Huawei unveils 245TB AI SSDs to bypass costly HBM restrictions

Huawei has unveiled three new AI SSDs designed to reduce reliance on expensive high-bandwidth memory (HBM), addressing supply restrictions that Chinese firms face in accessing advanced memory chips. The OceanDisk series includes the industry’s largest SSD at 245TB capacity, positioning solid-state storage as a partial alternative to costly HBM in AI workloads.

What you should know: Huawei’s new OceanDisk series targets the “memory wall” and “capacity wall” problems that currently bottleneck AI training and inference performance.

• The OceanDisk EX 560 delivers extreme performance with 1,500K IOPS write speeds, sub-7µs latency, and can increase fine-tunable model parameters on a single machine sixfold for LLM training.
• The SP 560 focuses on cost-effectiveness for inference scenarios, offering 600K IOPS while claiming to reduce first-token latency by 75% and double throughput.
• The LC 560 represents the largest SSD ever made at 245TB capacity with 14.7GB/s read bandwidth, targeting massive multimodal datasets in cluster training.

In plain English: Think of HBM as expensive, high-speed memory that AI systems use to process data quickly—like having a small but lightning-fast workspace. SSDs are larger but slower storage devices, like having a huge filing cabinet. Huawei is essentially creating super-fast filing cabinets that can partially substitute for the expensive workspace, helping AI systems handle more data without breaking the budget.

The big picture: This launch represents Huawei’s strategic response to U.S. restrictions on advanced HBM chip exports to Chinese companies, emphasizing domestic NAND flash technology over imported memory solutions.

How it works: Huawei’s DiskBooster software coordinates AI SSDs with both HBM and DDR memory to expand pooled memory capacity twentyfold.

• Multi-stream technology aims to reduce write amplification, potentially extending drive longevity in AI workloads.
• The approach follows a “system supplementing single points” philosophy, where different storage layers balance out HBM limitations rather than replacing it entirely.

What they’re saying: “The increasingly severe ‘memory wall’ and ‘capacity wall’ have become key bottlenecks to AI training efficiency and user experience,” said Zhou Yuefeng, vice-president and head of Huawei’s data storage product line.

• “This creates challenges for the performance and cost of IT infrastructure, affecting the positive AI business cycle.”

Why this matters: While these SSDs won’t fully replace HBM in LLM training, they offer Chinese AI companies a domestic alternative to reduce dependence on restricted imports, though the effectiveness of this approach remains to be proven at scale.