I have built a rig that it is used for Graph Database analysis on Threadripper utilising Samsung 980 PRO's
(detailed spec here: https://pcpartpicker.com/b/dT7TwP )
I have have two arrays with 2X512 GB Drives each.
In the first one, that is also the system/boot disk, the NVMe disks are on the motherboard with direct access to the CPU.
I am getting a 11.5GB/s - 110k IOPS seq read but just 2GB/s - 90k IOPS seq write speed (Samsung magician figures) - having a 8microsecond lag
The second array, it is mounted in the M2.DIMM port (using a special card provided with my ASUS ZENITH II EXTREME)
I am getting a 10.5GB/s - 115k IOPS seq read but up to 8GB/s - 105k IOPS seq write - having an 11 microsecond lag!
The lag it is to be expected because Array 2, is connected to the CPU via the Chipset; Array 1 is directly connected to the CPU (see the diagram of TRX40 in the link provided above).
The question is how on earth the write speed in the first array is so low. Both disks are just about 30% of their capacity.
The only difference is that Array 1 has been formatted in Windows using 4k, whereas Array 2 has been formatted using 256k (same with AMD CTS).
To perplex more the issue, Array 1 used to have a 10GB/s seq write when I installed it. It just somehow, after some bios and drivers updates that the speed has collapsed.
Any ideas?
