Huawei's Atlas 960 Super Pod cuts power use by over 550 kW per machine
Artificial intelligence workloads are driving rapid growth in data‑center power demand. Analysts estimate that by 2030 AI facilities could consume more than 945 terawatt‑hours each year, with roughly 80% of that energy spent moving data rather than performing calculations. The high cost and environmental impact of this inefficiency have prompted hardware designers to look for ways to streamline…
Key points
- Huawei's Atlas 960 Super Pod reduces power consumption by over 550 kW per machine using the Unified Bus Protocol.
- AI data centers may use over 945 TWh annually by 2030, with 80% of energy spent on data movement.
- Pirium architecture’s unified memory access, distributed processing, and copper‑optical cable mix boost compute‑per‑watt efficiency.
Huawei’s Pirium architecture tackles the problem with a Unified Bus Protocol that replaces multiple legacy communication standards with a single, integrated channel. The design adds unified memory access, distributes processing decisions away from a central CPU, and combines copper cables for short links with optical fiber for longer runs. These changes lower latency, reduce redundant data conversions, and improve overall energy efficiency.
In practice, the Atlas 960 Super Pod, built on Pirium, reports a reduction of more than 550 kilowatts of power per machine while delivering higher compute‑per‑watt performance. Huawei has open‑sourced the protocol and hardware specifications, inviting other chipmakers and server vendors to adopt the approach and help the AI industry meet sustainability goals.
New Atlas 960 Super Pod Cuts AI Energy Usage by 550 kW
geeky-gadgets.com · 4 October 2026
Loading the full article…
This text was published by geeky-gadgets.com and written by Julian Horsey. It is reproduced here with attribution so you can read it in full; the rights remain with the publisher. Read it at the source ↗
The headline, key points and digest above were generated by Digest AI's editorial model from the linked sources. Automated summaries can contain errors: the sources are the record. Spotted a mistake? Tell us. Published by Martin K., who runs Digest AI and handles corrections.
More in Hardware & Compute
All →- Nvidia raises Shield TV Pro price by $100 to $299.99 due to AI-driven component costs · 5 src
- Malaysia’s Data center colocation market grows 18.6% annually through 2031 · 1 src
- AM Intelligence orders 20,000 NVIDIA Rubin GPUs for India and Malaysia · 1 src
- User runs Qwen Flash Next Q4 on Mac Mini M5 with SSD streaming · 1 src
- NVIDIA launches 64 GB DGX Spark desktop AI system starting at $4,999 · 12 src
Comments
via GitHub Discussions