Kertari Supercomputer

The Kertari Supercomputer is the most powerful single supercomputer complex in operation on Aron. Located on the outskirts of Charlestown in The United Federal States of Lucithia, the system forms the computational core of the broader Kertari Research Complex. Construction of the primary compute structures was completed on 6 April 2025, with phased commissioning concluding later that year.

Kertari operates as a commercial supercomputing facility. Compute time is leased to public and private entities under structured agreements, with primary long-term allocations held by the UFSL Government, Atlas Dynamics, and Aquita Technologies. The facility is powered directly by the Charlestown Nuclear Powerstation, providing a stable high-capacity energy supply suitable for sustained multi-megawatt computational loads.

Site and Structural Layout

The Kertari Research Complex occupies a secured perimeter zone beyond Charlestown’s urban boundary. The compute campus consists of twenty hardened buildings arranged in three operational sectors: the North Arc, Central Arc, and South Arc. Each building is internally designated under the K-Block schema (for example, K-01 Halden, K-02 Merrow, K-03 Talith), and contains three independent compute halls, dedicated electrical substations, and localized cooling infrastructure.

The distributed multi-building design was selected to reduce systemic failure risk, contain thermal zones, and permit maintenance without full-facility downtime. Subterranean optical and power conduits interconnect the buildings, forming a unified computational fabric.

System Architecture

Kertari comprises sixty clusters distributed across its twenty buildings. Each cluster contains 212 compute nodes, for a total of 12,720 nodes across the facility.

Each node integrates a Kreiger K5700 processor implementing the CAG64 architecture alongside a KRAM RTA 7000 Super accelerator. The Kreiger K5700 operates at a 3.6 GHz base frequency and provides 48 CPU cores (24 performance cores and 24 efficiency cores). The CPU cache hierarchy includes approximately 3 MB of aggregate L1 cache, 48 MB of L2 cache, and roughly 128 MB of shared L3 cache connected via chiplet fabric.

Main memory per node consists of 256 GB of DDR5-6400 across eight channels, delivering approximately 410 GB/s of sustained bandwidth.

The KRAM RTA 7000 Super accelerator provides approximately 30,000 parallel compute units operating at up to 2.0 GHz boost frequency. Each accelerator includes 192 GB of stacked HBM3e memory delivering roughly 3 TB/s of memory bandwidth, supported by an on-package L2 cache of approximately 64 MB. Accelerator throughput emphasizes wide SIMD and matrix execution engines rather than extreme clock frequency.

At theoretical mixed-precision peak, the complex achieves approximately 4.4 exaflops. Total aggregate DRAM bandwidth approaches 5.2 petabytes per second, while combined accelerator memory bandwidth exceeds 38 petabytes per second. Realized performance depends on workload characteristics, arithmetic intensity, synchronization frequency, and cross-building communication overhead.

As of 2027, Kertari is recognized as the highest-capacity single supercomputer complex on Aron.

Storage and Data Fabric

The facility shares a distributed parallel storage center with approximately 80 petabytes of NVMe-based capacity. To reduce cross-building traffic and minimize latency penalties, each of the twenty buildings contains an independent ~2 petabyte NVMe cache tier used for dataset staging and checkpoint buffering.

An internal optical backbone links clusters within buildings at high bandwidth and low latency, while cross-building communication traverses a dedicated subterranean fiber network. Interconnect topology and scheduling policy are central to maintaining efficiency for tightly coupled large-scale simulations.

Power and Thermal Management

Kertari is powered directly by the Charlestown Nuclear Powerstation, ensuring stable long-duration energy supply independent of municipal grid fluctuations. Each node is equipped with dual 3,000-watt high-efficiency power supplies, permitting sustained loads approaching 6 kilowatts under full computational utilization.

Thermal management is based on high-efficiency closed-loop liquid cooling at the node level, supplemented by redundant airflow systems within each compute hall. Each K-Block contains localized heat exchange and pumping systems, allowing isolation of thermal incidents and independent maintenance operations.

Software Environment

All compute nodes operate a lightweight deployment of the Neo kernel known as Neo Feather. This configuration minimizes scheduler overhead and allows direct hardware access to accelerator subsystems. Neo Feather instances at Kertari are validated for compatibility with Atlas-aligned storage interfaces and identity frameworks.

Workloads commonly employ K# for accelerator kernels, with host code written in performance-oriented compiled languages such as Aegis or Harbor. UFEL is frequently used for structured data interchange across cluster boundaries.

Operational Role

Kertari supports climate modeling, energy grid simulation, materials research, biomedical computation, and large-scale economic modeling. Allocation policies balance governmental research mandates with commercial leasing contracts.