Bilingual

45°C cooling design cuts data center water use to near zero
a month ago
- NVIDIA's Rubin AI infrastructure uses 100% liquid cooling at up to 45°C, eliminating fans and reducing energy and water use in data centers.
- Liquid cooling can cut cooling energy costs by 4% per degree increase, saving millions annually in hyperscale facilities and reducing water consumption by up to 100%.
- The system operates in a closed loop, enabling chiller-less operation in favorable climates and allowing waste heat recovery for nearby buildings.
- Fully liquid-cooled servers enable higher rack density, reducing space and noise compared to traditional air-cooled systems.
- NVIDIA's design simplifies cooling with single inlet/outlet routing, making liquid cooling essential as chip power densities increase.
Study: Cerebellum helps AI ignore the ordinary for more efficient computing
15 days ago
- Conventional AI continuously processes data even when nothing changes, leading to high energy waste.
- Researchers developed a cerebellum-inspired memtransistor that ignores expected inputs and quickly detects unexpected events with significantly lower energy use.
- The device identified abnormal heart rhythms within one-fifth of a heartbeat with over 98% accuracy, using about 10,000 times fewer operations than conventional AI.
- It uses molybdenum disulfide in an asymmetric design to switch between excitatory and inhibitory modes, mimicking the cerebellum's neural balance.
- This technology enables fast, always-on AI for wearable health monitors, autonomous vehicles, robotics, and cybersecurity, reducing reliance on energy-intensive data centers.
- The study was published in Nature Communications, and future work aims to enhance the device to learn and adapt like the cerebellum over time.
Paris is harnessing the Seine to replace air-con
13 days ago
- Paris developed a district cooling network using chilled water distributed through underground pipes to buildings, reducing reliance on individual AC units.
- The system uses cold Seine river water to absorb heat from building water via heat exchangers, returning warmer water to the river with minimal ecological impact.
- Initiated in the 1990s by Engie, the network is now operated by Fraîcheur de Paris, with plans to triple its size by 2042 to cover more buildings across the city.
- District cooling improves energy efficiency and reduces urban heat island effects compared to traditional air conditioning, though it requires high initial investment.
- Similar systems exist in cities like Stockholm and Toronto, but replication depends on factors like cooling demand density, infrastructure, geography, and costs.
Do data centers only seem bad for the climate because we can see them?
10 days ago
- The climate impact of emissions is often invisible, making visible sources like data centers seem disproportionately harmful compared to hidden sources like individual microwaves.
- Data centers are more energy-efficient than distributed personal computing, similar to how a hypothetical national microwave would be more efficient than individual ovens, but visible concentration draws criticism.
- People make three mistakes: not comparing benefits per emissions, favoring 'normal' but less efficient alternatives, and focusing on visible targets while ignoring larger but invisible emissions from sectors like cement or livestock.
- Globally, data centers use 1.5% of electricity and are responsible for 0.48% of emissions, making them relatively small contributors compared to other industries that lack visible 'evil buildings.'
- Data centers are valuable tools that can help address climate change through optimization, and the priority should be greening the energy grid rather than cutting computing.
The Economic Mirage of Local LLMs
7 days ago
- The argument that self-hosting LLMs saves money is inaccurate when considering utility and hardware costs, unless prioritizing data security.
- Local inference on high-end GPUs like an RTX 4090 consumes 0.025 kWh for 18k tokens, costing ~$0.0095 in electricity, which is 2.5x to 3.5x higher than centralized API costs such as DeepSeek V4 Flash at ~$0.0038.
- Idle power consumption for local systems adds significant costs: a dual-GPU Nvidia desktop idling at 90W costs $898 over three years, making total ownership 26x higher than API costs for 1M tokens/day.
- Local hardware struggles due to memory bottlenecks during inference, with VRAM limits on GPUs (e.g., 32GB on RTX 5090) and bandwidth restrictions on unified memory (e.g., 546 GB/s on M4 Max), slowing token generation for large models.
- Key developments needed include native ternary quantization in hardware to reduce energy use and smart-sleep runtime management to minimize idle power consumption.

energy efficiency

45°C cooling design cuts data center water use to near zero

Study: Cerebellum helps AI ignore the ordinary for more efficient computing

Paris is harnessing the Seine to replace air-con

Do data centers only seem bad for the climate because we can see them?

The Economic Mirage of Local LLMs