Foundations — NVLink and GPU interconnects

6.3.4 · D1 · Hardware › Interconnects, Buses & SoC › NVLink and GPU interconnects

Pehle tumhe parent note NVLink and GPU interconnects padhne se pehle ek chota sa set of symbols aur units mein fluent hona hoga jo woh bina warning ke throw karta hai. Yeh page har ek ko bilkul kuch bhi nahi se build karta hai, ek aisi sequence mein jahan har idea sirf pehle waale ideas pe lean karta hai. Agar Hinglish version tumhare liye aasaan hai, toh the Hinglish note ko iske saath padhna.

1. Ek bit, aur "ek wire data carry karna" ka asli matlab

Ek single wire ki picture banao. Time left se right ki taraf chalta hai. Wire ki voltage do levels ke beech bounce karti hai. Har chota time-slot ek bit hold karta hai.

YEH figure KYA dikhata hai: ek voltage-versus-time trace jo equal slots mein cut hai, har slot ek 1 ya 0 hai. KYUN hume isko chahiye: parent note mein har number — 20 GT/s, 16 Gb/s, 20 GB/s — sirf yeh count karne ka ek tarika hai ki yeh kitne slots per second guzarte hain. Yeh picture sahi kar lo aur baaki sab arithmetic hai.

$1\text{ byte}=8\text{ bits}⟹1\text{B}=8\text{b}$

2. Speed count karna: "per second" family

Ab hum count karte hain ki kitne bits (ya symbols, ya bytes) ek point se har second guzarte hain. Har prefix sirf roughly ek hazaar se multiply karta hai. Hum wahi clean powers-of-ten convention use karte hain jo parent note use karta hai.

Teen kyun? Kyunki ek wire wiggle (ek transfer) hamesha exactly ek useful bit carry nahi karta. Woh gap hi agla idea hai.

3. Transfers vs. payload: kyun 20 becomes 16

OVERHEAD KYUN EXIST KARTA HAI (8b/10b): agar tum 0s ki ek lambi run bhejte, toh wire flat baith jaati aur receiver ki clock drift kar jaati — woh track kho deta hai ki slots kahan start hote hain. Toh sender har 8 real bits ko ek 10-bit pattern mein rewrite karta hai jo wiggling guarantee karta hai. Woh hai 8b/10b encoding. Tum har 8 ke liye 2 extra bits pay karte ho → tum fraction $\frac{8}{10}$ rakhte ho.

YEH figure KYA dikhata hai: ek 25 GT/s raw stream ek "8-of-every-10" filter se funnel ho raha hai, 20 Gb/s payload ke roop mein bahar aata hai. KYUN topic ko isko chahiye: exactly isliye datasheet "25" kehta hai lekin honest bandwidth "20" hai. Multiply-by-eight-tenths hi poora trick hai.

4. Lanes aur links: wires bundle karna

Ek link ki speed pane ke liye, apni lanes add karo:

$8\text{lanes}\times 20\text{Gb/s}=160\text{Gb/s}=20\text{GB/s per direction}$

YEH figure KYA dikhata hai: do wires ek lane form kar rahe hain, aath lanes ek link mein stack hain, aur GT/s → Gb/s → GB/s → link tak ka arithmetic ladder. KYUN: parent note mein har headline "20 GB/s" yahi ladder ek baar chadha hua hai. Ladder ek baar dekh lo aur tumhe kabhi panic mein re-derive nahi karna padega.

5. Ek saath do directions: bidirectional / full-duplex

$20\text{GB/s (out)}+20\text{GB/s (in)}=40\text{GB/s bidirectional per link}$

6. Bandwidth aur latency: ek link ke do rulers

Yeh dono independent hain. Ek firehose (huge bandwidth) phir bhi lamba ho sakta hai (kuch latency). Parent note inhe multiply karta hai:

7. Bandwidth ko time mein convert karna: transfer formula

Parent note ke saare timing estimates ek hi rearrangement se aate hain: time = amount ÷ speed.

Prerequisite map

Equipment checklist

Self-test: right side cover karo aur reveal karne se pehle answer do.