SupremeRAID™ GPU-Accelerated RAID Performance Evaluation
Benchmark comparison of SupremeRAID™, Linux MD RAID and a single NVMe-oF over a GP-SPARK direct-attach connection
Key Finding: SupremeRAID™'s Advantage in Parity RAID
On RAID 5 / RAID 6 write performance, SupremeRAID™ (GPU offload) substantially outperforms MDRAID (CPU software RAID).
SupremeRAID™ throughput expressed as a multiple of MDRAID throughput.
Sequential Write Bandwidth Comparison
SupremeRAID™ vs. MDRAID at the same RAID level. MB/s; bs=2M, numjobs=40, iodepth=32, 60s per pattern.
The chart covers the parity levels only (RAID 5 / RAID 6); RAID 0 and the single-NVMe references are in the table below.
Performance Summary (Sequential, MB/s)
| Configuration | Intel Read | Intel Write | DGX Read | DGX Write |
|---|---|---|---|---|
| SupremeRAID™ RAID 0 | 8080 | 4731 | 7805 | 4804 |
| SupremeRAID™ RAID 5 | 7906 | 2926 | 7725 | 2930 |
| SupremeRAID™ RAID 6 | 8129 | 2389 | 7687 | 1964 |
| MDRAID RAID 0 | 7990 | 4811 | 7401 | 4821 |
| MDRAID RAID 5 | 7202 | 660 | 7399 | 446 |
| MDRAID RAID 6 | 7280 | 536 | 7467 | 343 |
| Single NVMe-oF | 2060 | 1448 | 1923 | 1438 |
| Single NVMe (Local) | 6302 | 2402 | 11700 | 12700 |
RAID 0 shows near-parity across both approaches (no parity computation); the SupremeRAID™ advantage becomes pronounced on RAID 5/6.
dd Throughput Cross-Validation
100 GiB file write × 5 runs, average of the steady-state runs 2-5 (GB/s), cross-validating the fio trend.
| Configuration | SupremeRAID™ | MDRAID |
|---|---|---|
| RAID 5 | ≈ 2.95 – 3.00 GB/s | ≈ 0.85 – 1.60 GB/s |
| RAID 6 | ≈ 2.00 – 2.13 GB/s | ≈ 0.76 – 1.35 GB/s |
Values are approximate readings from the source report's chart, spanning both test environments; indicative of the trend only.
GP-SPARK - Test Environment
- CPU
- Intel Xeon Platinum 8380 ×2 (40 cores/socket, 2.30 GHz)
- Memory
- 256 GB
- OS Drive
- 1.92 TB NVMe SSD
- GPU
- NVIDIA A40 (Driver 610.43.02)
- Connection
- Direct-attach to GP-SPARK
- OS
- Ubuntu 24.04.4
- CPU
- NVIDIA Cortex-X925 3.9 GHz (10 cores)
- Memory
- 128 GB (Unified Memory)
- OS Drive
- 4 TB NVMe SSD
- GPU
- NVIDIA GB10 (Driver 580.159.03)
- Connection
- Direct-attach to GP-SPARK
- OS
- Ubuntu 24.04.4
GP-SPARK - Test Methodology
- I/O Engine
- libaio (direct=1)
- Block Size
- 2 MiB
- File Size
- 20 GiB / job
- Parallel Jobs
- 40
- I/O Depth
- 32
- Runtime
- 60s per pattern (30s cooldown between)
- Patterns Tested
- seq_write / rand_write / seq_read / rand_read
- Command
- dd if=/dev/zero of=/data/data-N.dat bs=1M count=100K
- Transfer Size
- 100 GiB / run
- Repetitions
- 5 consecutive runs
- Aggregation
- Average of steady-state runs (2nd-5th)
Configurations tested (both environments)
- SupremeRAID™ RAID 0 / 5 / 6 (GPU-accelerated RAID)
- Linux MD RAID 0 / 5 / 6 (software RAID)
- Single NVMe-oF (comparison baseline)
- Single NVMe (local disk, reference only)
Notes
- The figures above are taken from the HPCTECH Corporation report "SupremeRAID™ GPU-Accelerated RAID Performance Evaluation".
- Actual performance varies with host platform, SSD model, network topology and OS tuning.
- For a PoC in your own environment or for the full test report, please get in touch with us.