GRAID: High-Performance GPU-Accelerated RAID on GNU/Linux
Executive Summary
This article is the result of research that I did into hardware-accelerated Redundant Array of Independent Disks (RAID) using Graphics Processing Units (GPUs). In the next few years, I want to build a new local large language model (LLM) workstation for my business, and one of the key features that I'm looking for is fast disk performance and redundancy using some form of hardware or software RAID with NVMe solid-state drives. I was inspired to write this article today because one of my favorite YouTube channels, Level1 Techs, just released a video discussing the GRAID hardware-accelerated GPU RAID solution.
GRAID Technology, founded in 2019, developed the SupremeRAID™ solution to address a major performance bottleneck in modern data centers: the inability of traditional RAID controllers to keep pace with the speed of Non-Volatile Memory express (NVMe) solid-state drives (SSDs). As an out-of-path (an architecture where a processing component operates outside the main data flow to avoid creating a bottleneck), software-defined RAID system, GRAID's technology leverages a proprietary driver to offload intensive RAID computations to a dedicated GPU. This architecture allows data to move directly from the Central Processing Unit (CPU) to the NVMe drives, bypassing the conventional bottleneck and freeing up the CPU for core application workloads.
This GPU-accelerated approach provides superior performance, with documented speeds exceeding 100 gigabytes per second (GB/s) for sequential reads and millions of Input/Output Operations Per Second (IOPS), while maintaining robust data protection. For administrators working with GNU/Linux, particularly Red Hat Enterprise Linux (RHEL), the driver installation is straightforward and can be managed via Dynamic Kernel Module Support (dkms), ensuring compatibility with kernel updates.
Compared to both hardware and software RAID solutions, GRAID offers a compelling alternative by providing higher performance, greater efficiency, and a simpler, more cost-effective design, making it an ideal choice for data-intensive environments. The company's mission is to eliminate Input/Output (I/O) bottlenecks and they have achieved this by developing a solution that is both innovative and practical.
Keywords: GRAID, SupremeRAID, GPU-accelerated RAID, NVMe SSDs, GNU/Linux, Red Hat Enterprise Linux, RHEL, Dynamic Kernel Module Support, dkms, data center, I/O bottleneck, software-defined RAID, hardware RAID, software RAID, NVIDIA, Input/Output Operations Per Second, IOPS, terabytes per second, GB/s, High-Performance Computing, HPC, AI, machine learning
```
GRAID Article Definitions & Abbreviations
├── CPU: Central Processing Unit, the main processor of a computer
├── Data Center: A facility used to house computer systems and associated components, such as telecommunications and storage systems
├── dkms: Dynamic Kernel Module Support, a framework for rebuilding kernel modules automatically when the kernel is updated
├── GB/s: Gigabytes per second, a unit of data transfer speed
├── GNU/Linux: An operating system based on the Linux kernel
├── GRAID: High-Performance GPU-Accelerated RAID solution by GRAID Technology, Inc.
├── GPU: Graphics Processing Unit, a processor designed for parallel computations
├── HPC: High-Performance Computing, the use of supercomputers to solve complex problems
├── I/O: Input/Output, the communication between a computer and external devices
├── I/O Bottleneck: A performance issue where a system's speed is limited by the rate of data transfer between components, such as the CPU and storage
├── IOPS: Input/Output Operations Per Second, a performance metric measuring the number of read/write operations a storage device can perform per second
├── local large language model: A large language model designed to run on a local machine instead of a remote server or cloud.
├── NVMe: Non-Volatile Memory express, a communication protocol for solid-state drives over a high-speed PCIe bus
├── Out-of-Path: An architecture where a processing component operates outside the main data flow to avoid creating a bottleneck
├── RAID: Redundant Array of Independent Disks, a storage technology combining multiple drives for improved performance, redundancy, or both
├── RHEL: Red Hat Enterprise Linux, a commercial Linux distribution
├── Software-Defined RAID: A RAID system where the control functions are implemented in software rather than a dedicated hardware controller
└── SSD: Solid-State Drive, a storage device using integrated circuits instead of spinning disks
```
```
GRAID: GPU-Accelerated RAID on GNU/Linux
├── Company History: GRAID Technology
│ ├── Foundational Concept (2018):
│ │ └── Engineers identify performance bottleneck of NVMe SSDs
│ │ └── When used with traditional RAID controllers
│ ├── Incorporation (2019):
│ │ └── GRAID Technology, Inc. is founded
│ │ └── To develop the first-ever GPU-based RAID solution
│ ├── Product Launch (2021):
│ │ └── SupremeRAID™ is officially introduced
│ │ └── Receiving multiple industry awards and positive reviews
│ ├── Key Milestones:
│ │ ├── Strategic partnerships with NVIDIA and other hardware vendors
│ │ ├── Expansion into global markets (APAC, EMEA)
│ │ └── Continuous software updates for new kernels and features
│ └── Mission:
│ └── To eliminate the I/O bottleneck in modern data centers
│ └── Using cutting-edge software and GPU technology
├── Technology: SupremeRAID™
│ ├── Architecture:
│ │ ├── Software-Defined RAID
│ │ └── Out-of-Path Data Flow
│ │ ├── Data bypasses CPU/Controller
│ │ └── Moves directly from CPU to NVMe SSD
│ ├── Core Component:
│ │ ├── GPU (NVIDIA, etc.)
