Dell Quad M.2 PCIe card.

JustLong

Gawd
Joined
Jun 24, 2002
Messages
782
Got my hands on this bad boy today. Unfortunately there is no hardware RAID. Though with storage spaces we saw over 8GB/s read and 6GB/s write speed.


Storage spaces offered better sequential performance
M2SPACES.png

Creating a dynamic disk/traditional software RAID had better random performance.
M2DYNAMIC.png
 
Part number and specs? Can you build a SS with 2 of these cards and 8 drives?

EDIT: Also, that looks like Raid0. Any numbers for either parity or mirror in SS?
 
Last edited:
Wow, those look like Ramdisk speeds!

Edit - Oh, is there a heatsink that goes over the m.2 drives like the HP version?
 
Last edited:
BTW those results are not accurate. You need to add more threads and even than CDM will not report the true speed. CDM is CPU limited and doesn't work well at those speeds. You want to use another program that isn't limited by CPU like AIDA or AS SSD but note those don't liek virtual drives..aka softeware RAID or RAM disks. They won't scan drives like that...at least i couldn't get them to work with my RAM disk or RAID.

BTW you should be over 10GBps with that set up in certain tests.

basically dont trust CDM with anything over 2GBps...its a trash program. So you getting different results with 2 different RAID/storage? solutions doesnt mean shit with that program. Don't use it to judge if one option is faster than the other. You need a better benching utility.
 
Wow, I just have one question, is it bootable. I would buy one in a heart beat if it's bootable on a motherboard that doesn't have any m.2 slots.
 
anyone know what the thread was that the giuy was trying to get a 750 to replace the EVO his company was working on for like mapping data? This thing would be perfect for him. I can't find that thread to post this
 
Part number and specs? Can you build a SS with 2 of these cards and 8 drives?

EDIT: Also, that looks like Raid0. Any numbers for either parity or mirror in SS?
I'll dig up a part number tomorrow. Though I believe it wont work in most systems due to the way the card is implemented. Just RAID 0, Ill post numbers for other configurations when I get a chance.
Wow, those look like Ramdisk speeds!

Edit - Oh, is there a heatsink that goes over the m.2 drives like the HP version?
No heat sink, just a shroud to guide air across the SSDs.

BTW those results are not accurate. You need to add more threads and even than CDM will not report the true speed. CDM is CPU limited and doesn't work well at those speeds. You want to use another program that isn't limited by CPU like AIDA or AS SSD but note those don't liek virtual drives..aka softeware RAID or RAM disks. They won't scan drives like that...at least i couldn't get them to work with my RAM disk or RAID.

BTW you should be over 10GBps with that set up in certain tests.

basically dont trust CDM with anything over 2GBps...its a trash program. So you getting different results with 2 different RAID/storage? solutions doesnt mean shit with that program. Don't use it to judge if one option is faster than the other. You need a better benching utility.
Thanks for the input. I'll give those a shot. Just to note, the results were consistent through multiple executions. Also this is being tested in a Dual 6 Core Workstation so there is a decent amount of CPU resources for CDM to use.

Wow, I just have one question, is it bootable. I would buy one in a heart beat if it's bootable on a motherboard that doesn't have any m.2 slots.
Sorta. It came configured with 1 of the M.2 SSDs as the boot SSD, but since there is no hardware RAID there is no way to use more than 1 SSD for the OS. As I mentioned above I don't believe it will work in most systems. I'll test it in a non Dell motherboard to see if it works, posts, and detects more than 1 M.2 Slot.
 
I'll dig up a part number tomorrow. Though I believe it wont work in most systems due to the way the card is implemented. Just RAID 0, Ill post numbers for other configurations when I get a chance.

No heat sink, just a shroud to guide air across the SSDs.


Thanks for the input. I'll give those a shot. Just to note, the results were consistent through multiple executions. Also this is being tested in a Dual 6 Core Workstation so there is a decent amount of CPU resources for CDM to use.

Sorta. It came configured with 1 of the M.2 SSDs as the boot SSD, but since there is no hardware RAID there is no way to use more than 1 SSD for the OS. As I mentioned above I don't believe it will work in most systems. I'll test it in a non Dell motherboard to see if it works, posts, and detects more than 1 M.2 Slot.
with my expierence with a 1650v3 is that even when you do 2-3 threads most CDM uses is 1.5 cores/threads. So even when i selected 6 threads i still saw a different in benchmarks the more MHz I gave my CPU. The results will be consistent but not accurate.


BTW even though its just a fan it should be more than enough. It is a much larger fan than HPs so it should be just fine. It looks not as fancy or low profile though.

How much was the case?
 
