Samsung 950 Pro M.2 NVMe Performance

[rubylaser]

I just picked up a Samsung 950 Pro M.2 512GB for my workstation at the office. It's an MSI X99 SLI Plus with an i7-5820K overclocked to 4.5Ghz, 32GB, and two GTX 770's. I didn't see anyone on the forum with a benchmark for this drive yet, so I thought I'd provide at least a CrystalDiskMark. I'm very impressed at the performance of this drive for the cost of the drive. Let me know if you'd like to see anything else.

Edit: This is with Queue Depth 32, and 1 worker threads for both sequential and random io.

 

[Patrick]

I just picked up a Samsung 950 Pro M.2 512GB for my workstation at the office. It's an MSI X99 SLI Plus with an i7-5820K overclocked to 4.5Ghz, 32GB, and two GTX 770's. I didn't see anyone on the forum with a benchmark for this drive yet, so I thought I'd provide at least a CrystalDiskMark. I'm very impressed at the performance of this drive for the cost of the drive. Let me know if you'd like to see anything else.

I picked up a 512GB model also. I can probably post more numbers later today if anyone wants.

 

[rubylaser]

I picked up a 512GB model also. I can probably post more numbers later today if anyone wants.

I'd love to see some more numbers and hear from others if there are better benchmarks to run.

 

[vanfawx]

Would either @rubylaser or @Patrick be able to do a FIO on the raw device using the following FIO config? This is for random write 4k performance. I'd love to compare against my numbers for Intel 750's under Linux.

[Measure_RW_4KB_QD256]
ioengine=libaio
direct=1
rw=randwrite
norandommap
randrepeat=0
iodepth=64
size=25%
numjobs=4
bs=4k
overwrite=1
filename=/dev/nvme0n1
runtime=5m
time_based
group_reporting
stonewall

 

[rubylaser]

Would either @rubylaser or @Patrick be able to do a FIO on the raw device using the following FIO config? This is for random write 4k performance. I'd love to compare against my numbers for Intel 750's under Linux.

[Measure_RW_4KB_QD256]
ioengine=libaio
direct=1
rw=randwrite
norandommap
randrepeat=0
iodepth=64
size=25%
numjobs=4
bs=4k
overwrite=1
filename=/dev/nvme0n1
runtime=5m
time_based
group_reporting
stonewall

I'm sorry, I can't do that, I'm running my OS on this disk in Windows now. Maybe Patrick can take a run at it for you.

 

[Naeblis]

i did not get that write speed, however i am using a PCIe adapter. also could you bump up your queue depth?

 

[rubylaser]

i did not get that write speed, however i am using a PCIe adapter. also could you bump up your queue depth?

What do you want to see the queue depth at?

 

[Naeblis]

that depends on the # of your CPUs, my system had 56 in it. . let me get some pics

 

[rubylaser]

that depends on the # of your CPUs, my system had 56 in it. . let me get some pics

I have an i7-5820k, so just 6 physical cores and 6 hyperthreaded (this is just workstation at the office)

 

[vanfawx]

I'm sorry, I can't do that, I'm running my OS on this disk in Windows now. Maybe Patrick can take a run at it for you.

No worries Based on the numbers you provided, it looks like this does just a bit over 100k 4k IOPS at QD32.

 

[rubylaser]

Here are two more tests.

Queue Depth of 32, and 12 Worker Threads


Queue Depth of 256 and 12 Worker Threads

 

[Naeblis]

I guess the PCIe adapter really impacts the writes

 

[rubylaser]

I guess the PCIe adapter really impacts the writes
View attachment 1126

That is strange. I wouldn't think it would make that much of a difference. Did you install Samsung's NVMe driver for Windows?

 

[Patriot]

AHCI SM951

 

[rubylaser]

AHCI SM951

That looks faster for some operations and slower for most others. It seems like the 950 Pro NVMe isn't much of a performance increase over the previous AHCI version. Thanks for sharing!

 

[Naeblis]

I was hopeful. Installed & rebooted, no change

 

[rubylaser]

I was hopeful. Installed & rebooted, no change

Sorry, I hoped it would have helped out Is yours the 512GB version as well?

 

[Patriot]

That looks faster for some operations and slower for most others. It seems like the 950 Pro NVMe isn't much of a performance increase over the previous AHCI version. Thanks for sharing!

I don't think it is 3dnand either, so this is 840pro nand, on pcie.

 

[Keljian]

Sorry- the 950 pro has been sold as v-nand, so if you are right that is cause for a class action

 

[rubylaser]

You are both right. The 950 is advertised as 3-D V-NAND.