RAID 0 NVME Read Performance Same as Non-RAID

The pure read performance issue, e.g. only at single-drive levels has been duplicated using all connection combinations of: M2_1 + M2_2, M2_1 + M2_3, M2_2 + M2_3 and lastly M2_3 + M2_4. Uniquely write performance is only on par with expectations when using M2_1 (CPU bus) + M2_2 (Chipset bus). Contrariwise write performance drops to single drive speed when testing all other connection combinations.

To prevent SSD throttling, active cooling (fan/heatsink) has been used on all connections except for M2_3 + M2_4. They are too close to support active cooling. Used ASRock's passive system board heatsink in that instanced. The included thermal resistors were used on all test scenarios to monitor real-time heat confirming all tests were performed well within design specs, e.g. to prevent throttling. Additionally, the room was maintained at constant 68 degrees Fahrenheit throughout testing.

The inability to scale read performance in all combinations using RAID 0 indicates that the controller is not effectively parallelizing read operations between the two SSDs regardless of the bus CPU/Chipset or connections: M2_1, M2_2 and so on.

All hardware/software are new, built exclusively for this rig.

The BIOS is running version 3.20
A RAID 0 array is comprised of two Gen 4 NVME SSD's setup in BIOS.
Win 11 was installed using Microsoft's public Windows 11 Installation Media while following the instructions noted in AMD RAID Driver v 9.3.3.00117 e.g. using rcbottom.inf, rcraid.inf and rccfg.inf during advanced setup to ensure the single 8TB single volume is recognize during first time setup. Windows OOBE was setup without issue.

Post OOBE the following core apps/drivers were installed (in order):
AMD Chipset Drivers Revision Number 7.02.13.148
AMD RAID Installer Revision Number 6.10.09.200
AMD Software: Adrenalin Edition Adrenalin 25.3.1
ASRock Motherboard Utility ver:4.1.12
MediaTek Bluetooth driver ver:1.1037.0.424
MediaTek Wireless Lan driver ver:5.3.0.1825
Nahimic3 utility ver:1.10.2_APO4
Realtek high definition audio driver ver:6.4.0.2395_UAD_WHQL
Realtek Lan driver ver:10.071.0425.2024
APP Shop ver:2.0.0.6
ASRock Polychrome RGB ver:2.0.191

Hardware:
Systemboard ASRock X870E Taichi
CPU AMD Ryzen 7 9800X3D
RAID 0 NVME Read Performance Same as Non-RAID

CPU Cooler Cooler Master 360 Atmos Close-Loop AIO Liquid Cooler
Memory TeamGroup T-Force Delta DDR5 Ram 32GB Kit (2x16GB) 7200MHz x2 = 64GB
Power SupplyCorsair HX1200i
Graphics - Using iGPU on the CPU for testing, no add-in GPU for now.
Storage Samsung 990 Pro 4TB NVMe Gen 4 SSD x2 (8TB)

Here are some example test results on three tests highlighting the issue with CrystalDiskMark v8.0.
[table]
[tr][th]Read Seq MB/s - 1MiB[/th][th]Write Seq MB/s - 1MiB[/th][th]Bus Config[/th][th]Comments[/th][th]Test Config[/th][/tr]
[tr][td]7143.922[/td][td]7001.150[/td][td]M2_2 and M2_3 and PCIe Gen 4x4[/td][td]RAID 0: Chipset only bus[/td][td]NVMe settings/profile[/td][/tr]
[tr][td]7342.190[/td][td]13748.016[/td][td]M2_1 and M2_2 and PCIe Gen 5x4 > Gen 4x4[/td][td]RAID 0: CPU and Chipset bus[/td][td]NVMe settings/profile[/td][/tr]
[tr][td]7148.922[/td][td]6921.065[/td][td]M2_3 and M2_4 and PCIe Gen 4x4[/td][td]RAID 0: Chipset only bus[/td][td]NVMe settings/profile[/td][/tr]
[/table]

the nvme drives firmware is updated?

this is BIOS/UEFI RAID?

anyway i would blame:
-os/drivers
-nvme drives firmware
-motherboard/bios/hardware

I will give that a go and reply with results. Great suggestion!

Thank you all for helping debug this issue.

It appears the initial suggestion was AI-generated, based on the style and format in the original post that referenced "Can I use a Fedora Live CD to mount...". While it was helpful to some extent, the proposed solution ultimately did not work as expected due to several limitations. Fedora LiveCD/USB environments come in various versions, and commands can behave differently depending on the specific combination of version and environment used.

For example, the following command failed when executed in the Fedora 42 Live environment (booted via USB media):

sudo dnf install dmraid

The failure message was:
"failed to resolve the transaction" and "no match for argument dmraind:"

To provide additional context: Fedora 42's Live CD/USB environment does not include `dmraid` by default, nor does it allow installation via the above command. A subsequent search within the booted environment, executed as follows:

sudo dnf search dmraid
...returned:
"no matches found."

The search was performed after the install command failed, simply to confirm whether the `dmraid` package might already exist in the environment?”but it does not.

So, since this approach appears to be a dead end (at least thus far), are there any alternative methods to non-destructively test RAID 0 read performance?

For reference, when testing in the native environment (Windows 11), the issues are consistently related to read commands, irrespective of hardware connection configurations. However, write commands only produce at RAID 0 performance thresholds when one SSD is connected directly to the CPU bus and the other is on the chipset bus. These findings have been consistent across all tested configurations.

NDRE28 wrote: I am running my RAID-0 setup on an AMD AM5 platform. One of my SSDs is connected directly to the CPU, while the other 2 are connected to the chipset. OS: Windows 11 Enterprise v24H2. In my case, everything runs as expected, in terms of performance. RAID-0 did improve the transfer speeds.

Regarding RAID 0 performance, are you achieving close to 2x read/write speeds compared to a single SSD? This presumes at least one array is in a 2x NVMe/SSD RAID 0 config. Ideally, the performance should approach the theoretical maximum achievable across the bus and scale as additional, identical SSD's are added to a RAID 0 array.

I?™m particularly interested in whether you?™re obtaining optimal results, especially in comparison to my own setup. Interestingly, I previously purchased (and later returned) an 8TB WD NVMe drive that exhibited slower performance than a single Samsung NVMe. Theoretically, it should have been significantly slower than two Samsung NVMe drives configured in RAID 0, yet the actual performance difference was negligible.

At present, the results seem inconclusive. I?™d love to hear your insights on this.

My primary goal is to build a RAID 0 array using two Samsung 990 Pro 4TB NVMe SSDs, resulting in a single 8TB volume optimized for read/write performance in line with the hardware capabilities?”specifically, nearly doubling bandwidth for reads and writes. Theoretically, this setup should achieve read speeds of approximately 14,900 MB/s and write speeds of about 13,800 MB/s on par with Gen 5x4 single non-RAID NVMe's

However, RAID 0 read performance has failed to effectively utilize parallelization, with read speeds mimicking single-drive performance as measured by both Windows 11?™s winsat tool and CrystalDiskMark. For instance, across three CrystalDiskMark test runs?”each consisting of five repetitions?”the average read performance was 7,211 MB/s, irrespective of the slot configuration.

Interestingly, write performance closely aligned with real-world RAID 0 expectations, reaching approximately 13,748 MB/s when utilizing the M2_1 and M2_2 slots with PCIe Gen 5x4 with the Gen 4x4 configuration. In contrast, other connection and bus combinations delivered suboptimal write performance, averaging around 6,961 MB/s. This includes all tested physical configurations, such as M2_2 with M2_3 (PCIe Gen 4x4) and M2_3 with M2_4 (PCIe Gen 4x4).