Glossary
Flash storage vs HDD
Flash storage vs HDD is the comparison between solid-state storage built on NAND flash and hard disk drives that record data magnetically on spinning platters. The two media differ by orders of magnitude in random access speed and differ substantially in cost per terabyte, capacity per device, power use and failure behaviour.
Why the media choice matters at scale
At a few hundred terabytes, choosing flash or disk is mostly a performance decision. At tens of petabytes it becomes the largest single factor in hardware cost, power draw and rack count, and it shapes how the system behaves during rebuilds, full-dataset reads and growth. Most large estates end up running both, and the real decision is which data goes where.
How the two media work
A hard drive reads by moving a head to the right track and waiting for the sector to rotate under it. At 7,200 RPM one rotation takes 8.33 ms, so the average wait for rotation alone is 4.17 ms; with head movement a random read takes 5 to 15 ms. Once positioned, a drive streams data well, which is why disks remain strong at sequential work.
NAND flash has no moving parts. A read senses charge in a page of cells in tens of microseconds, and a drive serves many reads in parallel across its dies. Writes are slower than reads, pages are erased in large blocks before reuse, and cells wear with writing, as covered under flash storage endurance.
Flash and HDD compared
| Property | Flash (NVMe SSD) | HDD |
|---|---|---|
| Random read latency | Tens to low hundreds of microseconds | 5 to 15 milliseconds |
| Random IOPS per drive | Hundreds of thousands or more | Around 100 to 200 |
| Sequential throughput per drive | Several GB/s | Roughly 200 to 300 MB/s |
| Largest capacities | Above 100 TB (QLC) | Tens of TB |
| Purchase cost per TB | Higher | Lower |
| Wear mechanism | Limited write endurance | Mechanical wear, no write-count limit |
| Behaviour under mixed load | Largely unaffected by access pattern | Random access collapses throughput |
Throughput per terabyte and rebuild time
Hard drive capacity has grown much faster than hard drive speed. A hypothetical 20 TB drive streaming at 250 MB/s offers 12.5 MB/s per stored terabyte and takes (20 × 1012) ÷ (250 × 106) = 80,000 seconds, about 22 hours, to read end to end. A 30.72 TB NVMe drive at 6 GB/s reads end to end in about 85 minutes. The same gap applies to random access: around 200 IOPS across 20 TB is 10 IOPS per terabyte.
Cost comparisons flip with the unit. Measured per terabyte stored, disk is cheaper. Measured per random operation, flash wins by a wide margin: delivering 100,000 random reads a second from disk at around 150 IOPS per drive takes about 670 drives, a load one NVMe drive can carry. The meaningful comparison is the cost of meeting both the capacity and the performance requirement of a dataset.
What flash and HDD mean for petabyte-scale storage
Placement follows access pattern. Capacity that is read sequentially or rarely, such as backup targets, archives, video libraries and the history of a data lake, stays economical on disk, where streaming throughput across many drives adds up. Small objects, metadata, active AI training sets and random-read services belong on flash, where disk would spend most of its time seeking.
Large hard drives are turning into sequential devices. Each capacity generation spreads the same number of random operations over more terabytes, so a disk tier that stores more per drive serves fewer random requests per terabyte. When analytics or AI workloads start hitting data that used to be cold, request density rises and that data moves to flash.
Disk throughput depends on keeping access sequential. A hard drive streaming one large read loses most of its throughput once a second stream forces it to seek back and forth, which is why disk tiers depend on software that lays data out and schedules reads to keep each drive streaming.
Rebuild windows grow with disk size. While a large hard drive is being reconstructed, the data it held runs with reduced protection and the rebuild consumes bandwidth the workload would otherwise use. Flash shortens that window even as flash drives grow, because reads are so much faster.
Space and power shift the balance in constrained sites. High-capacity flash fits more terabytes into each rack unit than disk, which favours flash where floor space or power is the limit, while disk keeps the lower purchase price per terabyte.
Software that runs on standard servers lets both media sit in one system as separate tiers or as hybrid servers, with storage tiering moving data between them.
Flash, disk and drive rebuilds in Scality RING
RING runs on standard x86 servers fitted with flash, disk or both: hybrid servers mixing flash and hard drives are common among large deployments, and RING XP is the all-flash configuration on NVMe. When a drive fails, RING writes data across the remaining drives in that server and rebuilds only the data that was written, so rebuild time on large hard drives tracks how full the failed drive was.