OCP Global Summit showcases liquid-cooled, high-density AI rack designs
Редакция ЦОД:Главное
Exhibits at the Open Compute Project Global Summit in San Jose included high-density server racks, AI accelerators, networking components and liquid-cooling designs aimed at hyperscale data-center deployments. The most detailed systems shown were a Supermicro-Arm Open Rack Wide platform, Google’s Ironwood TPU rack and AMD’s Helios compute tray.
- Supermicro and Arm showed a direct-liquid-cooled double-wide rack rated for a maximum 200 kW energy power budget.
- The Supermicro-Arm rack integrates 168 compute nodes, 336 Arm AGI CPUs and up to 45,696 CPU cores.
- A fully configured Google Ironwood rack includes 64 liquid-cooled TPUs in 16 four-chip modules.
- Google’s Mount Diablo sidecar was presented in connection with denser racks and a future transition to 800 VDC distribution.
- AMD’s Helios tray combines four Instinct MI455X GPUs with an EPYC Venice CPU and Pensando networking technology.
The Open Compute Project Global Summit in San Jose featured infrastructure designed for high-density AI data centers, including server racks, accelerators, networking hardware, liquid-cooling components and cold-plate designs.
Supermicro and Arm displayed a liquid-cooled Open Rack Wide system using Arm AGI processors. The double-wide rack contains 168 compute nodes and 336 Arm AGI CPUs, with up to 45,696 CPU cores. Its direct-to-chip liquid-cooling design supports a maximum energy power budget of 200 kW per rack.
Google Cloud showed an Ironwood TPU rack configured with 64 liquid-cooled TPU chips across 16 four-chip compute modules. Google also presented its Mount Diablo power sidecar, intended for denser compute racks and a prospective transition to 800 VDC power distribution. The 800 VDC reference was framed as a future direction rather than an implemented deployment.
AMD displayed a liquid-cooled Helios compute tray incorporating four Instinct MI455X GPUs, an EPYC Venice CPU and Pensando networking technology. According to AMD CEO Lisa Su, each compute tray weighs approximately 160 pounds, underscoring the mechanical handling implications of increasingly integrated AI infrastructure.
Taken together, the systems illustrate competition across Arm, Intel and AMD CPU platforms as well as between specialised TPU platforms and GPU-based AI systems. They also suggest that rack-level engineering—particularly power ingress, heat rejection, internal networking and equipment handling—could become increasingly consequential for operators deploying dense AI capacity.
The exhibits point to an industry shift from discrete servers toward integrated rack-scale AI systems, where power delivery, direct liquid cooling, component weight and high-speed internal networking are central design constraints. For data-center operators, this could increase the importance of liquid-cooling readiness and, over time, support for higher-voltage DC power distribution.
- — / 19 октября 2026 г. · АнглийскийПервоисточник →
Материал подготовлен редакцией ЦОД: Главное на основе указанных первоисточников.