Processing-in-memory (PIM): Question bank — Hinglish deep dive

6.5.5 · D5 · Hardware › Advanced & Emerging Architectures › Processing-in-memory (PIM)

Is page par use hone wale symbols aur terms (quiz se pehle inhein banao)

Kisi bhi trap se pehle, yahan har woh letter aur abbreviation saaf alfazon mein diya gaya hai jo is page par use hota hai. Skip mat karo — traps tab hi samajh aate hain jab ye earn kar liye ho.

Formula Crossbar ke peeche ki ek equation:

$I = G V$ Ek wire ka grid socho: vertical columns input voltages carry karti hain, horizontal rows currents collect karti hain. Har crossing par ek memristor cell hoti hai (dekho Memristor / ReRAM Crossbars).

$V_{j}$ = column $j$ par apply ki gayi voltage (ye vector $V$ ka ek input number hai).
$G_{ij}$ = row $i$ aur column $j$ ke crossing par cell ki conductance (1/resistance) — isme matrix $G$ ka ek weight store hota hai.
$I_{ij}$ = woh current jo us cell ne row $i$ par push ki.

Ohm's law (kyun: ek memristor cell bas ek programmable resistor hai, aur $I = G V$ resistor law hai): $I_{ij} = G_{ij} V_{j}$ . Toh har cell ek weight ko ek input se multiply karti hai. Kirchhoff's current law (kyun: ek shared wire par aane wali currents add honi chahiye): row $i$ par total current $I_{i} = \sum_{j} G_{ij} V_{j}$ hai. Woh sum hi ek dot product hai — matrix–vector product ki ek row. Saari rows mein poora grid $I = G V$ ek single analog settling step mein compute karta hai. Isliye physics multiply-accumulate "free mein" karta hai.

Upar diya roofline sketch dekho: sloped teal line bandwidth roof hai ( $B$ se limited performance); flat orange line compute roof hai ( $R$ se limited). Jahan dono milti hain woh ridge point hai. Ek workload ka arithmetic intensity $I = C / N$ use left (bandwidth-bound → achha PIM target) ya right (compute-bound → PIM zyada nahi deta) mein rakhta hai. Yeh picture har trap ke liye yaad rakho.

True ya false — justify karo

PIM ka main benefit yeh hai ki woh arithmetic units (ALUs) ko faster banata hai.

False. PIM ki jeet data movement khatam karna hai, faster arithmetic nahi; compute-bound kaam ke liye wahi slow-ya-fast ALU math karta hai chahe kuch bhi ho. Dekho Memory Wall.

Agar ek workload compute-bound hai (high arithmetic intensity), toh PIM phir bhi ek bada speedup dega.

False. Jab

I = C / N

high hota hai, data movement pehle se waqt ka ek chhota sa hissa hoti hai, toh use hatane se zyada fayda nahi hota — roofline mein tum ridge ke right ho, flat compute ceiling ke neeche, bandwidth wale ke neeche nahi. Dekho Roofline Model.

Processing-using-memory (PUM) crossbars bilkul waisa hi bit-exact result dete hain jaise ek digital multiplier.

False. Equation

I = G V

kaagaz par exact hai, lekin conductances

G_{ij}

drift karte hain, aur ADCs/DACs quantization aur noise add karte hain, toh PUM results approximate hote hain — neural nets ke liye theek, IEEE-exact finance ke liye galat.

Ek bada cache add karna PIM jaisa hi kaam karta hai.

False. Cache CPU side par hoti hai aur sirf reused (temporal-locality) data ke liye helpful hai; PIM trip ko poori tarah khatam karta hai aur isliye streaming, low-reuse data mein bhi help karta hai jo cache capture nahi kar sakti.

PIM memory ki internal bandwidth

B_{in t}

external bus bandwidth

B

se zyada ho sakti hai.

