Seagate Firecuda 530 2TB + ASRock WRX80 Creator R2.0 throttled performance

[Kish]

Hey all. I have an ASRock WRX80 Creator R2.0 with a Threadripper Pro 5955WX (16C/32T). I’m testing out the Seagate Firecuda 530 (2TB variant) as it seems like the best NVMe gen4 SSD right now (for write/endurance).

When the drive is used in a gen4 (4x4x4x4x) PCIe bifurcation card, I get the full 7.4GB/s read, 6.9GB/s write speed from the drive (Q8T1).

However, when used in the motherboard M2_2 (slot 2), I’m only getting 6.4GB/s read/write speeds (Q8T1):

Temps are within spec (39C idle, ~60C load writing), so there is no throttling. I feel like it’s hitting some arbitrary hard-coded wall here.

Now, I understand that M2_2 (slot 2) is connected to the WRX80 chipset, and not to the CPU directly (as with M2_1 (slot1) and the PCIe slots). But slot 2 should still have a gen4 x4 connection:

Furthermore, the WRX80 chipset itself is connected to the CPU via a gen4 x8 connection, so there should be plenty of bandwidth available. I have some SATA3 drives connected (which also go through the WRX80 chipset), but I can’t imagine those even remotely saturating a gen4 x8 chipset link. I don’t have anything else connected to the WRX80 chipset.

I want to use this drive in the mobo M2_2 (slot 2) as a super fast caching drive for VM vDisks (since no redundancy). And save the PCIe lanes / bifurcation card for more important things.

What’s causing the bottleneck here?

[Blackbox514]

I don’t find it unreasonable to think that any device that is connected via the chipset as opposed to directly to the CPU will be subject to additional latency, no matter which platform is used.

The abundance of PCIE lanes is pretty much the point of investing in a Workstation class system, so why would you want to share lanes for high speed storage?

Personally, I would use one of the “nerfed” x 8 slots using a dual bifurcation card rather than m.2 through the chipset.

[Quension]

If it is just latency, then theoretically a comparison with higher queue depths would come out more equal.