DEFEATING DMI 3.0
Traditionally, there has always been a bottleneck in large scale SSD performance which is DMI 3.0 and this performance bottleneck occurs when trying to move data through the systems chipset. Trying to use the SSD7101A-1 through the chipset would limit the output to PCIe 3.0 X4 and decrease the data speeds significantly.
Add to that the fact that you are also sharing DMI 3.0 with other SSDs, USB and SATA ports and things can get very frustrating. With the SSD7101A-1 running via X16 travel, we are pushing full data speeds directly to the CPU, rather than the chipset.
HighPoint’ performance claims only 128Gb/s on paper, but we know we are speaking to throughput of at least 7GB/s and around 1 million IOPS. To add to this, HighPoint has stated that two of these devices can be paired up in RAID for an output of 256Gb/s.
HighPoint provides a RAID driver for the SSD7101A-1 as well as a NVMe RAID Manager which can be downloaded from the link provided by clicking the just above. In our testing, we have had better performance without their driver, but the software is actually very useful.
From the ‘Manage’ menu, your SSDs are identified and creating a RAID volume is quick and easy. RAID can be created in types 0,1,5 and 10 but 5 and 10 do not support TRIM. Clicking ‘Maintenance’ on the right side will allow you to shut down RAID on your SSDs. We can see that I have created RAID 0 from these drives, but if I were not to have done this, the drives would all still be identified and perform at their listed performance from the OS, each as its own logical drive.
We found the biggest benefit of the ‘Setting’ menu was the fact that we could confirm when the device was running in X16 and Gen 3. Different motherboards can be a real PITA at times, and also, we couldn’t manage to boot from the RAID. To go one step further, turning on VROC on our ASRock motherboard displayed the SSDs we were using, but also listed that they weren’t compatible with Intel VROC. We are pretty certain that running four Optane SSDs through the SSD7101A-1 would allow boot right from the UEFI.
This shot is of our OS Disk Management screen with the four SSDs being used in a JBOD configuration.
I wonder if a Threadripper build could run this with dual GPUs?
Waiting for it to arrive. I am told it can…and then some.
I heard only intel boards are listed as supported at their site?
Wow, I look forward to your piece on it.
Comparing this w/ the native TR raid drivers coming soon.
I nurse a hope that amdS lanes are cleaner than intels. Certainly there are a lot of ifs and buts attached to pre x299 mobo pcie3 x16 slots.
The normal spec for TR is 3x native m.2 ports on the mobo, so 3x ssd in raid on TR is very democratic – available to most wise TR buyers if u boot on sata ssd – no expensive controller needed.
From raid benches i have seen, 3x raid 0 seems a sweet spot in the diminishing returns table.
No it wont. TR is limited to 2x 16 lane cards sadly.
Hey! Great review. So this unit is suppose to be bootable? Just possible issues with your particular motherboard?
Nice review! Could also review the native nvme raid on threadripper with the drivers that will be released on 25th of September. Amd claims that it supports up to 7 nvme drives on raid.
I dont recall details as of no use to me, but i did research bootable raid on the net, & the upshot is u can boot on anything, there is just an arcane procedure to set it up – like a clean install of win from a flash stick onto a suitably pre-configured array ~.
I gather from comments the article hasnt clarified that only intels are listed in the supported moboS, or so I hear.
It seems odd, as its perfect for lane rich TR/epyc, bar the high cost & possibly superfluous raid capabilities (tr & epyc can perform raid unaided by controller)
The usual frustrating refrain re raid nvme is “its already so fast, u wont notice the difference” – true, but your system will, and u will notice that.
To me the exciting concept is virtual memory. For the first time ever, the boundaries between ram and storage are getting blurred by such arrays.
For discerning parts of this virtualising process (as we see here, iops are no better on raid), intelligent 8GB/s paging/juggling of big slabs of contiguous data from such storage arrays up and down the speed hierarchy of the cache pool, could be extremely useful.
As a sign of things to come, we see the radeon vega pro ssg, with 2TB of raid nvme attached tot the gpu card – a snip at $7kus. The onboard nvme raid can be used as a cache extender for the gpu onboard 16GB hbm2 ram.
All vega cards can use hbcc to do same cache extending, only via the more handicapped pcie3 bus to system resources like these arrays or spare system memory.
This auto expand and contract of cache at system level is a revolution for coders. They can simply assume ~unlimited gpu cache, and let the system handle the details.
AMD “slides” on hbcc claim a bump in gpu ram of 2-3x can be simulated, un-noticed by users, by adding spare system ram to the hbcc vega gpu cache pool.
I posted thoughts on what boil down to – add a small fast non boot array to your system, and soon a use will come along & u will be glad u did.
So here is what i figure for a cheap rig to test the water.
