1.Ultra-High Performance Requirement Scenarios

The main scenario is to meet ultra-high performance read/write requirements, with primary applications as follows:

Rapid Transmission and Storage of Satellite Data in the Remote Sensing Industry

Satellites have fixed communication windows with ground stations each day. The faster the storage write speed within a given time frame, the more data can be transmitted. Current solutions rely on building large-scale distributed storage systems (spanning multiple data centers). By using all-flash storage as front-end cache combined with backend traditional storage, resource consumption (power, carbon emissions, and maintenance costs) can be significantly reduced. Additionally, high-performance storage can also be deployed in a distributed manner, greatly enhancing overall business capacity.

High-Concurrency Transactions in New Retail

During promotional periods, new retail businesses experience ultra-high concurrency in transactions. Queries and updates to database data typically employ parallel real-time computing methods, creating a strong demand for high-performance storage.
For ultra-high-concurrency orders (primarily counting), caching technology can temporarily store generated orders before gradually writing them to the database via task queues. Meanwhile, e-commerce platforms only record orders and do not perform real-time inventory checks. Therefore, database pressure (since all database data is stored in storage) can be alleviated using multi-cluster, multi-queue, and multi-caching techniques.

MPP (Massively Parallel Processing) Computing

Due to its architecture, MPP computing restricts each node to accessing only its own resources, resulting in isolated resource utilization. Since MPP typically handles large-scale data computations, the processing speed of each node directly impacts the overall computation speed for complex mathematical problems. Thus, improving single-node performance significantly enhances MPP efficiency (e.g., national supercomputing centers). In known single-node systems (including personal computers), the bottleneck is not computational resources but storage resources. Therefore, all-flash storage can effectively boost computational efficiency.

2.Business Performance Improvement

In IT operations, over 80% of issues are database-related, and database access is often the most time-consuming part of business processes. Since database data is stored in storage, the read/write efficiency of storage determines database performance, which ultimately affects end-user experience. Thus, improving storage performance directly enhances overall business performance.
Business systems typically consist of network devices, servers, middleware, databases, applications, and storage. In practice, computational resources are often abundant or even excessive, while storage systems—especially mechanical storage—have inherent performance limits due to the physical nature of HDD platters. All-flash storage, which uses SSDs, effectively overcomes these limitations, significantly improving business performance.
Moreover, all-flash storage provides notable benefits for transactional businesses (e.g., banking, e-commerce). These businesses typically involve high concurrency, placing heavy demands on database queries and writes. Since storage IOPS directly impacts database concurrency performance, the high performance of all-flash storage can substantially enhance transaction processing capabilities.

3.Localized Computing Power Unleashed

During the transition to domestic IT solutions, many localized products (e.g., servers, middleware, databases, and business systems) must replace existing ones. However, domestic hardware often lags behind x86 architectures in performance, leading to insufficient computing power, business queues, and backlogs. GP high-performance all-flash storage, leveraging RDMA and NVMe-oF technologies, enables:

• Direct storage of business data on NVMe SSDs
• Reduced CPU involvement in data storage
• More CPU resources allocated to computational tasks

This compensates for the computing power deficit in domestic hardware while providing high throughput, high concurrency, and low latency—ensuring rapid data access and improved business performance for localized deployments.

4.Ultra-High Performance Database Appliances

Most business access issues stem from data queries and retrieval. The three critical performance metrics for database appliances are storage IOPS, latency, and throughput. Traditional HDDs and mainstream SSDs still rely on SATA or SAS interfaces, limiting data transfer speeds due to controller bandwidth and storage path constraints. Additionally, CPU overhead in storage operations is high. Improving database appliance I/O, throughput, and latency traditionally requires costly hardware upgrades.
By deeply integrating database compute nodes with all-flash storage, data leverages RDMA (zero CPU overhead) to travel directly from compute nodes (via PCIe and RDMA NICs) to storage nodes. This reduces the storage path from 7 nodes (in traditional setups) to just 4 nodes, slashing latency from milliseconds to under 20 microseconds. With single-node performance reaching 72GB/s bandwidth and 16 million IOPS, the overall performance of database appliances improves by over 10x. Combined with high-density scalability, high availability, and reliability, this easily supports complex database demands in critical industries.

5.AI Training Acceleration Solution

AI applications are becoming increasingly widespread, with large-scale models like GPT-3 achieving unprecedented success. ChatGPT has already transformed multiple industries. As model sizes grow explosively, the massive data access and high-concurrency demands in training have become core challenges for AI development. At SIGMOD 2022, researchers experimentally validated the training efficiency of major models and identified storage systems as the bottleneck, specifically:

• High IOPS pressure from massive small files
• Insufficient bandwidth for large-scale data access
• CPU involvement in data access limiting computing power

To address these issues, Green Computing's self-developed NVMe-oF all-flash storage product, ForinnBase GroundPool (GP), implements ASIC-level offloading of NVMe, NVMe-oF, and RDMA protocols. It combines NVMe's high performance, zero-copy, and scalable networking with GPU direct access and high-performance shared storage. GP delivers high capacity, bandwidth, IOPS, scalability, and low latency/power consumption while supporting GPU Direct Storage and compute-storage separation architectures. This provides an efficient, flexible, and cost-effective AI training acceleration solution, offering a one-stop upgrade for AI enterprises.

These solutions significantly improve AI training efficiency, reduce waiting times, enable secure centralized data storage, and comply with space/power constraints.

6.General Use Cases

In conventional data backup and migration scenarios, traditional mechanical storage suffers from long backup times and low efficiency (e.g., a bank requiring 1.5 days for a backup cycle every 3 days due to excessive data). Ultra-performance storage dramatically improves backup speeds, enabling daily backups and ensuring data security.
For data migration, traditional speeds depend on storage write performance. Ultra-performance storage accelerates migration, ensuring faster system deployment and business readiness.