Storage Review: Graid Technology Makes AI Storage Fast AND Resilient with SupremeRAID™ AE

Read the Full Report from Storage Review Here: Storage at GPU Speed: Benchmarking SupremeRAID™ AE for AI

Storage Review put SupremeRAID™ AE (by Graid Technology) on a Dell PowerEdge R770 with dual NVIDIA H100s and 16 Micron 61.44TB E3.S SSDs to see if GPU powered RAID can deliver both speed and resiliency for AI without burning a PCIe slot.

SupremeRAID™ AE runs the RAID engine on a small slice of the GPU, aggregates NVMe into a single namespace, and supports NVIDIA GPUDirect Storage for direct paths between storage and GPU memory. In our lab, the single RAID 5 pool hit about 183.60 GB/s random 1M reads and roughly 54.23 GB/s random 1M writes on FIO. GDSIO peaked near 88.5 GiB/s on reads and around 45.9 GiB/s on writes, showing that GPU ingest is the practical limiter more than the array.

During a VLM inference run, staging a 172 GB batch resulted in only about a 4 percent tokens per second dip, indicating minimal GPU overhead for AE. The build also shows why AE matters for real deployments. You keep slots free for high speed IO, you get a large unified namespace, and with 16 x 61.44TB drives you are near a petabyte raw in 2U.

Full Report: https://www.storagereview.com/review/…

More from Graid Technology: https://www.graidtech.com/product/sup…

00:00 Why RAID on the GPU

00:47 SupremeRAID AE on H100 overview

01:26 Platform and drives: R770, dual H100, 16 x 61.44TB E3.S

02:19 Do we sacrifice GPU cycles for RAID

02:57 Server teardown and slot economy

04:00 Enterprise AI in 2U and storage adjacency

04:28 Micron 6550 ION performance profile

06:26 AE on any platform and slot layout notes

07:20 OCP NICs, PCIe lanes, and IO options

08:55 Test focus: inference impact vs hero numbers

09:13 Results: about 183.6 GB/s sequential read

10:05 GDSIO ingest results

11:09 Inference impact about 4 percent

12:34 Key takeaways

#nvme #raid #ai

‍