Auteur Sujet: Skeleton GrapheneGPU to cut Data Centre Energy Use by 44% (Lu 369 fois)

letsar · « **le:** 03 juin 2025 à 10:52:19 »

" Estonian firm Skeleton Technologies launches GrapheneGPU, a new system that reduces power requirements and aims to boost hyperscale computing performance

Skeleton Technologies has launched GrapheneGPU to target energy inefficiency in AI data centres through what the company describes as ‘a fundamental shift in infrastructure scaling’.

The Estonian energy storage firm says its new system reduces AI energy consumption by 45% and lowers power connection requirements by 44%, whilst delivering 40% more computing performance measured in FLOPS.

With a completed validation under power profiles specified by hypescalers, initial shipments of the system are planned from Skeleton's German facility in June 2025. The company will subsequently expand manufacturing to the US in the first quarter of 2026 to meet international demand.

“As AI computing and energy demands double annually, AI data centres are facing a critical bottleneck,” says Taavi Madiberk, CEO of Skeleton Technologies. “GrapheneGPU delivers up to 40% more computing with the same energy footprint, while cutting both capital and operating costs by reducing grid upgrade needs, energy waste and cooling.”

Addressing challenges in GPU power fluctuation

AI data centres currently consume much more energy than they require, according to Skeleton, with inefficiencies arising from grid capacity limitations, infrastructure constraints and intermittent outages.

As we already know, the International Energy Agency (IEA) has already projected global electricity consumption by AI data centres will reach 945 TWh by 2030, with the US accounting for approximately 240 TWh and Europe consuming over 150 TWh.

With these sustainability concerns in mind, the fluctuating power demand of GPUs inevitably presents operational challenges for data centre operators. GPUs cycle between 0% and 100% power consumption within seconds, creating sharp spikes during processing followed by idle periods.
GrapheneGPU key facts:

Grid connection reduced by up to 44%
Energy consumption cut by up to 45%
GPU performance gains of up to 40%

Data centres typically manage these fluctuations through artificial loads, typically called dummy loads, during idle phases to maintain constant power draw and prevent grid instability or equipment damage.

This approach results in energy losses of up to 45% as waste heat, providing no computational benefit. Peak power availability remains constrained by limited grid capacity, with new substations and transmission lines requiring up to seven years to construct.

To tackle these concerns, GrapheneGPU stores energy during idle periods and releases it during peak demand to smooth these power fluctuations. Skeleton Technologies says this eliminates dummy loads, reduces cooling requirements and enables more efficient GPU operation.

Likewise, the system combines peak-shaving supercapacitors with AI-driven control software, which are powered by Skeleton's patented Curved Graphene material.

How Skeleton Technologies hopes to expand production to meet demand

The company’s new system crucially maintains compatibility with existing infrastructure, enabling deployment without major facility modifications. It is also designed to reduce both capital investments and operational costs proportionally by cutting energy use and peak power demand.

Skeleton Technologies is eager to prove that data centres can achieve higher computing density without requiring additional power generation or expensive grid upgrades. Already operating manufacturing facilities in Germany and Finland and headquartered in Tallinn, Estonia, the company has raised more than €300 million (US$338.62m) in capital since inception.
Key fact

Skeleton Technologies is building the largest supercapacitor factory in the world and expects a production capacity of 30 million cells a year

Self-described as a European leader in high-power, fast-charging energy storage for AI data centres, Skeleton Technologies also focuses on the grid, transportation and industrial applications across automotive, industrial equipment, truck fleet and defence sectors.

Looking ahead, the company is eager to reduce CO2 emissions and improve energy efficiency, hoping that GrapheneGPU will offer a pathway to better managing data centre power.

Taavi adds: “Powered by our patented Curved Graphene, this is a fundamental shift in how AI infrastructure can scale – sustainably and economically.” "

Le lien de l'article : https://datacentremagazine.com/it/skeleton-graphenegpu-to-cut-data-centre-energy-use-by-44

Leon · « **Réponse #1 le:** 03 juin 2025 à 11:50:32 »

Ca semble assez fumeux comme concept. Beaucoup de blabla difficilement compréhensible, qui ressemble plus à du marketing qu'autre chose.
Genre il suffit d'ajouter un stockage par ultracap à côté de chaque GPU pour diminuer leur consommation de 40% rien que ça.
Je dois être bête, je n'ai pas compris.
Le concept de "GPU artificial load", ça existe vraiment? C'est vraiment nécessaire pour faire tourner de l'IA sur GPU? C'est la première fois que j'en entend parler...

Leon.

simon · « **Réponse #2 le:** 03 juin 2025 à 12:43:48 »

Citation de: Leon le 03 juin 2025 à 11:50:32

Le concept de "GPU artificial load", ça existe vraiment? C'est vraiment nécessaire pour faire tourner de l'IA sur GPU? C'est la première fois que j'en entend parler...

Apparament oui, j'en ai déjà entendu parler (mais je suis loin du domaine).
De ce que j´ai compris, les clusters de GPU consomment tous en même temps, et tirent une puissance assez incroyable. Lors des transitions repos<>travail, ca génère un impact de charge assez important pour faire sauter des protections. Donc ils font tourner les GPU à vide pour avoir une consommation relativement stable... si c'est vrai, c'est pitoyable.

Auteur Sujet: Skeleton GrapheneGPU to cut Data Centre Energy Use by 44% (Lu 369 fois)

letsar

Skeleton GrapheneGPU to cut Data Centre Energy Use by 44%

Leon

Skeleton GrapheneGPU to cut Data Centre Energy Use by 44%

simon

Skeleton GrapheneGPU to cut Data Centre Energy Use by 44%