Memory & StorageLOT 261122 · 4 min read

VAST Data: Petabyte SSDs Put RAID Data-Loss Odds at 1 in 11

VAST Data's SC26 simulation of a 900 PB petabyte-SSD system shows enclosure-local declustered RAID carries 1-in-11 odds of data loss over 10 years. Its fix: system-wide erasure coding.

By Grace Kim800 words

Key takeaways

  1. VAST's 900 PB simulation puts data-loss odds at 1 in 11 over 10 years with enclosure-local declustered RAID and petabyte SSDs.
  2. A 900 PB build from 15 TB SSDs would draw 5.6 MW; 122 TB drives cut this to about 400 kW.
  3. VAST's own scheme reaches 1 in 62 billion odds; its stated minimum is 1 in 54 million.
  4. Metadata kept as three copies rates 1 in 58 million over 10 years versus 1 in 78 trillion for data drives.
  5. Micron shipped the 245 TB 6600 ION in May; SanDisk, Solidigm and Kioxia have 245-256 TB drives arriving.
VAST says petabyte SSDs will outgrow today’s parallel file system RAID
VAST says petabyte SSDs will outgrow today’s parallel file system RAID — Diosming / unsplash

VAST Data has put the odds of data loss on a 900 PB petabyte-SSD system at 1 in 11 over 10 years when enclosure-local declustered RAID is used. The figure comes from a discrete event simulation built on real failure data, presented by the company's principal technical strategist, Glenn Lockwood, at an Exhibitor Forum talk during Supercomputing 2026 (SC26) in the HPC System Hardware and Architecture session.

The core problem is rebuild time. A 122 TB drive takes twice as long to rebuild as a 60 TB drive at the same per-drive speed, because per-drive capacity is outpacing bandwidth and "the performance of each individual drive does not change," Lockwood said. Longer rebuild windows mean more time for second and third failures that destroy data.

Why operators will still want the big drives

Lockwood's case for petabyte SSDs rests on power. He prices storage at about 28 W per SSD regardless of capacity, including its share of surrounding infrastructure, plus 20 W per GB/s of bandwidth. Building 900 PB from 15 TB SSDs would take more than 63,000 drives and draw about 5.6 MW — "the equivalent of 3,100 or so of Blackwell Ultra GPUs" — to deliver 190 TB/s of bandwidth. He noted that "real AI applications in our experience have never approached needing anything close to that level of performance" from storage.

A 122 TB configuration reaches almost 8,000 drives, 8 TB/s aggregate bandwidth, and roughly 400 kW — less than a tenth of the power. Within that 400 kW budget, 15 TB SSDs top out at about 250 PB with nothing left for performance. The savings can go to compute.

What the simulation found

The simulated system used 7,920 SSDs at 122 TB each across 180 enclosures of 44 drives. Lockwood presented these results:

  • RAID 5 single parity: 100% chance of data loss over 10 years, though "no one uses single parity anymore."
  • Double parity: 1 in 150 odds, described as "state of the art for many open source file systems." He named Lustre and BeeGFS.
  • Enclosure-local declustered RAID — the approach used by DDN EXAScaler and HPE Cray ClusterStor, per Lockwood — with petabyte drives swapped in: 1 in 11.
  • Global declustering plus extra parity, his stated minimum: 1 in 54 million.
  • VAST's own scheme, not detailed: about 1 in 62 billion.

Lockwood called the usual reaction to a 245 TB drive accurate: it is "a big blast radius." His recommended fix is system-wide erasure coding with at least three parity strips. Global declustering alone helps only slightly, and it enlarges the fault domain because stripes spread over every drive make two random failures likelier to hit the same stripe. VAST already publishes quadruple parity.

How the incumbents protect data

DDN's DCR, in one published configuration, uses RAID 6 in an 8+2 array for a 24-NVMe SSD pool, with each stripe spread over 10 of the 24 drives. DDN has conceded in the past that larger systems may need more than DCR. HPE's GridRAID declusters parity per enclosure half — two 12-drive pools per 24-drive enclosure. Perlmutter at Berkeley Lab's NERSC, where Lockwood once worked, used GridRAID with 8+2+1 RAID 6 per 12-drive Lustre object storage target; the +1 spare spreads rebuild writes across all surviving drives. In every case, protection stops at two failures within one or two dozen drives, and only those drives can assist a rebuild.

Metadata is the weakest link

VAST keeps metadata — the index of where every file piece dwells — on separate fast SSDs as three replicated copies rather than parity, surviving only two failures. With 122 TB drives, data drives rate 1 in 78 trillion over 10 years, against 1 in 58 million for metadata. Rebuilds in the index area start about 12 seconds after a failure, versus about two minutes for data. VAST power-cycles a dropped drive and hopes first; waiting an hour for data drives "wouldn't make a big difference," while 20 minutes or more on a metadata drive carries "extremely high" risk.

The drive pipeline

Micron began shipping the 245 TB 6600 ION in May, drawing up to 30 W at maximum load. Kioxia's LC9 appears in a Dell 2U server with 40 drives totaling 9.8 PB. SanDisk's 256 TB UltraQLC was due in the first half of this year, and Solidigm targets roughly 245 TB later this year. Multiple roadmaps reach petabyte-class drives, including SanDisk's UltraQLC, an SK hynix QLC target, and Samsung at an unspecified date. Lockwood's power math supports the shift, but he argued the data must still be compared openly and fairly to settle the best path forward for a problem he considers unavoidable.

Tom's Hardware (Source)

Filed under

Share this article:

More from Grace Kim

Grace Kim

Show full bio

Market editor covering people, launches and day-to-day practice in Semiconductors, chip manufacturing, hardware, AI compute at Chip Dispatch.

1,213 articles

Read also

« Previous article