concentrate hoti hai, isliye upper links ko choke avoid karne ke liye wider hona chahiye. Mesh mein koi single funnel point nahi hota — traffic kai parallel links mein spread hoti hai, isliye uniform width kaafi hai.

Recall

Designer high-bisection hypercube ke bajaaye low-bisection ring kyun choose kar sakta hai? Rings ka degree 2 hota hai (tiny routers), minimal wiring, aur low power — ideal jab traffic light ho aur power/area budget tight ho. Peak bandwidth waste hoti hai agar tum kabhi itna traffic generate hi nahi karte ki use karo.

Recall

Modern nodes mein NoCs ki taraf move mein gate delay nahi, wire delay kyun drive karta hai? Transistors shrink aur speed up hote rahe, lekin wires utni acchi tarah scale nahi hue; ek lamba cross-chip wire ab traverse karne mein kai cycles cost karta hai. NoCs lambi global wires ko short router-to-router hops mein break karte hain, har signal path ko fast rakhte hain. Yeh System-on-Chip (SoC) Design ka ek core constraint hai, jahan floorplan, logic nahi, timing budget set karta hai.

Recall

Torus ko physically mesh se zyada lay out karna mushkil kyun hai, jabki same degree hai? Iske wraparound links ko chip ke ek edge se bilkul opposite edge tak jump karna hota hai — bahut lambe wires jo delay, power, aur routing congestion badhate hain, mesh ke uniformly short neighbour links ke unlike.

Recall

Do topologies jinka identical diameter ho phir bhi wildly different real latency kyun ho sakti hai? Hop count per-hop cost ignore karta hai: ek lamba wraparound ya fat-tree link ke upar hop ek short mesh hop se zyada cycles leta hai, aur congestion queueing delay add karta hai. Diameter steps count karta hai, time nahi — bilkul wohi distinction jis par Latency and Throughput Trade-offs built hai.

Recall

Ek topology ka huge bisection bandwidth $B_{b}$ hone ke bawajood slow kyun lag sakta hai? $B_{b}$ ek cut ke paas peak throughput ceiling hai; ek single packet phir bhi puri diameter hops pay karta hai. Throughput aur latency alag axes hain, isliye wide-throat network phir bhi ek akela packet kaafi dur travel kara sakta hai.

Edge cases

Recall

$1 \times N$ mesh (ek single row) ka diameter aur bisection kya hai? Ek single row linear array hi hai: diameter $= (N - 1) + (1 - 1) = N - 1$ (koi doosra dimension nahi hai ise shorten karne ke liye), aur ek mid-cut sirf ek link severs karta hai, toh $B_{b} = 1 \cdot B_{link}$ — essentially ek khula hua ring.

Recall

$N = 2$ par ring ka diameter formula kya hoga? $⌊ 2/2 ⌋ = 1$ : do nodes ek hop apart hain — sahi hai, kyunki 2-node ring sirf ek single bidirectional link hai (do "directions" coincide ho jaate hain).

Recall

$2 \times 2$ mesh ke liye, kya "interior degree 4" rule apply hota hai? Nahi — $2 \times 2$ mesh mein har node corner hai, isliye sab ka degree 2 hai. Degree-4 interior sirf tab appear hota hai jab grid kam se kam $3 \times 3$ ho; chhote grids mein sab boundary hote hain.

Recall

$k = 1$ hypercube kya hota hai, aur kya $lo g_{2} N$ diameter phir bhi hold karta hai? $k = 1$ se $N = 2^{1} = 2$ nodes milte hain jo ek link se jude hain; diameter $= k = lo g_{2} 2 = 1$ hop — consistent. Sabse chhhota hypercube sirf do connected nodes hai.

Recall

Fat-tree mein, same switch ke neeche do leaves ke beech traffic kya carry hoti hai — kya yeh kabhi root tak pahunchti hai? Nahi — aisi traffic lowest common switch par wapas mud jaati hai, isliye nearby communication sasta hota hai aur kabhi fat root links ko burden nahi karta. Sirf cross-subtree traffic oonchi climb karti hai; yahi locality hai jis wajah se coherence traffic locality-aware placement prefer karta hai.

Recall

Ek single isolated node (N=1) ki bisection bandwidth kya hai? Koi cut nahi banaya ja sakta — koi do halves exist nahi karte — isliye bisection bandwidth undefined/zero hai; yeh metric meaningful tabhi banta hai jab network mein split karne ke liye kuch ho.

Recall

Kya odd-length ring diameter formula change karta hai? Nahi — $⌊ N /2 ⌋$ phir bhi hold karta hai; e.g. 7-node ring ka diameter 3 hai, kyunki floor correctly us half-hop ko drop karta hai jo odd count otherwise imply karta.