Beta Q4 2026
maxRAID Grid
Performance
at any scale.
Scale‑out block storage, built for cloud providers and AI workloads. Multiple nodes, one system. Add nodes and keep growing — rebalancing is automatic.
Architecture
Pools that span the grid.
A grid is square: N nodes with N NVMe drives each form N storage pools. Each pool is a K+M erasure‑coded maxRAID array with one drive in every node, on its own diagonal — in the example above, ten nodes with ten drives form ten 8+2 pools. To grow, add nodes in grids of the same geometry. Spares sit outside the square: a spare drive in every node, a spare node, or both.
Spares, standing by.
Add a spare drive in every node, a spare node, or both. When something fails, rebuild starts at once — shortening the time pools run degraded.
Redundancy from the grid itself.
Protection comes from parity across servers — without buying mirrored media.
No central metadata server.
Each pool runs independently and performs like a local array. The grid is managed with maxRAIDOne.
Performance
Add nodes.
Add performance.
Each storage pool performs like a local array. Clients stripe each volume across pools — and with no central metadata server in the data path, every pool you add contributes its full performance.
Storage pools
Grow when you need to. Performance grows with you.
IOPS add up the same way: up to 2 million per pool — 20 million across the ten 8+2 pools of ten nodes with ten SSDs each.1
One system. Room for every tenant.
Local‑array speed.
No metadata bottleneck. Each array performs as if it were in the server.
Built for AI.
NVMe‑oF over TCP and RDMA, GPUDirect, and shareable multi‑mount volumes.
Grows with you.
From 50 TB to more than 10 PB. Use ultra‑large QLC SSDs to reach multi‑petabyte capacities.
Size your grid.
Move the sliders. Performance adds up pool by pool, until the network becomes the limit.
One 10 × 10 grid · 10 pools per grid, one per drive slot · spare node at right
- IOPS
- 20M
- Storage pools
- 10 8+2 each
- Usable capacity
- 1.23 PB
- before compression
- Hardware
- 11 nodes
- 100 NVMe drives in pools, plus 1 spare node
- Survives
- 2 failed drives in every pool
- Rebuild
- Starts at once
- onto the spare node
Illustrative estimates from design targets. For a real sizing of your workload, talk to an engineer.
Get a sizingEconomics
Less flash for every
usable terabyte.
Keeping three copies of every block leaves a third of your flash usable. A 10+2 parity array leaves 83% usable — and, like three copies, it survives any two drive failures.2
| 1.8 PB usable · 15 TB SSDs | Ceph, three copies | maxRAID Grid | maxRAID Grid, typical data |
|---|---|---|---|
| Drives | 361 | 144 | 110 |
| 5‑year SSD cost | $2.55M | $508K | $388K |
| Performance | ~4M IOPS · ~16 GB/s | 24M IOPS · 240 GB/s | 22M IOPS · 220 GB/s |
A fraction of the flash.
With highly compressible data, 64 SSDs and $226K.
Compression and thin provisioning.
In‑line, so every byte goes further.
Near‑theoretical wear.
Support for QLC and PLC SSDs with large indirection units.
Your hardware.
White‑box x86_64 servers, standard Ethernet and modest RAM.
Resilience
Servers fail.
Service doesn’t.
Failed SSDs are replaced automatically. Failed servers fail over to hot spares through virtual target IP addresses — automatically, and invisibly to applications.
Design targets1
Node + drive
Tolerates both node and drive failures.
< 30 s
Failover, from failure to service.
0
In‑flight data lost during failover.
Full speed
Writes stay at full speed in degraded mode.
At a glance
Built for cloud providers and AI workloads.
- Performance
- Up to 20 GB/s and 2M IOPS per storage pool, read and write, when the network can carry it; volumes stripe across pools
- Protection
- You choose the parity: +2 or more per pool. Each pool survives as many failed drives as it has parity drives — and losing a whole server costs each pool only one. With 8+2 pools, as in the 10 × 10 example, that means any two drives per pool, or two whole servers. With a spare, rebuild starts at once.
- Scale
- 50 TB to 10+ PB
- Protocols
- NVMe‑oF (TCP, RDMA, GPUDirect) and iSCSI; shareable multi‑mount volumes for AI
- Multi‑tenancy
- Thousands of volumes per server; QoS by bandwidth or IOPS, burstable
- Data services
- Snapshots, backups, in‑line compression, thin provisioning
- Self‑healing
- Automatic SSD replacement; server failover to hot spares via virtual target IPs
- Integrations
- Kubernetes CSI driver, at general availability; Proxmox VE plugin planned
- Hardware
- White‑box x86_64 servers, standard Ethernet; NVMe SSDs (TLC, QLC, PLC) for full performance
Scale without compromise.
maxRAID Grid enters beta in Q4 2026, with general availability targeted for Q1 2027.
- maxRAID Grid is pre‑release. Performance and resilience figures are design targets — up to 20 GB/s and 2 million IOPS per storage pool, adding with each pool a volume is striped across, provided the network can carry the traffic — and final figures will be published at general availability.
- All figures are modeled, not measured: a planning example for 1.8 PB of usable storage on 15 TB SSDs. maxRAID Grid configurations: 12 × 12 (no compression); 10 nodes × 11 drives — ten 8+2 pools plus a spare drive in every node — for typical data that compresses by 30%; and 8 × 8 for highly compressible data (60%).