SciTransfer
ENERGIA/PIM · Project

High-Performance Memory Chips to Slash Data Centre Energy and Processing Costs

digitalPilotedTRL 7

Imagine if your computer's brain had to walk across town every time it needed a piece of information from its filing cabinet; that's how current servers work. This technology puts the brain directly inside the filing cabinet, so the work happens where the data lives. It stops the constant, energy-hungry commuting of data, making everything run much faster and cooler.

By the numbers
20x
faster performance
12x
more efficient
10x
cheaper
80%
energy consumption related to data movement
The business problem

What needed solving

Data centres waste 80% of their compute energy simply moving data between memory and the CPU. This creates a structural bottleneck that slows down AI and Big Data processing while driving up electricity costs.

The solution

What was built

A semiconductor-based Processing-In-Memory (PIM) chip and associated control software, including a module that fits into standard server memory slots.

Audience

Who needs this

Cloud Infrastructure ProvidersAI Hardware IntegratorsBig Data Analytics FirmsServer Manufacturers
Business applications

Who can put this to work

Cloud Computing
enterprise
Target: Hyperscale Data Centre Operator

If you are a data centre operator dealing with skyrocketing electricity bills and CPU bottlenecks — this project developed Processing-In-Memory (PIM) technology that makes servers up to 20x faster and 12x more efficient.

Artificial Intelligence
mid-size
Target: AI Model Training Provider

If you are an AI provider dealing with slow processing speeds for Big Data — this project developed PIM modules that fit into standard memory slots and are 10x cheaper than conventional memory setups.

Cybersecurity
SME
Target: Secure Data Processing Firm

If you are a security firm dealing with data leakage during transfer between CPU and memory — this project developed a new security scheme to make PIM elements secure enclaves.

Frequently asked

Quick answers

How does this affect the total cost of ownership?

PIM equipped servers have proven to be 10x cheaper for Big Data and AI applications compared to those using conventional memory.

Can this be deployed at an industrial scale?

Yes, the modules are designed to fit into standardized server memory slots and the team has worked with prime server partners like Intel, QCT, and Supermicro for integration.

What is the IP or licensing status?

Based on available project data, the technology is developed by SAS UPMEM, but specific licensing terms are not provided.

How difficult is the hardware integration?

Integration is simplified because the technology leverages existing industry protocols and fits into standard server memory slots.

What is the timeline for deployment?

The project period ran from 2022-03-01 to 2024-02-29, indicating the design and integration work is now complete.

Consortium

Who built it

The project is led by a single French SME, SAS UPMEM, which maintains 100% industry control. This lean structure allowed for direct collaboration with major hardware giants like Intel, QCT, and Supermicro, ensuring the technology is compatible with global server standards rather than remaining a theoretical academic exercise.

How to reach the team

Contact SAS UPMEM in France for PIM module integration

Next steps

Talk to the team behind this work.

Request a technical deep-dive on PIM integration for your data centre