Get a TR w/ 3x nvme ports. preserve these precious native raid supported ports for your raid and find another way of booting.
factors are:
the drives are dear. u have been generous with ports/lanes, nand improves fast, u dont need space (multiples of existing ram are nothing vs ssd array capacities), only speed,…
If the progression in specs from the 950 pro to the now oldish 960 pro are any guide, we will soon see nvmeS which saturate their allotted 4GB/s pcie3 bandwidth – in both read and write.
at that point, you can get the most from your 3x array, so its tempting to go cheap til then.
while the samsung 960 pro 500GB ssd is the current king for this, 3x ~$370us? is a lot, and still well shy of the magic 4GB/s at ~3.2 GB/s SeqRead & 2.2GB/s SW.
The samsung evo retail range are good, slower and cheaper, and also come in much cheaper 256GB size, and may stripe to some very respectable speeds, yet keep outlays minimal until the really cool ssdS arrive, at which point, you have 3 x very salable 256GB ssd ~desktop boot drives.
then we come to the matter of oem versions of these same drives, which are available out there even cheaper.
Really cheap? There are even such panasonic 128GB nvme ssdS – dunno, some say they lack sufficient cache, oth, maybe raid 0 has the effect of tripling the apparent cache?
NB I am not suggesting u rely on this storage. I am suggesting it as a scratch workspace for temp files and virtual page files as above. Its like volatile dram, if suddenly lost, the job can be recovered. This is the coding norm for such files. Debate on possible failure rates (~anachronism from hdd days imo) is beside my point.
So yeah,w/ a 3 x nvme port TR w/16GB & even a future 4GB VEGA card, the above array could realistically provide ~12GB/s read and write in a year or so, 9 & 5.25 GB/s now w/ 960 pro, & for not much money (compared to large ram certainly), 256GB panasonics should yield 7GB & 4.5GB/s.
(remember sata ssdS are ~500MB/s & HDD is about 120MB/s when they finally find the file)
In perspective, most agree,few games suffer from using 8 lanes vs 16 lanes for demanding gaming, and 8 lanes equates to 8GB/s, so an 8GB array is like system memory bandwidth to a gpu card.
Apart from the drives, it wont cost a nickel extra on your planned TR, maybe even a smaller spend on “precautionary” memory.
There are those who slag off bios raid & some who praise/slag off software raid. Dunno. Its hard to trust anything when so many attitudes seem remnants of another era.
A curious one is some round on windows software raid for the huge? overheads it introduces, which begs the question, does it matter when we now we all 🙂 have cores doing nothing.
It puzzles me that there is a huge demand for such cards as these, but they almost dont exist. You can get a pcie3 x4 lane adaptor for a single nvme ssd cheap as chips ($15) anywhere, but ~nothing bigger exists i can find, even lacking raid.
Why is it so simple to make a dumb 1x adaptor, and so hard to make a dumb 2x (8 pcie3 lane) or 4x (16 lane) adaptor? Thats all thats needed for AMDs bios raid or MS software raid.
fyi
comparison results on an x99? intel mobo, 6x 960 pro 1GB invarious raid arrays
https://imgur.com/a/a68Sd
There always seem to be some glitches in all of them, but patterns emerge, raw numbers aside.
The table gives a good idea of the sort of multiple in raw speed that can be expected i the eco system can handle these radical bandwidths.
one pattern is there seems to be a big drop of in read gains for the fourth drive in an array – ?, maybe its just me?
another pattern that stands out, is write speeds, as alluded to in the article.
In some apps, fast read may be meaningless if writes are slow, and writes lag seriously on single ssdS.
The numbers show tho, that writes profit far more from striping than read speeds, and it may pay a user to keep adding drives to the array, even when write gains are dropping off.
another curiosity I noticed elsewhere, was they tested the arrays 4x 960 pro individually, and the variance was quite extreme – ~2800MB/s vs 3200MB/s read? Its very damaging to add sub par drives to an array, asthe lowest speed drive dictates the pace for the whole array.
This card is passe now on the TR4 platform as you get FREE RAID from AMD:
Der8auer hits 28GB/s on TR4 in I/O meter and 8xNVMe Samsung 960 Pro!
https://www.youtube.com/watch?v=9CoAyjzJWfw
ASUS Hyper M.2 x16 Card (less than 100$)
https://www.komplett.se/product/945185/datorutrustning/lagring/haarddisk/ssd-pci-express/asus-hyper-m2-x16-card
ASUS X399 Rog Zenith
Change the PCI slot into 4x4x4x4x mode from original 16x mode.
1900X can push 64 pci-lanes.
Done.
As someone who dual boots different OS that don’t necessarily support motherboard RAID, that’s really the target market left for the PCIe RAID controllers.
can you use it with just one or two m2 ssd in normal or raid configuration?
Yes you can.