│ │ └── Proprietary Driver
│ │ └── Installation Process on GNU/Linux (RHEL-specific)
│ │ ├── Prerequisites:
│ │ │ ├── Compatible GPU (e.g., NVIDIA)
│ │ │ ├── **Kernel Headers (kernel-devel package)**
│ │ │ │ └── **sudo dnf install kernel-devel**
│ │ │ ├── **EPEL repository for dkms**
│ │ │ │ └── **sudo dnf install epel-release**
│ │ │ └── **Dynamic Kernel Module Support (dkms)**
│ │ │ └── **sudo dnf install dkms**
│ │ ├── Step 1: Download Driver
│ │ │ └── From GRAID's website or support portal
│ │ ├── Step 2: Install via dkms
│ │ │ └── **sudo dkms install <graid_module_name>/<version>**
│ │ │ └── **Ensures driver rebuilds automatically with kernel updates**
│ │ ├── Step 3: Verify and Load Module
│ │ │ └── **lsmod | grep graid**
│ │ │ └── **modinfo graid**
│ │ └── Step 4: Configure
│ │ └── **Command-line utility: graid**
│ │ ├── **graid create raid10 --level 10 --name my_array /dev/nvme0n1 /dev/nvme1n1...**
│ │ └── **graid status**
│ │ └── **graid show --array my_array**
│ │ └── **Configuration file: /etc/graid/config.conf**
│ │ └── **Device Naming: /dev/graid/my_array**
│ └── Processing Method:
│ └── Offloads RAID Computations (parity, striping)
│ └── To GPU's parallel processing cores
│ └── Frees up CPU resources
├── Performance Metrics & Advantages
│ ├── Speed:
│ │ ├── Up to 100+ GB/s sequential reads
│ │ ├── Millions of IOPS
│ │ └── Unlocks full NVMe drive potential
│ ├── Efficiency:
│ │ ├── Minimal CPU utilization
│ │ └── Lower power consumption than comparable hardware RAID
│ ├── Scalability:
│ │ └── Linear performance increase with more drives
│ └── Data Protection:
│ ├── High-performance RAID 5/6
│ └── RAID 0/1/10
├── Target Environment
│ ├── Operating System:
│ │ └── **Red Hat Enterprise Linux (RHEL)**, CentOS Stream, Fedora
│ ├── Hardware:
│ │ └── NVMe SSDs (PCIe Gen 3/4/5)
│ ├── Use Cases:
│ │ ├── Enterprise & Data Centers
│ │ ├── High-Performance Computing (HPC)
│ │ ├── AI / Machine Learning
│ │ ├── Video Editing & Production
│ │ └── Big Data Analytics
└── Competitive Comparison
├── vs. Hardware RAID:
│ ├── No dedicated ASIC/controller bottleneck
│ └── Lower cost and simpler design
└── vs. Software RAID (mdadm):
├── Superior performance (no CPU overhead)
└── Faster rebuild times
```
References:
- Level1 Techs. "Is Hardware Raid TRULY Dead? SupremeRAID SR-1010, Reviewed". Retrieved 20250922 from https://www.youtube.com/watch?v=HCtK52_q91A
- Linus Tech Tips. "I predicted the future! - GRAID Graphics Accelerated Storage". Retrieved 20250922 from https://www.youtube.com/watch?v=12l326jQML8
- Graid Technology Inc. "Dell HPC: Using GPU Accelerated RAID to Maximize the Performance and Usable Capacity of NVMe Flash". Retrieved 20250922 from https://www.youtube.com/watch?v=4ouFIHef3rY