Dashboard

AIStor benchmarks8-node cluster

The exascale data store for the AI enterprise.

Two production-style sweeps of MinIO AIStor on 192× Solidigm D5-P5336 122 TB QLC NVMe and dual 400 GbE Spectrum-X fabric. Headline numbers below.

Explore benchmarks Cluster spec TCO calculator

Validated stack

Peak GET

267.76GiB/s

8 clients · 256MiB

vs fabric ceiling72%

Node scaling · single-NIC

Peak PUT

120.61GiB/s

7 clients · plateau

vs fabric ceiling32%

Node scaling · single-NIC

Peak sustained GET

195.24GiB/s

128MiB · c=192

vs fabric ceiling66%

Object size × concurrency · dual-NIC

Peak sustained PUT

133.71GiB/s

1GiB · c=192

vs fabric ceiling45%

Object size × concurrency · dual-NIC

GET TTFB median

3–5ms

Best 2 ms · all 8 cells

Burst GET (1-second)

272.43GiB/s

Headroom above sustained median

Sustained run

4.0hours

0 errors across 40 cells

2026-04-29

Node scaling

Add one client at a time on a single-NIC fabric — show how throughput scales 1→8 client nodes.

GET ramp 1→8 clients39 → 268

256MiBc=32/clientEC:4 (4 data + 4 parity, 24 sets × 8 drives)kernel 7.0.1-1.el10.elrepo

2026-04-28

Object size × concurrency

Dual-NIC sweep matrix — every combination of 4 object sizes and 8 concurrency points.

GET vs concurrency · 1 GiB75 → 185

4 obj sizes8 conc pointsEC:3 (3 parity, 12 sets × 16 drives, sdc=auto)kernel 6.12 (TractorOS RHEL 10.1 bootc)

Cluster at a glance

The hardware behind both runs

Full spec

CPU: Intel Xeon 6781P · 80c single-socket
MinIO: AIStor 2026-03-11 (DEV dual-NIC)
Raw capacity: ~23.5 PB across 192 drives
Fabric: Dual 400 GbE · MTU 9000
Erasure coding: EC:4 (4 data + 4 parity, 24 sets × 8 drives) · EC:3
Kernels: 7.0 (today) · 6.12 (matrix)
OS: TractorOS RHEL 10.1 bootc
Total errors: 0 across 40 cells