Diamond-Cooled AI Servers Are Here

The core of this technology is synthetic diamond, which has the highest known thermal conductivity of any material, dissipating heat five times faster than copper. This innovation is additive, meaning it complements existing air and liquid cooling systems to further boost thermal management. Akash Systems' "Diamond Cooling" can reduce GPU and High Bandwidth Memory (HBM) temperatures by up to 10°C (18°F). This prevents thermal throttling, leading to up to 22% more floating-point operations per second (FLOPs) per watt and a 15% improvement in token throughput in high ambient temperatures of around 120°F. The servers, manufactured by Mitac Computing, pair the diamond cooling with AMD's Instinct MI350X GPUs, which are built on a 3nm process and feature 288 GB of HBM3e memory. The system also incorporates two 5th Gen AMD EPYC CPUs and AMD Pensando Pollara 400 AI NICs. The MI350X GPU itself has a maximum power draw of 1000W, highlighting the extreme thermal challenge. This leap in thermal management is critical as rack densities in AI data centers are soaring, with some exceeding 120kW per rack, pushing traditional air-cooling methods to their limits. Inefficient cooling not only creates performance bottlenecks but also drives up operational costs, as data centers are projected to consume a significant percentage of global electricity. The technology was first developed and deployed in partnership with NASA for use in satellites, where conventional air cooling is not an option. Akash Systems, backed by venture capital firms like Khosla Ventures and Founders Fund, is now applying this space-proven technology to terrestrial data centers. For each server, the company claims the cooling technology can generate up to $1 million in incremental value over a four-year period. Akash has already received a $300 million launch order for these systems and has also delivered diamond-cooled servers equipped with NVIDIA H200 GPUs to cloud provider NxtGen AI.