iNc · In-memory computing

Technical comparison

Compare SRAM-PIM, DRAM-PIM and NAND-iNc across architecture, energy and capacity.

Multi-dimensional comparison

Dimension	SRAM-PIM	DRAM-PIM	NAND-iNc
Data Final Location	Temporary Cache	Temporary Memory	Permanent Storage (Data Destination)
Data Movement Path	SSD→DRAM→SRAM	SSD→DRAM	Zero Movement (In-place Computing)
Solves Root Problem	Local Optimization	Partial Relief	Fundamental Solution
Energy Efficiency Gain	Limited Optimization	2-4x	100x+
Storage Capacity	KB-MB Level	GB-TB Level	TB-PB Level
Cost per GB	$1000+/GB	$10/GB	$0.1/GB
Data Persistence	Volatile	Volatile	Non-volatile (Permanent)
Write Endurance	Unlimited	Unlimited	50K+ cycles (LDPC Protected)
Edge AI Suitability	Insufficient Capacity	High Power	Perfect Match (Low Power, High Capacity)
Manufacturing Process	High Precision	Standard DRAM	Standard Nand Flash (No Change)

The Difference Between In-Memory Computing Power and Traditional Computing Power

Two Completely Different Measuring Sticks

Traditional Computing: TOPS/GOPS

The CPU/GPU Measuring Stick

TOPS (Tera OPS) / GOPS (Giga OPS) essentially counts: how many 'multiply + add' digital operations can this chip perform per second.

First move data from storage to the chip
Compute on digital MAC units
Move results back to storage

TOPS/GOPS is like counting how many screws a factory worker can tighten per hour, regardless of whether the materials are on-site.

But in reality, much time is spent 'moving data' - this overhead isn't reflected in TOPS/GOPS.

In-Memory Computing Power

Turning the Warehouse into a Factory

In-memory computing does something different: instead of counting 'screw-tightening' speed, it computes directly within the storage array - no data movement needed.

Traditional: Data sits in storage, must be moved to CPU/GPU for computation

In-Memory: Compute directly in the storage array, no data movement

In-memory computing doesn't hire more workers (increase TOPS), it transforms the warehouse itself into a factory - goods don't need to be moved around, naturally faster and more energy-efficient.

Data flow comparison

Data movement versus in-place computing

Traditional architecture

Storage

DRAM

CPU / GPU

DRAM

Storage

1 / 5 · Loading data from disk to memory

Five stages, including returning results through DRAM.

NAND-iNc

Computation inside the storage array without moving data to the CPU.

One step, minimal data movement

Three fundamental differences

Different Physical Implementation

Traditional TOPS: Uses digital MAC units with clock frequencies, easy to count.

In-Memory: Uses currents and conductance in storage cells for 'analog computation' - many operations complete instantly across the entire array, no clear 'operations per second' concept.

Different Bottleneck Locations

Traditional chips: Often have high compute power, but are bottlenecked by 'memory bandwidth' and 'data movement power consumption'.

In-Memory: Almost no data movement, computes directly in storage - what's truly fast is task completion time and energy consumption, not just TOPS numbers.

Different Suitable Tasks

Traditional TOPS/GOPS: Suitable for general-purpose computing, measures 'pure compute power'.

In-Memory: Especially suited for vector search, similarity matching, embedding search - tasks with 'massive data, high movement overhead' - can do 10-100x more work with the same energy.

Simple Understanding

Traditional Power (TOPS/GOPS)

Counts 'how many digital operations per second', assuming data is already next to the processor.

In-Memory Computing Power

Doesn't emphasize operations per second, but 'how much faster and energy-efficient can we complete the same task without moving data'.

Intuitive Understanding

For many AI search tasks, an in-memory chip uses less power and time, equivalent to 10-100x 'effective compute power' of traditional chips - not a simple TOPS conversion, but overall efficiency from a new architecture.