Evidence
Energy Whitepaper
Executive Summary
As organizations deploy AI-driven Agent and Actor workloads at scale, traditional computing infrastructure struggles with cost, power, and efficiency constraints. PointOfData.ai’s Actor Infrastructure Platform (AIP) was purpose-built for low-resource environments — originally designed for ARM-based single-board computers, IoT devices, and sensors — yet delivers server-class performance with a fraction of the energy demand.
Key findings from measured benchmarks:
- Neural Memory DB on energy-optimized devices (e.g., Orange Pi) can achieve the same performance at up to 50% of the cost while using up to 75% less energy.
- Energy efficiency scales across compute environments because AIP’s architecture eliminates multiple software layers, uses an optimized C implementation, and incorporates highly efficient storage and retrieval mechanisms.
Introduction
In today’s computing landscape, energy efficiency and performance are often seen as opposing goals. AIP demonstrates that both can be achieved simultaneously. Originally developed to operate on low-resource hardware — including smartphones, smart sensors, and single-board computers used for monitoring and process control — AIP was engineered with careful attention to programming languages and code design to maximize efficiency.
Over time, AIP has consistently delivered high-speed performance on minimal hardware, including widely used ARM-based SBCs such as the Raspberry Pi and Orange Pi, as well as RISC-V modules. These low-power devices draw only a few watts, yet AIP performs at levels comparable to more complex software running on traditional servers and desktops. When deployed on standard computing hardware, AIP consumes significantly less energy than comparable software solutions.
Test Conditions
Measurements for the AIP Neural Memory database were conducted under the following conditions:
- Activity logging level: Maximum
- CPU: ARM Cortex A-76, single-core operation (no threading) at 2.3 GHz
- Storage: PCIe-attached NVMe, 2 TB Western Digital “Black”
- Operating system: Linux 22.04
Assumptions applied in calculating energy usage:
- Measured CPU time accurately reflects the duration of operations.
- A single A-76 CPU core consumes 750 milliwatts/sec (0.75 joules) at full load.
- Reported timings exclude the time required to write timing logs to storage.
- Each event object includes ten keyword facets and four location dimensions.
- The database under test contains 1,000 event objects.
Comparison Methodology
To quantify energy efficiency, AIP’s Neural Memory database engine was compared against four common SQL/noSQL databases using atomic database operations. Energy usage, measured in joules (watt-seconds), was derived from CPU power consumption and operation execution time. For the SQL/noSQL databases, energy data was sourced from a research study utilizing DBJoules.
Two datasets were used for testing:
- Netflix Users Database — synthetic dataset of 25,000 users with columns including User ID, Name, Age, Country, Subscription Type, Total Watch Time, Favorite Genre, and Last Login.
- SMS Collection — 5,574 messages with User ID and Message Text columns.
Test systems:
- System 1: AMD Ryzen 5 3500U CPU, 2.10 GHz, 4 cores, 8 GB RAM
- System 2: Intel Core i5-1135G7 CPU, 2.40 GHz, 4 cores, 8 GB RAM
- AIP System: Orange Pi 5 Plus, 4 ARM A-55 cores at 800 MHz, 4 ARM A-76 cores at 2.3 GHz, 32 GB RAM
Equivalent AIP operations require at least three tables with associated indexes in traditional SQL/noSQL engines, making the database measurements best-case comparisons.
Energy Usage Results
SQL/noSQL Databases
On the Netflix Users dataset, INSERT operations ranged from 0.81–1.77 joules and SELECT operations from 0.73–1.79 joules across MySQL, Postgres, MongoDB, and Couchbase on the two test systems. On the SMS Collection, INSERT ranged from 0.66–2.39 joules and SELECT from 0.63–2.61 joules.
AIP System
| Operation | Avg CPU Time (sec) | Joules |
|---|---|---|
| Store Event and All Facets | 0.0069 | 0.0052 |
| Search for Events, Retrieving All Facets | 0.0240 | 0.0180 |
Comparative Analysis
When comparing AIP to the best-performing SQL/noSQL engines:
| Dataset | Operation | Best DB Joules | AIP Joules | AIP % of Best DB |
|---|---|---|---|---|
| Netflix | INSERT/STORE | 0.8100 | 0.0052 | 0.64% |
| Netflix | SELECT/FIND | 0.7300 | 0.0180 | 2.46% |
| SMS | INSERT/STORE | 0.6600 | 0.0052 | 0.78% |
| SMS | SELECT/FIND | 0.6300 | 0.0180 | 2.85% |
Adjusting for CPU power efficiency relative to the Core i5 (approximately 2.8× the power of the ARM A-76 at similar clock speeds), AIP would consume at most 1.79% of the energy for inserts and 8.0% for searches compared to conventional engines. Considering that each event object in AIP effectively replaces three SQL tables, the adjusted figures drop to 0.7% for insertions and 2.3% for searches — confirming energy savings of at least 75%, and in some cases, up to 97%.
Observations
While AIP demonstrates exceptional energy efficiency, retrieval operations are slightly slower than insertions, unlike conventional SQL/noSQL databases. Optimizing search performance could further enhance AIP’s efficiency without compromising its low-power advantages.
Conclusion
The evolution of computing toward more sustainable, efficient, and intelligent systems demands a fundamental shift in how software and hardware interact. AIP and its Neural Memory Database embody this shift by demonstrating that high performance does not have to come at high energy cost.
Testing has shown that AIP can deliver equivalent or superior performance at up to 50% lower cost and up to 75% less energy consumption when deployed on energy-optimized devices like the Orange Pi. Its modular, actor-based design allows workloads to be intelligently distributed across heterogeneous environments.
In an era where energy consumption is as critical a metric as processing speed, PointOfData.ai’s AIP stands as a proof point that efficiency is the new frontier of performance.
