By Nathan Brooks

DapuStor's J5060 dual-mode SSD claims a write latency below 8 microseconds in its pseudo-SLC region, matching Kioxia's purpose-built SLC FL6 on the same PCIe Gen 4 bus. The 30.72-terabyte drive pairs that cache with bulk QLC capacity in a single part, an arrangement the Chinese vendor is pitching at AI training workloads.

The trade-off is plain to see: it is still QLC hardware pretending to be SLC. Whether the disguise holds depends on the workload, and on how carefully one reads the endurance figures.

What does the dual-mode approach actually buy?

The pSLC region delivers more than seven times the random-write IOPS of a QLC-only configuration, according to DapuStor. In 4 KB random-write tests, the company reports an average write latency under 8μs. Kioxia's real-SLC FL6, also PCIe Gen 4, sits at 8μs. Micron's similar XTR runs closer to 15μs. The J5060 thus lands in SLC territory without the SLC price tag.

The endurance picture carries the same caveat. True SLC cells tolerate between 50,000 and 100,000 write cycles; QLC cells manage around 1,000. DapuStor says its pSLC region lifts endurance by more than 25× over QLC alone. A genuine improvement, admittedly, though still a long way from genuine SLC.

How is the cache carved up?

A single 30.72 TB J5060 can be configured with 400 GB (8%), 800 GB (14%), or 1.2 TB (20%) of pSLC, with the remainder as QLC. The ratio is fixed at provisioning time and applies uniformly across every drive in an array. Three firmware tweaks underpin the scheme: isolation between QLC and pSLC workloads, optimisation of the SLC reserve to cut write amplification, and region-aware scheduling so latency-sensitive data travels the shortest path through the device.

Why bother mixing rather than separating?

DapuStor frames the answer as one of consolidation. In a 24-drive chassis of dual-mode parts, write traffic distributes across every drive. With one SLC drive fronting 23 QLC drives, the same traffic funnels through a single PCIe link. Bursty AI training can saturate that one link while the rest of the array sits idle. Aggregate throughput suffers accordingly.

The second argument concerns failure domain. A failed SLC drive in the segregated layout forces the cluster to reconstruct its contents from QLC peers, a recovery DapuStor describes as taking days and degrading cluster performance. A failed dual-mode drive only loses its own pSLC region. Rebuild volumes shrink, recovery time shrinks with them, and the rest of the cluster stays in service.

So what is the verdict?

The J5060 cannot match true SLC for either raw speed or endurance, and DapuStor does not pretend otherwise. The case for it rests on density per drive, fewer parts to provision, and a single SKU covering both journal and bulk roles. For AI labs scaling large training clusters, that combination may finally outweigh the purity argument.

Capacity is winning the argument, for now. Endurance purists will need to keep reading the data sheet before signing the purchase order.