In brief
- Compal showcased an HGX Rubin NVL8-based direct liquid-cooled rack solution at SC26 in Chicago.
- The HGX Rubin NVL8 rack can support up to 72 NVIDIA Rubin GPUs across as many as nine servers.
- Compal showed the NVIDIA Vera CPU Compute Tray and its rack-scale system integration capabilities for these processors.
- The Vera CPU rack solution can accommodate up to 32 trays, for a total of 256 NVIDIA Vera CPUs.
Compal Electronics showcased its liquid-cooled rack solution based on NVIDIA HGX Rubin NVL8 at Supercomputing 2026 (SC26), held at McCormick Place in Chicago on 18 November. At the event, which ran from 17 to 19 November, the company also presented the NVIDIA Vera CPU Compute Tray and its system integration capabilities for these processors. Compal’s announcement focuses on bringing compute resources together with power and cooling requirements in high-density AI infrastructure.
What capacity is the HGX Rubin NVL8 rack designed to support?
Compal’s HGX Rubin NVL8-based solution can bring together up to nine OG231-2-L1 direct liquid-cooled servers in a single rack. Together, these servers offer capacity for up to 72 NVIDIA Rubin GPUs. Each server features eight Rubin GPUs and NVIDIA NVLink 6 technology.
In its published specifications, NVIDIA lists NVFP4 inference performance of up to 400 petaflops for an eight-GPU HGX Rubin NVL8 system, with total NVIDIA NVLink 6 Switch bandwidth of up to 24 TB/s. NVIDIA projects up to 10 times higher token throughput than HGX B200 for certain agentic AI inference workloads. This performance comparison depends on the conditions specified by NVIDIA and is not presented as a result that applies to all applications.
Alongside direct liquid cooling, the rack architecture includes rack-level power infrastructure, centralized system management and leak protection mechanisms. The solution therefore treats not only GPU capacity but also the power and thermal requirements of high-density systems as part of the design.
How is the Vera CPU solution configured?
At SC26, Compal presented the NVIDIA Vera CPU Compute Tray and highlighted engineering work spanning module design and system integration through to rack-scale infrastructure. The company’s OX130-8N-L1 Vera CPU rack solution can support up to 32 compute trays, for a total of 256 NVIDIA Vera CPUs. Each tray contains eight Vera CPUs and one NVIDIA BlueField-4 DPU for networking and infrastructure acceleration.
According to NVIDIA’s technical information, a Vera CPU rack can support more than 22,500 concurrent CPU execution environments. NVIDIA also reports up to 1.8 times the performance and up to four times the energy efficiency, under the stated comparison conditions, for Vera CPUs versus an x86 CPU baseline of its own choosing in certain agentic AI and simulation workloads. These are vendor-reported results for specific workloads and comparison conditions.
What does this mean for architecture and visualization workflows?
The systems announced are aimed at large-scale AI, scientific research and high-performance computing requirements rather than individual architectural design or visualization workstations. As such, they do not introduce a new architectural software feature or archviz tool. For organizations operating large-scale compute infrastructure, they demonstrate how GPU and CPU resources can be integrated with power, cooling and system management.
For architecture firms and visualization teams, any potential benefit depends on access to this kind of infrastructure and the suitability of their workloads. Compal’s announcement provides no information about compatibility with specific architectural software, licensing terms, pricing or general availability. It is therefore more accurate to view the solution as high-density compute infrastructure rather than an alternative to a desktop workstation. Any investment assessment should consider rack-level power and cooling requirements alongside GPU and CPU capacity.
Sources
1 sourceSource texts are not republished; short quotes are marked, everything else is our own summary and commentary.
Compal’s announcement is a reminder that increasing AI infrastructure capacity is not simply a matter of adding more GPUs. Bringing power, cooling and system management together in a single rack design is a meaningful approach for large-scale compute deployments, but that does not mean it is a directly usable solution for architecture and visualization firms in Turkey.
For firms, key decision-making details such as pricing, availability, licensing and software compatibility have not yet been announced. It is difficult to assess a hardware investment until this information is clear. For now, it is more appropriate to view the announcement as an example of system design for high-density AI infrastructure, not as a ready-to-use archviz product.
Frequently asked questions
How many GPUs does Compal’s HGX Rubin NVL8 rack support?
The rack can support up to 72 NVIDIA Rubin GPUs across as many as nine servers. Each server has eight GPUs.
How many processors can Compal’s Vera CPU rack solution accommodate?
The OX130-8N-L1 solution can support up to 32 compute trays, for a total of 256 NVIDIA Vera CPUs. Each tray contains eight CPUs and one NVIDIA BlueField-4 DPU.
Have pricing and a release date been announced for Compal’s solutions?
The source does not specify pricing or a general availability date. Compal showcased the systems at SC26.



