AI Compute Center Solutions × WS7000 Storage Acceleration
A next-generation disaggregated all-flash storage acceleration platform for AI compute centers / AI factories: extreme IO performance, GPU-direct acceleration and linear scaling form a low-latency, high-throughput, highly reliable storage backbone for large-model training, inference and agent workflows.
Built for the AI compute center
WS7000 adopts a disaggregated architecture in the same lineage as NVIDIA G3 ICMS, building a dedicated compute-storage layer across GPU HBM, DRAM and Network SSD. ASIC / FPGA / DPU high-performance IO Modules schedule compute and storage independently, breaking the bottleneck of traditional tightly-coupled designs.
For long-context inference cache management and agent-workflow context sharing, it delivers storage capability on par with the NVIDIA Rubin platform — enabling up to 5x throughput and 5x energy-efficiency gains.
Eight solutions for the compute center
One disaggregated platform, an optimal answer for each compute-center workload.
Training cluster acceleration
Up to 70M IOPS, 300GB/s bandwidth and 20μs-class latency per system stably feed large GPU clusters — accelerating sample loading, gradient sync and checkpoint persistence to shorten convergence.
Inference & long-context KV cache
Ultra-low latency and high-concurrency IO with ASIC/FPGA/DPU-accelerated KV-cache management deliver low-jitter, high-throughput online serving for long contexts and concurrent sessions.
Agent workflow context sharing
A shared context pool for multi-agent, multi-session workflows on the dedicated compute-storage layer lifts context hit rate and access efficiency under heavy load.
GPU-direct storage acceleration
Full NVMe-oF / RDMA / RoCEv2 / GDS support with BlueField-3/4 DPUs lets GPUs reach storage directly, bypassing the CPU, simplifying topology and cutting data-path overhead.
Disaggregated resource pooling
Compute and storage scheduled independently and provisioned on demand, with up to 12 × 200GbE links for multi-tenant linear scaling and elastic supply.
Green efficiency & TCO
End-to-end NVMe-native design and hardware IO acceleration deliver 5x throughput and 5x efficiency, lowering per-token cost and facility power to optimize total cost of ownership.
7×24 high-availability backbone
Active-Active controllers + Cross Link interconnect + dual-port NVMe SSD + Multi-pass verification keep business uninterrupted and data intact through single-point failures.
Brownfield upgrade & product mix
WS5000 delivers integrated Ascend-ecosystem appliances; WS7000 targets extreme-performance acceleration for compute centers. Tiered together, they upgrade brownfield clusters smoothly.
Four core capabilities
Hardware acceleration + NVMe-native + GPU-direct + Active-Active HA.
Hardware IO acceleration engine
Protocol parsing, data verification, cache scheduling and multi-path forwarding offloaded to ASIC/FPGA/DPU hardware, cutting CPU usage and software latency.
End-to-end NVMe-native
Data paths built entirely on NVMe, supporting only NVMe SSD and NVMe SCM with no conversion loss — ready for a 5–10 year data-center roadmap.
GPU-direct collaboration
Full NVMe-oF / RDMA / RoCEv2 / GDS support connects to BlueField-3/4 DPUs for GPU-direct storage acceleration and faster data movement.
Active-Active high availability
Active-Active controllers sync and load-balance over Cross Link; dual-port NVMe and Multi-pass verification sustain 7×24 operation.
WS7000 specifications (vendor spec)
Blueprint specs; final figures per the factory datasheet.
| Item | Specification |
|---|---|
| Model | WS7000-2401 |
| Throughput | 300 GB/s |
| IOPS | 70 million |
| Network bandwidth | 2.4 Tbps |
| Access latency | 20 μs-class |
| Drives / type | 24 bays · NVMe U.2 SSD (single / dual-port) |
| Max drive capacity | 250 TB |
| PCIe expansion | 6 × PCIe 5.0 x16 (Gen 4/3/2/1 compatible) |
| Front-end ports | Up to 12 × 200GbE · NVMe-oF over RoCE v2 / TCP / IB |
| IO Module | FPGA / ASIC / BlueField-3 / DPU / Retimer |
| Management | IPMI / SNMP / Redfish · web UI · remote ops |
WS5000 × WS7000 positioning
Select and combine by build stage and workload profile.
| Dimension | WS5000 | WS7000 |
|---|---|---|
| Positioning | Ascend-ecosystem all-flash storage appliance | Extreme-performance acceleration for AI compute centers |
| Core scenarios | Integrated training / inference, domestic stack | Large-scale training, long-context inference, agent context |
| Architecture focus | Disaggregation + deep Ascend tuning | Disaggregation + end-to-end NVMe + GPU-direct |
| Build stage | Greenfield delivery, brownfield retrofit | Compute-center scale-out, extreme-performance upgrade |
Benchmark it on your own workload
2 live demo units are ready for immediate PoC. Let the data do the talking.