True. DRAM ke andar,

k

banks mein se har ek

B_{bank}

par parallel stream karta hai, toh

B_{in t} = k \cdot B_{bank}

single external bus

B

ko bahut peeche chhhod sakti hai — yahi untapped parallelism hai jo PIM harvest karta hai. Dekho DRAM Organization (banks, rows).

Kyunki

E_{m o v e} \approx 100 - 1000 \times E_{o p}

hai, PIM se energy savings usually raw speedup se zyada hoti hai.

True (aksar). Lambi high-capacitance wires movement ko time se bhi zyada energy mein dominate karati hain, toh movement kam karna total energy dramatically gira sakta hai tab bhi jab timing speedup modest ho.

PIM Von Neumann Bottleneck ko poori tarah hata deta hai.

False generally. PIM bus par trips kam karta hai lekin rarely poori tarah khatam karta hai — results, control, aur non-local data abhi bhi travel karte hain; yeh bottleneck ko delete karne ki jagah narrow karta hai.

Ek memristor crossbar matrix size chahe jo bhi ho,

O (1)

analog steps mein matrix–vector product compute karta hai.

Multiply-accumulate ke liye True. Ohm's law + Kirchhoff sabhi

G_{ij} V_{j}

products ek settling step mein sum karte hain, lekin inputs load karne aur DACs/ADCs se outputs padhne mein abhi bhi

O (n)

lagta hai, toh end-to-end free nahi hai. Dekho Memristor / ReRAM Crossbars.

Near-memory PIM (PNM) ke liye ReRAM jaisi fundamentally nayi memory technology chahiye.

False. PNM sadharan logic (ALUs, chhote cores) use karta hai jo conventional memory ke physically paas rakha hota hai — jaise ek HBM 3D-Stacked Memory stack ke logic die par. Koi exotic device nahi chahiye.

Error dhundho

"Ek workload

1 0^{6}

ops karta hai aur

1 0^{6}

bytes move karta hai, toh iska arithmetic intensity 1000 ops/byte hai."

Error: intensity

I = C / N = 1 0^{6} /1 0^{6} = 1

hai, 1000 nahi. Shayad claimant ne galat scale ke byte count se divide kiya; hamesha

C

aur

N

ko same units aur orders of magnitude mein rakho.

"HBM 3D-stacked hai, aur 3D-stacking hi ise PUM (using-memory) design banati hai."

Error: 3D-stacking PNM (near-memory) enable karta hai — sadharan logic ek nearby die par hoti hai. PUM ka matlab hai memory array ki apni analog physics se compute karna; stacking akela PUM nahi hai. Dekho HBM 3D-Stacked Memory.

"Near-memory PIM ka speedup

B / B_{in t}

hai, external divided by internal bandwidth."

Error: ratio ulta hai. Speedup

\approx B_{in t} / B = k \cdot (B_{bank} / B)

— tum chahte ho ki badi internal bandwidth upar ho taaki speedup 1 se zyada ho.

"Kyunki PIM banks mein parallelize karta hai, banks badhaate jaane par iska speedup bina limit ke badh sakta hai."

Error: Amdahl's Law kaat deta hai — residual compute term

C / R

(rate

R

par ALU ka time) aur un-parallelizable serial fractions speedup cap kar dete hain; bottleneck ke baad banks add karna diminishing returns deta hai.

"Analog crossbars noisy hain, toh real computing ke liye bekar hain."

Error: overgeneralization. Approximate results error-tolerant workloads jaise Neural Network Accelerators ke liye bilkul theek hain; "noisy" ≠ "bekar," bas application domain narrow ho jaata hai.

"PIM free hai — hum bas DRAM chip par kuch logic bolt kar lete hain."

Error: DRAM ek aisa process use karta hai jo saste dense storage ke liye optimize hai aur logic ke liye bura hai; compute add karne par area, heat, aur memory density ka cost aata hai, saath hi programming-model aur coherence problems bhi hain. Yeh trade-off hai, free lunch nahi.

Why questions

Bandwidth-bound job ke liye ALU faster karna "kuch nahi" kyun karta?

