High-Speed Dataset Streaming
Feed datasets to your GPU cluster at maximum speed — cutting training cycles and enabling low-latency model inference.
Service
AI Workload Storage:
Optimized for Generative AI
Scalable storage tailored to build and run generative AI on Kobayashi AI Cloud.
01Features
AI Storage Features
Rapid Checkpointing
Maximize AI infrastructure goodput with high-speed shared storage: accelerate checkpoint read/write during multi-host training.
Multi-Modality Ready Storage
Store unlimited unstructured data (text, video, etc.) to streamline multi-modal AI training workflows.
02Object
Kobayashi Object Storage
Fully S3-compatible object storage for:
- Storing static data
- Serving external consumers
- High-performance data streaming for MLOps workflows
Seamlessly migrate data across storage classes (within the same service) to align with your data strategy.
Standard class
- Capacity-optimized buckets
- Ideal for large unstructured data volumes
- Cost-efficient for static data
- Unlimited scalability
Enhanced class
- Performance-optimized buckets
- Perfect for GPU data streaming & checkpointing
- Up to 2 GiB/s write throughput per GPU*
- Unlimited scalability
03Filesystem
Kobayashi Shared Filesystem
A high-speed shared filesystem built exclusively for AI workloads — delivering scalable performance for parallel AI compute and unlimited capacity scalability. It’s the go-to choice for training & inference workflows, combining cost efficiency, ease of use, and a robust feature set.
- All-flash NVMe high-performance tier
- 500+ GB/s aggregate read throughput**
- Native integration with the Kobayashi Cloud Platform
Shared Filesystem
All-flash storage node
All-flash storage node
All-flash storage node
All-flash storage node
Performance & capacity scaling
04Block
Block Network Storage: VM Boot & Workload Volumes
Block network volumes built for booting and running virtual machines. Choose from 3 tiered options — tailored to your performance, reliability, and cost requirements:
- 3x Mirrored Block Volumes
(max reliability for critical workloads) - Erasure-Coded Block Volumes
(balanced performance & cost efficiency) - Non-Replicated Block Volumes
(cost-optimized for non-critical use cases)
* Performance varies based on bucket data structure, write concurrency, and upload process configuration.
** Burst performance measured in a 254-GPU/254-CPU host cluster (during real customer multi-modal training workloads).