Not long ago, specifying memory and storage for a video security system was straightforward. Calculate your camera count, multiply by bitrate, add retention days and size storage accordingly. Done.

That model no longer works. The growing adoption of AI inference at the edge has transformed cameras and edge gateways from recording devices into active computing platforms that run detection models, process multiple streams simultaneously, generate metadata and make real-time decisions. The workload has changed entirely, and the design assumptions have to change with it.

From passive recording to active analytics

Traditional video surveillance operated on a simple write-and-retrieve model. Video data was ingested into storage; humans reviewed it later. Memory requirements were modest. Storage was essentially a buffer. Memory and storage considerations were minimal or not a priority.

AI-enabled systems work differently. A modern edge camera or gateway isn't just capturing video; it's continuously running inference on that video, often across multiple streams at once. It may be executing person detection, license plate recognition and behavioral analytics simultaneously, each model competing for memory bandwidth and processing resources. Metadata from those inferences has to be written alongside the raw video. Models themselves must be stored and periodically updated.

The result is a fundamentally different concurrency profile. It's not about how much data you're storing anymore. It's about how many things are happening at the same time and whether your memory and storage architecture can consume them without introducing latency or degrading inference accuracy. 

Edge DRAM: Bandwidth is the critical variable

When system architects think about DRAM in edge AI video applications, their instinct is to think in terms of capacity: How many gigabytes do I need? That's the wrong starting question. Bandwidth is what determines whether real-time inference is actually possible.

A single AI-enabled camera processing two or three streams at moderate resolution may require 4GB or more of DRAM. Add sensor fusion, combining video with audio, thermal, or environmental data and requirements climb further. But it's the bandwidth, not the raw capacity, that governs inference latency and responsiveness. A system with adequate capacity but insufficient bandwidth will bottleneck on inference, with real consequences in time-sensitive or safety-critical applications.

LPDDR5 (low-power double data rate fifth generation), the high-bandwidth, low-power memory standard common in advanced mobile and embedded devices, has emerged as the practical standard for edge AI video deployments precisely because it addresses both sides of this equation. It delivers on bandwidth for demanding AI workloads while maintaining the low power consumption that compact camera housings and edge gateways demand. Thermal constraints are a real design consideration in always-on deployments; memory that runs hot creates reliability problems over time in devices that never go offline.

Storage: Four workloads running at once

Storage in AI-enabled video security systems is more complex than it appears, because there are actually four distinct workloads running concurrently, each with very different characteristics.

• Continuous video recording
  A single HD camera stream can consume several gigabytes per hour; multi-camera deployments with higher resolutions can significantly increase total storage requirements, depending on resolution and retention requirements.
  

• AI metadata and event logging
  Every inference generates structured data, including detected objects, classifications, timestamps and confidence scores that have to be written and retained alongside the video. This is a high-frequency, sustained write workload.
  

• Model hosting
  AI models range from lightweight classifiers of 50MB or less to complex multi-task models of several hundred megabytes. Edge systems often host multiple models for different detection scenarios and storage has to be immediately accessible with low read latency for inference to function correctly.
  

• Model updates
  As AI capabilities improve, models are updated in the field. That means write cycles that go well beyond simple data logging, a workload that many storage specifications don't adequately account for. 

Taken together, these workloads demand NAND flash storage (the technology inside SSDs, e.MMC and microSD cards) engineered specifically for sustained endurance. Consumer-grade storage isn't designed for this kind of continuous, mixed-workload operation. Industrial-grade storage, rated for wide temperature ranges and sustained write cycles delivers the ruggedness, consistency and quality that systems intended to run reliably over multi-year deployments in real-world environments demand.

Designing for what's coming, not just what's here

The cameras being deployed today will be running more sophisticated AI models two or three years from now. That means future-readiness has to be a design input, not an afterthought.

Smart design starts with thinking in workloads, not just capacities. That means stress-testing memory and storage specifications against peak concurrency rather than average load, and factoring in endurance ratings and life cycle support for deployments in harsh or uncontrolled environments. It also means planning for hybrid architectures, where lightweight analytics run locally and deeper processing is offloaded to the cloud when connectivity allows. That kind of distributed model still places significant demands on local storage, even when the cloud handles much of the work.

Micron has built its industrial-grade memory and storage portfolio around these challenges, engineering solutions that deliver reliability, ruggedness, quality and performance for always-on AI video systems. As edge cameras continue to evolve into intelligent analytics platforms, the memory and storage foundation underneath them becomes a critical design decision, not an afterthought. Getting it right is where intelligent system design begins.