Editorial illustration for AMD's Helios Rack Uses Liquid Cooling to Capture 89% of Rack Heat
AMD Helios Rack Captures 89% Heat With Liquid Cooling
AMD is betting that the next round of the AI infrastructure race gets won or lost at the rack level, not the chip level. The company used its latest architecture reveal to introduce the CDNA 5 architecture, the Instinct MI455X GPU, and a rackscale system called Helios that ties 72 GPUs together with EPYC "Venice" processors and Pensando networking. The headline engineering detail: Helios uses liquid cooling designed to capture 89% of the heat generated inside the rack, a figure AMD points to as evidence that thermal management, not just raw silicon speed, has become the bottleneck for training and running large models.
That framing matters because foundation models keep growing, context windows keep expanding, and the old approach of stacking faster individual GPU servers is running out of road. AMD's pitch is that compute, memory, networking, power, cooling and software need to be engineered together as one open, standards-based platform rather than assembled piecemeal. The ROCm software stack is meant to be the glue holding that system together, from the silicon up through the rack.
Earlier AI systems were typically built by connecting servers containing eight GPUs. AMD Helios shifts the architectural boundary from the server to the rack.
Why this matters
For anyone building AI infrastructure at scale, the cooling math matters as much as the silicon. AMD says its Helios rack pulls 89% of heat out through direct liquid cooling before it hits the room, with leak detection built into the rack rather than bolted on later. If that number holds up outside AMD's own lab, it changes how data center operators think about power density and floor space when they're trying to cram 72 GPUs per rack.
But this is AMD's figure, from AMD's blog, ahead of any third-party benchmarking or real deployment data. We've seen vendors round up thermal claims before. The bigger open question is whether ROCm, AMD's software layer, can actually keep pace with CDNA 5 and Helios on the training and inference workloads researchers care about, since hardware specs mean little if the software stack lags Nvidia's CUDA ecosystem.
Worth watching: independent thermal testing, actual ROCm performance benchmarks against MI455X, and whether hyperscalers commit real capacity to Helios racks rather than just evaluating them.
Related Reading
Open Source D&B rebuilds 642 million‑business database after AI agents hit limits
Open Source OpenAI's GPT Transcribe Cuts Error Rate to 3.31%, Improving on GPT-4o
Open Source Study: 6.7% of Deepfake Requests on Hugging Face Targeted Children
From Archives Cursor Open-Sources Mixture-of-Kittens, a Megakernel for MoE Training
From Archives White House tech official calls Chinese AI model theft "unacceptable
Common Questions Answered
What is the key innovation in AMD's Helios rack system for AI infrastructure?
AMD's Helios rack uses liquid cooling technology designed to capture 89% of the heat generated inside the rack, which is a significant advancement in thermal management for AI systems. This direct liquid cooling approach includes built-in leak detection within the rack itself, eliminating the need for external cooling solutions to be added later.
How does the architectural approach of Helios differ from earlier AI systems?
Earlier AI systems were typically built by connecting individual servers containing eight GPUs each, but AMD Helios shifts the architectural boundary from the server level to the entire rack level. This rackscale approach ties 72 GPUs together with EPYC Venice processors and Pensando networking in a single integrated system.
What components are included in the AMD Helios rackscale solution?
The Helios system integrates 72 CDNA 5 architecture Instinct MI455X GPUs, EPYC Venice processors, and Pensando networking components all within a single rack. These components work together with the liquid cooling system to create a cohesive AI infrastructure solution designed for maximum performance and thermal efficiency.
Why does the cooling efficiency of Helios matter for data center operators?
If AMD's 89% heat capture figure holds up in real-world deployments, it significantly changes how data center operators approach power density and floor space allocation when deploying 72 GPUs per rack. Efficient cooling directly impacts operational costs and the ability to maximize computational density in existing data center facilities.
Further Reading
- AMD attacks the rack with Helios systems that rival Nvidia's - The Register
- Schneider Electric, AMD unveil Helios AI rack design for hyperscale DCs - Economic Times Data Centers
- Motivair Cooling Technology Enables the AMD Helios Reference Design - Motivair
- AMD Helios™ - AMD
- The 246kW rack enters the design brief - DataCentral
