Analog Compute-in-Memory vs GPU
von Neumann bottleneck · data movement energy · inference efficiency
GPU — External memory, high data-movement cost
Analog CIM — Compute in memory, minimal movement
Data / weight tokens moving between units
⏸ Pause
↺ Reset View
Speed: 1×
GPU Data Movement
0
GPU Energy Used
0
Analog Energy Used
0
Efficiency Ratio (GPU / Analog)
1.0×
●
GPU: weights shuttle between
HBM
&
compute units
—
high energy per operation
·
●
Analog CIM: weights stay
in memory array
—
near-zero data movement