How much was the case?
I'm evaluating workstations for purchase so I got a few machines from Dell and this was included in one of the configs I'm testing, so currently nothing :)
So far the performance improvements this card offers has been near zero as we are CPU bound more than IO bound with how we use the workstations, even with 24 threads performing concurrent operations.
The synthetic testing was just for fun as the actual workloads are all that matter.
 
I'm evaluating workstations for purchase so I got a few machines from Dell and this was included in one of the configs I'm testing, so currently nothing :)
So far the performance improvements this card offers has been near zero as we are CPU bound more than IO bound with how we use the workstations, even with 24 threads performing concurrent operations.
The synthetic testing was just for fun as the actual workloads are all that matter.
Z-ON F-I say it with me Xeon Phi

Unreal engine needs to support Xeon Phi for servers. I could host my 30 tick 500 player server on valhala :/ intead of a 120 player 10 tick server :/

Hell you could probably do 1000 players with that 3 TFLOPs one that comes out in a few months. :/

I dont know why people like UE4 so much. server and game client are 1.5 cores so single thread :/

/end random rant
 
Thanks but I was aware of this for a while now. You even replied to it.

Need 2TB SSD for IO intensive work

I know i replied to it...i couldn;t find it for the life of me. I looked through all my watcched threads so i must not have watched it and just kept checking back or something. oh well at least you already knew so i dont need to feel bad :D

BTW if this works with 1TB 950 PROs you could try to get them to give this to you. That would be 4 TB of crazy IOs
 
I had my rep contact dell about this card, 3k for a 1tb version, and something very expensive (think it was 12 or 13k) for 3tb version. They wouldn't sell it unloaded.
 
I had my rep contact dell about this card, 3k for a 1tb version, and something very expensive (think it was 12 or 13k) for 3tb version. They wouldn't sell it unloaded.
dumb. The HP is like 400 right? For wahtever reason thats what i thought i read. 2K for a stupid RAID card?
 
My 32GB Ramdisk for comparison, thru SoftPerfect RAM Disk

LtoOE80.jpg


so2PqiS.png


Guessing this Dell drive is probably cheaper per GB, but not by much :p
 
My 32GB Ramdisk for comparison, thru SoftPerfect RAM Disk





Guessing this Dell drive is probably cheaper per GB, but not by much :p
CDM does not work for RAMDisks. Your RAM is depending on system 20-80GBps or more (in rare cases octa channel or quad 3200MTs)
 
The ram bandwidth of your system should be significantly higher than the 5 to 7GB/s that CDM shows however a ramdisk will have overhead in the OS and in the driver and CDM has to load the data from ramdisk to ram so 1/2 of your ram bandwidth is the limit without counting any overhead.
 
AIDA shows you real RAM speeds but AIDA does not test virtual drives....its BS so no good program exists on testing virtual drives like RAID or RAMDisks.

high MTs quad channel system >80GBps So skylake dual channel systems (3200 MTs) are ~40GBps
e5f2df75_image.jpeg

If you watch task manager CDM is single thread limited but you can add extra threads but CDM will never fully load a CPU. At best it'll use half. I tried on my 1650v3 and best i could do was 50% CPU with RAMDisk. CDM is a trash program on anything over 1-3GBps It even has issues reporting the 950 PRO in certain systems.


AS SSD is not coded like shit but like AIDA it can not due virtual drives like RAID or RAMDisks, which sucks.
 
I'm still not understanding, what is the problem with CDM?

And I don't think the transfer speed can really be debated, whether it's the Dell card or the RAM Drive. Quoting theoretical bus speeds as evidence as why the program is a bad benchmark doesn't make much sense. Those speeds don't exist in a vacuum when used in actual application.
 
I'm still not understanding, what is the problem with CDM?

And I don't think the transfer speed can really be debated, whether it's the Dell card or the RAM Drive. Quoting theoretical bus speeds as evidence as why the program is a bad benchmark doesn't make much sense. Those speeds don't exist in a vacuum when used in actual application.
see my updates to the post above above with spoiler. CDM is trash. Play around with Threads and watch your utilization.

also thats not bus speed....thats actual RAM.
 
So... no one makes quad, or similar M.2 NVME adapters but Dell and HP?

I've been looking, and all I see are single M.2 adapters.
 
AIDA shows you real RAM speeds but AIDA does not test virtual drives....its BS so no good program exists on testing virtual drives like RAID or RAMDisks.

high MTs quad channel system >80GBps So skylake dual channel systems (3200 MTs) are ~40GBps
e5f2df75_image.jpeg

If you watch task manager CDM is single thread limited but you can add extra threads but CDM will never fully load a CPU. At best it'll use half. I tried on my 1650v3 and best i could do was 50% CPU with RAMDisk. CDM is a trash program on anything over 1-3GBps It even has issues reporting the 950 PRO in certain systems.


