Protecting AI Data: Why GPUs Change the Landscape - SupremeRAID™ Ultra and InnoGrit N3X Breakthrough Parity Check RAID Performance for Massive Parallel AI I/O

Breakthroughparity RAID performancefor massive parallel AI I/O withSupremeRAID™Ultra and InnoGrit N3X April2026 Table of Content EXECUTIVE SUMMARY................................................................................................................................................2 THE HARDWARE ADVANTAGE: SUPREMERAID™2.0, NVIDIA RTX 2000E ADA, AND INNOGRIT N3X.............2 WHAT AI WORKLOADS DEMAND FROM STORAGE................................................................................................3 PERFORMANCE COMPARISONS...............................................................................................................................4 CONCLUSION.............................................................................................................................................................10 APPENDIX....................................................................................................................................................................11 Protecting Your Data in the AI Erawith SupremeRAID™and InnoGrit N3X|April2026 Executive Summary AI workloads are massive, highly parallel, and unforgiving. They generate intense bursts of small-blockand mixed I/O patterns across thousands of threads, queues, and datasets. At scale, this pressure turns This joint whitepaper demonstrates how SupremeRAID™2.0 redefines parity RAID performance for the AIera. By pairing 24x InnoGrit N3X SLC NVMe drives with the SupremeRAID™Ultra (powered by a 50WNVIDIA RTX 2000E Ada), we eliminate the traditional trade-off between data protection and performance. The results focus on the metrics that define AI infrastructure efficiency. SupremeRAID™deliversmillionsof random write operations per second with parity RAIDin optimal mode and sustainsunmatchedthroughput in degraded mode—scenarios where traditional software RAID often collapses. The result is The Hardware Advantage: SupremeRAID™2.0,NVIDIA RTX 2000EAda, and InnoGrit N3X This test platform represents the ideal architecture for modern AI servers, combining next-generation The RAID engine is theSupremeRAID™Ultra, powered by theNVIDIA RTX 2000E AdaGeneration GPU.Thissingle-slot, low-profileaccelerator operates within a highly efficient50Wpower envelope. Crucially, itis driven by theSupremeRAID™2.0 driver, which features a completelyredesigned I/O offload engine. Protecting Your Data in the AI Erawith SupremeRAID™and InnoGrit N3X|April2026 The storage media consists of24x InnoGrit N3X NVMe SSDs,powered byKIOXIA XL-FLASH™technology.These drives bridge the gap between DRAM and standard NAND, utilizingSLC (Single-Level Cell)architecture to ensureextreme low latencyand deterministic performance. By pairing theSupremeRAID™2.0 engine with the N3X's raw speed, we achieve a storage subsystem capable ofsaturating the PCIe bus with parity protection enabled, preventing the "tail latency" spikes that can stall Note:Official support for InnoGrit N3X drives will be included in the upcoming SupremeRAID™2.0 driverminor release. What AI Workloads Demand From Storage AI data platforms stress storage systems in ways that traditional enterprise applications do not. The 1.High parallel readsduring training and dataset shuffling.2.Write-heavy burstsduring checkpointing and logging.3.Continuous metadata activityfrom distributed data services. Parity RAID (RAID5/6) is attractive because it offers capacity efficiency at scale. However, the challengehas always been keeping parity RAID fast under heavy random writes and maintaining performanceduringdegraded mode(when a drive fails and reconstruction enters the data path). SupremeRAID™2.0 Performance Comparisons Test Descriptions The results below compareLinux MD (mdadm)and theSupremeRAID™2.0 Linux driveron an identical24-drive NVMe configuration. Each subsection presents the data followed by an analysis of its impact onAI workloads. Testing Environment oCPU:AMD EPYC 9755 128-Core Processor × 2oMemory:32 GB DDR5-6400 RDIMM × 24oGPU RAID Accelerator:SupremeRAID™Ultra (NVIDIA RTX 2000E Ada), single slot, low •Software oOS:Ubuntu 24.04.2 LTSoKernel:6.8.0-62-genericoRAID Implementations:▪Linux MD (mdadm) v4.3 •Configuration oOne RAID group with 24 physical drives (RAID5 and RAID6)oOptimal:All drives healthy 1.4K Random Read Performance Observation:Random read performance is essential for AI data loading and shuffling. While Linux MDperforms well in optimal states, its performance plummets by over 99% in degraded mode due to CPUbottlenecks during reconstruction. SupremeRAID™2.0, utilizing the GPU to handle I/O, sustains massive 2.4K Random Write Performance Observation:This is the most critical metric for metadata updates and checkpointing. Traditional RAIDcollapses under the "write hole" penalty, delivering sub-million IOPS. SupremeRAID™2.0 revolutionizesthis workload, delivering over6.4 million IOPSin optimal mode and sustaining nearly identical Protecting Your Data in