人工智能对存储的影响

AI Impact On Storage Rory Bolt, Senior FellowKIOXIA America, Inc. AITrend and Challenges ata CenterNAND Bit Demand by Detailed Workload •Focus : TrainingInference (Token/sec and Token/$) ►Agentic AI, Edge AI, and Physical AI 4 Key Directions for SSD Products Generative AI evolution has created and accelerated 4 direction for SSDs NVIDIAStorage-Next:GPU Memory Extension by NVMe SSDs •Addresses HBM expansion limitations& high costs •NVIDIA Storage-Next •GPU-initiated I/O (Software Stack: NVIDIA SCADA) •NVMe/PCIe®•Fine Grain I/O size (512 byte) EmergingAIUse-Case:Near-GPU Caching •Addresses inefficiency ofsmall data accesses oververy high-speed networks •Large, efficient transfers fromdata lake to load cache•Small reads serviced from local SSD•CPU-initiated I/O Another Topic: NVIDIA ICMS (Inference Context Memory Storage Platform) •ICMS: Storage Platform forKV cache•PCIe® Gen5/Gen6 TLCNVMeSSDs will be utilized inthe storage box High Level Requirements For SSDs 512B random read optimization High queue depthsLiquid coolingMulti-initiator accessLarger capacities 512B Random Read Optimization New ECC layoutsNot necessarily tied to IU size Hypothetical case: 25M IOPS = 40nS per I/OiftRead= 45uSec must overlap 1125 parallel IOs iftRead= 25uSec must overlap 625 parallel IOs 512B Random Read Optimization Increased Endurance Workloads appear fundamentally differentMotivation forpSLCandpMLC Overprovisioning as part of the solutionCaches vs long term storage 3 DWPD to 100 DWPD High Queue Depths Scheduling impactHead of line blockingOverall latency Liquid Cooling Data center efficiency Standardization of design? Multi-initiator Access Direct access to filesystem dataLeases on LBA ranges Mapping leases to initiatorsFast path enforcement Fragmentation? Data protection? Larger Capacities Data growth trends pSLCandpMLC SKU minimization and inventory KIOXIA develops SSDs that comply with NVIDIA Storage-NextSuper High IOPS SSD–KIOXIA GP Series Super High IOPS Emulator Testing (Up to 100+ MIOPS)Software Stack Enablement, System Study with KIOXIA GPSeriesPerformance Subject to change without notice. Product image may be different than actual product.PCIe is a registered trademark of PCI SIG High IOPS Emulator Phase 1: Maximum performancedone inAug 2025 Prerequisites: •GH/GB system•X86 based GPU system won’t generate 100M IOPSdue to narrow bandwidth between GPU and CPU Phase 2: Model device performance done in Sept 2025 ►Telemetry •IO counters to track IO statistics Phase 3: Running on SCADA •How apps behave under SCADA►Device experiments Emulator Demo (effect of higher latency profile) KIOIXA Super High IOPS Emulator–NVIDIA GTC Booth Demo Challenges of Increasing PCIe Bandwidth:High Bandwidth vs. Shorter Interconnect Lengths Thank You