Kyunki total time

N / B

movement term dominate karta hai; tiny

C / R

compute term (

R

ALU throughput hai) ko speed up karna bottleneck ko chhuuta nahi — movement ko khud attack karna hoga.

PIM specifically neural networks aur graph analytics ko kyun favor karta hai?

Dono bade low-reuse data (matrix–vector multiplies, edge traversals) par data movement se dominate hote hain jisme per-byte compute modest hota hai, yaani low arithmetic intensity

I = C / N

— exactly wahin jahan movement kam karna faayda deta hai. Dekho Neural Network Accelerators.

Internal DRAM bandwidth external bus se itni zyada kyun hoti hai?

DRAM row activation ek baar mein thousands of bits read karta hai, aur

k

banks parallel stream karte hain, lekin external bus

B

ek narrow shared pin-limited channel hai — PIM wide internal data par compute karta hai us se pehle ki woh narrow bus se funnel ho. Dekho DRAM Organization (banks, rows).

Crossbar mein ek shared row wire naturally summation kyun karti hai?

Kirchhoff's current law: ek common node par aane wali currents add hoti hain. Har cell

G_{ij} V_{j}

row

i

par inject karti hai, toh wire physically unhe

\sum_{j} G_{ij} V_{j}

mein sum kar deti hai — ek dot product free mein.

Roofline Model PIM use karne ka decision lene mein kyun help karta hai?

Yeh performance ko arithmetic intensity

I = C / N

ke against plot karta hai, bandwidth ceiling (sloped,

B

se set) vs compute ceiling (flat,

R

se set) dikhata hai; low-intensity workloads ridge ke left mein bandwidth roof ke neeche hote hain, aur yahi woh roof hai jise PIM uthata hai.

Streaming data ke liye memory wall fix karne ke liye temporal locality kyun kaafi nahi?

Caches sirf tab help karti hain jab data eviction se pehle reuse ho; streamed data ek baar hi touch hota hai, toh use kabhi fayda nahi milta — PIM reuse par rely karne ki jagah source par compute karke help karta hai.

Edge cases

Jab arithmetic intensity bahut high ho (jaise dense matmul,

I \approx 167

) toh PIM ka benefit kya hoga?

Workload compute-bound ho jaata hai (ridge ke right); movement par almost koi waqt nahi lagta, toh PIM ka speedup 1× ke kareeb pahunch jaata hai — CPU/GPU caches yahan pehle se jeet jaati hain.

Jab

k = 1

ho (sirf ek bank, koi internal parallelism nahi) toh kya hoga?

Tab

B_{in t} = k \cdot B_{bank} = B_{bank}

aur speedup

B_{in t} / B

mein exploit karne ke liye koi parallelism nahi; PIM roughly baseline pe degenerate ho jaata hai — jeet aati hai kai banks ek saath kaam karne se.

Agar ek workload purely serial control logic hai jisme almost koi data nahi (tiny

N

N / B

negligible hai, toh PIM ke paas bachane ke liye kuch hai hi nahi; yeh na bandwidth-bound hai na data problem — ek fast general CPU sahi tool hai.

Jab required numerical precision bahut high ho toh PUM crossbar result ka kya hoga?

G_{ij}

mein analog drift, device variation, aur ADC quantization ek accuracy floor set karte hain; kuch bits effective precision se aage crossbar requirement meet nahi kar sakta, toh digital compute ki zaroorat padti hai.

Infinitely many banks add karne par (

k \to \infty

) lekin ek fixed serial compute fraction ke saath limiting speedup kya hoga?

Amdahl's Law se yeh

1/ (serial fraction)

pe converge karta hai, infinity pe nahi — un-parallelizable

C / R

term aur coordination overhead ek hard ceiling banate hain.

Jab boundary intensity par workload clearly na bandwidth-bound ho na compute-bound, tab kya hota hai?

Tum Roofline Model ke ridge point ke paas hote ho; PIM partial win deta hai, aur decision raw speedup ki jagah secondary costs (energy, area, programmability) par depend karta hai.