Forum Discussion
Why NVMe Gen 5 RAID 0 on AMD/INTEL Collapses Your Read Speeds (And how Microsoft can fix it !)
As a preamble, I woud like to clarify that each time I did/do some research on that matter, the suggestions are always the same:
- if you want the performance, renounce booting, and drop RaiXpert2 in favor of other *existing* RAID 0 sofware implementation;
- if you want booting, renounce performance, and you are stuck with RaidXpert2;
the *ONLY* difference is the software, and this is *NOT* about any hardware bootleneck, even for such high throughput.
Look at this guy who fixed it for Linux:
https://github.com/joeytroy/amd-raid-driver
it is the proof it is perfectly doable.
So just like for Resizable BAR (Base Address Register) story, are we going to wait another decade for firmware providers to deliver the real performance that is already sitting in our desktop computers ?
Please advise.
========================================================
Here is below the block diagram of the MSI MAG B850M MORTAR WIFI (v1) MicroATX motherboard I bought in early 2025 for (in my opinion) its simple design suitable (optimized ?) for DirectX gaming on Windows 11 thanks to:
- direct CPU to DDR5 RAM connection (obvioulsy)
- direct CPU to PCIe 5.0 x16 GPU connection
- direct CPU to 2 x M.2 NVMe Gen5 x4 allowing RAID 0 configuration
- no PCIe lane sharing
Lately, more and more motherboards are offering direct CPU to 2 x M.2 NVMe Gen5 x4 allowing RAID 0 configuration.
However, I was quickly disappointed when I tested my 2 x M.2 NVMe Gen5 x4 RAID 0 volume performances as sequential reading was and still is shockingly slower than sequential writing !
Back then I opened a ticket directly to AMD customer support, but it went nowhere (I have the ticket # in my e-mails if needed...).
Recently, thanks to free Google AI mode that indexed/learned virtuallly all knowledge available on Internet, I investigated again that matter and I found the problem root cause that I am sharing now hoping it will bring someone's attention.
If you are running 2 (two) ultra-fast PCIe Gen 5 NVMe SSDs in RAID 0 using motherboard firmware controllers—such as AMD’s RAIDXpert2 or Intel's RST/VROC—on a modern desktop platform, you have likely run into a massive performance defect. Your write speeds scale linearly as expected (~24,000 MB/s), but your sequential read speeds collapse down to half (~10,000 MB/s) as expected, performing worse than a single standalone drive.
Here is a quick breakdown of why this happens, and why it is time to demand a software ecosystem overhaul.
The Root Cause: Legacy Driver Architecture Bottlenecking Next-Gen Hardware
Your motherboard lanes and CPU are perfectly fine. The issue is a severe execution flaw inside the closed-source proprietary driver stacks (Code:
rcraid.sys
for AMD and Code:
iaStorAVC.sys
for Intel).
- [] Why Writes Scale (~2x): Writing is asynchronous. The driver dumps the data straight into your system RAM cache or the SSDs' local DRAM/SLC caches instantly. The driver just acts as a non-blocking traffic cop, firing data to both drives simultaneously, a.k.a 'write and forget'.
- [] Why Reads Collapse (< 1.0x): Reading is synchronous. The driver must actively calculate exactly where the striped blocks live across both physical drives, pull them, and stitch them back together.
The "One Worker, Two Conveyor Belts" Bottleneck
Firmware-assisted RAID is not true hardware RAID; it is software RAID running at the CPU level. Kernel profiling reveals that these driver stacks use a single CPU worker thread to process synchronous read I/O routing.
The Analogy: Imagine two ultra-fast conveyor belts (your Gen 5 SSDs) dumping packages into a mailroom at a combined 30,000 MB/s. However, there is only one single worker (one CPU thread) tasked with reading the barcodes on every package to sort them in alphabetical order. The worker cannot physically move fast enough. Packages pile up, and the conveyor belts are forced to constantly grind to a halt to prevent an overflow. This stalls your entire PCIe pipeline.
When NVMe drives operated at Gen 3 speeds (~3,500 MB/s), a single CPU core was fast enough to hide this inefficiency. At Gen 5 speeds (~14,000+ MB/s), this legacy, single-threaded software design acts as an artificial wall, completely strangling the hardware before it can ever ramp up to its rated line speeds.
The Call to Action: AMD, Intel, and Microsoft Need to Fix Their Drivers
Consumers are paying premium prices for high-end motherboards and cutting-edge Gen 5 storage, yet the firmware and driver software ecosystem completely castrates the hardware capability under real-world read scenarios.
What needs to change:
- [] AMD & Intel Must Modernize Storage Stacks: The single-threaded architecture of
- Code:
- rcraid.sys
- and
- Code:
- iaStorAVC.sys
- is an outdated relic. Both chipmakers must redesign their proprietary driver engines to natively support multi-threaded polling, modern asynchronous interrupt handling, and modern direct-memory access when available. If modifying the low-level UEFI firmware execution layer presents safety or design challenges, the optimization must be implemented within the Windows kernel-level OS drivers.
- [] A Joint Effort with Microsoft: Windows already possesses highly optimized, multi-threaded storage schedulers (as seen in Windows Storage Spaces). AMD and Intel need to collaborate directly with Microsoft to allow firmware-initiated RAID volumes to hook cleanly into native, highly parallel Windows Kernel worker pools ( Code: storport.sys multi-queue integration).
Hardware engineering has outpaced software design. We shouldn't be forced to abandon motherboard RAID configurations or rely on complex OS-level workarounds just to unlock the storage throughput we already paid for. It is time for AMD, Intel, and Microsoft to patch this multi-generation oversight and deliver true parallel execution for high-bandwidth NVMe arrays.
1 Reply
- OwenLantyBrass Contributor
The read/write numbers alone don't establish a single-worker driver bottleneck. Capture the same sequential-read test with Windows Performance Recorder and compare CPU usage, storage queue depth and per-drive throughput; also note the RAID driver and firmware versions. That trace would give AMD or Intel a reproducible driver case and distinguish a CPU-bound path from caching or test-size effects.