iNc presentation

At the edge

Making real-time intelligence more reliable

Automatic braking · fault warnings · vital-sign monitoring

Faster decisions, fewer accidents and losses

Higher energy efficiency

Making AI viable on large numbers of devices

Several times to hundreds of times

Higher efficiency for power-sensitive devices

Less dependence on the cloud

Privacy and network resilience

Process data locally

Upload only the necessary results

On the training side

Making AI training more efficient

Breaking through the memory wall

Training energy reduced by orders of magnitude

AI development and environmental goals

Better aligned

At the societal level

More inclusive intelligent infrastructure

Narrowing the digital divide

Smaller cities · remote regions · traditional industries

Application scenarios

Internet of Things (IoT)

Intelligence near the sensor

Upload results only: lower bandwidth and cloud costs

Smart glasses and wearables

Local visual understanding

Navigation · assisted reading · real-time translation

Vehicles and mobility

ADAS / autonomous-driving edge inference

Lower latency and better real-time response

Robotics

On-device perception and planning

Smoother motion and greater safety

Smartphones

On-device AI acceleration

Image enhancement · speech recognition · real-time translation

PCs and tablets

Local inference of large models

Office assistants · coding assistants · creative tools

Smartwatches

Real-time health monitoring

All-day wear and continuous monitoring

These are

The uses of in-memory computing

A user asks:

How many TOPS can a computing SSD deliver?

It has no TOPS

And no GOPS

Because

TOPS = (2 × MAC × frequency) / 10¹²

MAC-unit count × operating frequency

1 MAC = 1 multiplication + 1 addition

In-memory computing works differently

Four computational processes in one step

No MAC units

No clock frequency

The TOPS formula does not apply

We can only estimate

Equivalent computing performance

Vector similarity search comparison

Traditional CPU/GPU

20 seconds / 100J

iNc solution

0.5 seconds / 2.5J

43× energy efficiency

47× lower latency

Eliminating the bandwidth bottleneck

You may not know

A harsh reality

L4 systems are equipped with

500+ TOPS

Actual utilization is only

20–40%

Data movement becomes an energy black hole

63.7%

Of power is used for DRAM ↔ NPU transfers

In-memory computing instead

Computes in the analog domain inside the array

Eliminating data movement at its source

For technical readers

Explore the whitepapers

Let us

Embrace new technology together