0 0 TB TB Usable Capacity 可用容量
EC 2+1 erasure coding — ~122 TB usable from a 3-node NVMe cluster. EC 2+1 糾刪碼保護，三節點 NVMe 叢集約 122 TB 可用容量。
Phison AI Data Platform Infrastructure Phison AI Data Platform 基礎設施
Phison Storage Server Phison Storage Server
Phison's distributed storage platform unifies Block, File, and S3 Object access on a single NVMe cluster — eliminating storage silos and delivering enterprise-grade resilience for AI and hybrid workloads. 群聯分散式儲存平台在單一 NVMe 叢集上統一 Block、File、S3 存取，消除儲存孤島，為 AI 與混合工作負載提供企業級韌性。
0 0 GB/s GB/s Read Throughput 讀取吞吐量
100Gb network and PASCARI NVMe SSD deliver aggregate read bandwidth for AI datasets. 100Gb 網路與 PASCARI NVMe SSD 為 AI 資料集提供聚合讀取頻寬。
100 100 K K IOPS IOPS
Block, File, and S3 Object access on a single cluster — no storage silos. 單一叢集同時支援 Block、File、S3 存取，消除儲存孤島。
Core Barriers in Enterprise Storage Infrastructure 企業儲存基礎設施的核心障礙
Organizations building AI infrastructure encounter fundamental storage challenges that slow deployment and increase total cost of ownership. 建構 AI 基礎設施的組織面臨根本性儲存挑戰，拖慢部署速度並增加總擁有成本。
Different Storage Needs for Every Use Case 每種使用情境需要不同儲存
AI training, inferencing, databases, and file workloads each demand different storage protocols — Block, File, and Object — forcing organizations to maintain separate storage silos. AI 訓練、推論、資料庫與檔案工作負載各需不同儲存協定，迫使企業維護獨立的儲存孤島。
Cost Balancing a Mix of Data Tiers 多層資料成本難以平衡
Balancing hot NVMe performance storage with cold archive tiers is complex. Organizations overprovision or underutilize expensive SSD capacity due to lack of unified visibility. 在高效能 NVMe 與冷資料層間取得平衡十分複雜，缺乏統一可視性導致過度配置或浪費。
Challenges Supporting Enterprise Migration 企業遷移困難重重
Migrating from legacy SAN/NAS to modern distributed storage requires downtime windows, data migration plans, and retraining of IT staff — delaying AI projects. 從舊有 SAN/NAS 遷移至現代分散式儲存需要停機視窗、資料搬遷計畫與人員再訓練，拖延 AI 專案。
Growing IT Bottlenecks for AI Workloads AI 工作負載的 IT 瓶頸日增
As GPU clusters scale, storage bandwidth and IOPS requirements grow non-linearly. Traditional storage architectures cannot keep up with the throughput demands of LLM training. GPU 叢集擴增時，儲存頻寬與 IOPS 需求非線性增長，傳統架構無法跟上 LLM 訓練吞吐量需求。
One Rich Hardware Profile, One Rich Access Mode 單一硬體架構，單一存取模式
Vendor lock-in forces teams to use proprietary access APIs, making it impossible to serve Kubernetes CSI, NFS, and S3 workloads from a single unified storage platform. 供應商鎖定迫使團隊使用專有 API，無法從單一統一儲存平台同時服務 K8s CSI、NFS 與 S3 工作負載。
Broad Data Backup and Disaster Recovery 資料備份與災難復原範圍廣泛
Enterprise storage must provide snapshot, replication, and EC redundancy. Legacy systems require separate appliances for each function, increasing operational overhead. 企業儲存須提供快照、複寫與 EC 冗餘，傳統系統每項功能需獨立設備，增加營運負擔。
ARCHITECTURE 平台架構
Phison AI Data Platform Storage System Architecture 群聯 AI Data Platform 儲存系統架構
- VM · K8s CSI · DB VM · K8s CSI · 資料庫
- [ Install Storage Agent ] [ 安裝 Storage Agent ]
- Linux App Linux 應用程式
- [ Install Storage Agent ] [ 安裝 Storage Agent ]
- External Clients via HTTPS Access 外部客戶端 HTTPS 存取
- [ Protocol Required Only ] [ 僅需協定支援 ]
Storage Node 1 儲存節點 1
Storage Node 2 儲存節點 2
Storage Node 3 儲存節點 3
3 node 節點 4× 15.36 TB
184 TB raw 原始 EC 2+1 (66%) ~ 122 TB 100Gb 100Gb NIC NIC ~20 GB/s ~20 GB/s Read 讀取 ~8 GB/s ~8 GB/s Write 寫入 ~500K ~500K IOPS (4K rand read) IOPS (4K 隨機讀)
Auto Rebalancing & Zero Downtime SSDs / Nodes Scale-out 自動重平衡與零停機 SSD / 節點擴充
Auto Rack / Host / SSD Failure Recovery 自動機架 / 主機 / SSD 故障復原
Web-based Resource Management Interface 網頁式資源管理介面
Unified Block / File / Object Data Access Plane 統一 Block / File / Object 資料存取平面
Snapshot / Disaster recovery / Data encryption 快照 / 災難復原 / 資料加密
ACCESS MODES 存取模式
Three Access Modes, One Unified Storage Platform 三種存取模式，一個統一儲存平台
Phison Storage Server converges Block, File, and Object storage onto a single cluster — eliminating silos without sacrificing performance or protocol fidelity. 群聯儲存伺服器將 Block、File、Object 儲存整合至單一叢集，消除孤島而不犧牲效能或協定保真度。
Block Raw Device
VM · K8s CSI · DB
Delivers raw block access with consistent low-latency IOPS for databases, virtual machines, and Kubernetes persistent volumes.
Standard File Access
Linux App · NFS · SMB
POSIX-compliant distributed file system. Supports NFS v3/v4.1 and SMB 3.x protocols for seamless integration with existing Linux and Windows applications.
S3 Object Storage
External Clients · HTTPS
AWS S3-compatible API for AI training datasets, model artifacts, and application data. Accessible from anywhere via standard HTTPS without additional agents.
TECHNOLOGY 核心技術
Core Storage Technology 核心儲存技術
The storage platform is designed from the ground up for enterprise AI workloads, combining hardware-level SSD optimization with software-defined distributed storage. 儲存平台從底層為企業 AI 工作負載設計，結合硬體級 SSD 優化與軟體定義分散式儲存。
EC 2+1 Erasure Coding Protection
Two-plus-one erasure coding tolerates simultaneous SSD and node failures with automatic rebuild. Data durability at 66% usable efficiency — superior to 3-way replication.
Automatic Failure Domain Rebalancing
Hot-add SSDs or storage nodes and the platform automatically redistributes data across the expanded cluster with zero downtime and zero manual intervention.
Unified External Plane
A single storage cluster serves Block, File, and S3 workloads simultaneously with per-volume policy controls for IOPS limits, snapshot schedules, and replication targets.
NOFFIT Cache-Tiered Caching
Intelligent flash-to-flash caching accelerates hot data access. Phison NVMe SSD firmware optimization ensures consistent latency under heavy mixed read/write workloads.
aiDAPTIVCache-Based Caching
Optional in-memory KV cache tier extends storage bandwidth for AI inference workloads, reducing GPU idle time caused by I/O-bound data loading.
Capacity Planning & SSD Management
Web-based dashboard provides real-time IOPS, throughput, latency, and capacity metrics. Predictive capacity planning alerts before utilization thresholds are breached.