AS SSD is not coded like shit but like AIDA it can not due virtual drives like RAID or RAMDisks, which sucks.
Getting way o/t but...

First, I can't possibly fathom how someone can get this angry over a benchmark program.

Secondly, you're saying CDM is "trash" and "coded like shit" and AS SSD is somehow better, yet it doesn't recognize virtual disks or RAID. I just downloaded it and tried it; not only does it not see the RAMDrive it won't even open if I switch over to hard drive emulation.
Whereas CDM and Anvil's work without HDD emulation, HD Tune and ATTO do not see it but the program at least STILL FUNCTIONS.

Results of ATTO and HD Tune with HDD Emulation. Clearly slower (as advertised)
3Jc8VDU.jpg


kIvErBc.jpg


I'm guessing you're just trolling. You have to be.
 
i am not angry at all...its just a trash program for fast drives and highly inaccurate.

Again AS SSD/AIDA are better as i said in regards to programming (performance/accuracy not features). It uses the CPU better and doesn't get CPU bottle-necked.

Whats the CPU utilization of those 2 programs. I wouldn't be surprised if they were single thread bottle-necked. Change your CPU MHz and you'll probably see them produce different numbers.

Here is an example i did with my 950 PRO. This is why CDM is trash for anything fast.

EDIT***Ops had the numbers reverse***
4.8GHz HW
Dropbox - 950 PRO 4-8 HW CDM.png
4.2GHz HW
Dropbox - 950 PRO 4-2 HW CDM.png

my drive is mostly full so i cant get a good full speed test with the newer version with the "threads" but the threads are highly broken and never load more than 50% of the CPU no matter what u do. The new version just barely handles 950 PRO but does not work for massive RAID arrays or RAMDisks. They cap out very easily.
 
Last edited:
Whatever. I'm out of this convo.

you can post stupid memes all you want. I could careless. Its not my fault your too stupid to understand why the program is crap....i honestly can't understand how you are this challenged.

BTW you made my day...really dont know how i can break this down any further barney style for you.

CDM is poorly optimized and reports inaccurate result. You can see this by playing with threads and CPU speeds....and seeing the results.

AS SSD/AIDA are not CPU limited due to crappy coding.

Not complicated...go test it....-_-
 
I wish they would use these SFF ports on high end motherboards, instead of the shit show of ports like on the ASRock X99 Extreme 11.

Yeah well these ports are used to carry PCIe lanes, (not sata/sas, like those connectors were originally designed for, and are used for in servers). But I agree, either these or SFF-8087 ports would be much better than a crapload of individual sata ports.
 
I have 2 of these cards and I am having problems getting them to boot windows. Windows PE will see the drives, format and place the wim and as soon as the system reboots into the OS it will BSOD.
 
Are you trying the cards in a non dell and are the cards using NVMe drives?
 
I'd take a simple motherboard that had tons of those SFF/U.2 ports onboard. They'd need additional controllers though, I'm not sure the PLX itself can handle those types of ports without the use of a Marvel or Intel chip.
 
Are you trying the cards in a non dell and are the cards using NVMe drives?

They are in a Dell workstation. They came already loaded with windows 7 on the NVMe drives but cant seem to get past the part after the reboot when it places the wim where it starts to run sysprep.
 
I'm evaluating workstations for purchase so I got a few machines from Dell and this was included in one of the configs I'm testing, so currently nothing :)
So far the performance improvements this card offers has been near zero as we are CPU bound more than IO bound with how we use the workstations, even with 24 threads performing concurrent operations.
The synthetic testing was just for fun as the actual workloads are all that matter.
is there used any PLX chips on this board?
 
I'd take a simple motherboard that had tons of those SFF/U.2 ports onboard. They'd need additional controllers though, I'm not sure the PLX itself can handle those types of ports without the use of a Marvel or Intel chip.
U.2 is just straight up PCIe - 4 lanes. No marvell or other controller chip required. You could use a PLX chip to add many ports, or just use the ports from the chipset.
 
U.2 is just straight up PCIe - 4 lanes. No marvell or other controller chip required. You could use a PLX chip to add many ports, or just use the ports from the chipset.

I wasn't sure if there was a port limit to the PLX or not
 
I wasn't sure if there was a port limit to the PLX or not
Depends on the specific PLX chip you use, they have chips that support up to 48 lanes, so you could have 16 lanes in and 32 lanes out to 8 SSD's, for example.
 
Depends on the specific PLX chip you use, they have chips that support up to 48 lanes, so you could have 16 lanes in and 32 lanes out to 8 SSD's, for example.

I've only seen those in extreme HPC cases but since those exist I'm sure they could tailor ones for other platforms.
 
Back
Top