VDURA, a provider of data storage software for artificial intelligence (AI) and high-performance computing (HPC), and Phison Electronics, a supplier of NAND flash controllers and storage solutions, have demonstrated a real-time data transfer rate of 5.1 petabytes (PB) per second at Supercomputing 2025 (SC25). The demonstration, performed live at Booth 2033 with Hafþór “The Mountain” Björnsson, involved moving 136.5 PB of data in 27 seconds. This transfer used Phison Pascari enterprise solid-state drives (SSDs) and VDURA’s flash-optimized parallel storage infrastructure to reach a sustained 5.1 PB per second, a speed level previously thought unattainable in AI and HPC environments, according to the companies.
The system powering the demonstration included VDURA’s DirectFlow client, a fully parallel file system, and the VeLO metadata engine, as well as Phison Pascari D205V NVMe drives. The companies report that this architecture allows for high throughput and metadata acceleration, preventing pipeline chokepoints and providing the durability required for AI factories and GPU-accelerated clusters. Automated tiering is also included, reducing manual tuning for storage management.
In the previous year’s SC24 event, a demonstration transferred 282.6 PB, focusing on capacity. For SC25, the emphasis shifted to speed, with the new challenge showcasing real-time transfer of hundreds of petabytes. The demonstration is intended to highlight the infrastructure’s applicability to hyperscale data centers and AI-focused GPU clusters, both of which demand extreme data throughput and consistent performance.
VDURA describes its platform as combining flash-first speed with hyperscale capacity and “12-nines” durability, targeting AI and high-performance computing environments. Phison’s Pascari product line provides high-performance enterprise SSDs for data-intensive workloads in AI, cloud, and hyperscale data centers. Both companies highlight the use of metadata acceleration and automated tiering as features designed to address the demands of large-scale data centers.
Source: VDURA
