Samsung Demos Its First BGA SSD: 1500 MB/s Read Speed and Tiny Package
anandtech.com
anandtech.com
The performance of this thing is insane. Ars benchmarked it at 2600MB/s read, 1500MB/s write for sequential data[1]. Most SSDs benchmark at less than 600 MB/s[2].
By comparison, a RAM Disk gets ~6000MB/s on the https://hardforum.com/threads/quad-channel-ram-disk-benchmar... benchmark.
[1] http://arstechnica.com/gadgets/2015/10/950-pro-review-samsun...
[2] http://www.guru3d.com/articles-pages/samsung-850-evo-ssd-rev...
[3] https://hardforum.com/threads/quad-channel-ram-disk-benchmar...
Example: sudo fio --filename=/home/olav/testfile --direct=1 --sync=1 --rw=read --bs=4k --numjobs=8 --iodepth=32 --runtime=60 --time_based --group_reporting --size=1G --name=journal-test
fahad@heffalump:~/fiotest$ sudo fio --filename=$PWD/testfile --direct=1 --sync=1 --rw=read --bs=4k --numjobs=8 --iodepth=32 --runtime=60 --time_based --group_reporting --size=1G --name=journal-test
journal-test: (g=0): rw=read, bs=4K-4K/4K-4K/4K-4K, ioengine=sync, iodepth=32
...
fio-2.2.10
Starting 8 processes
journal-test: Laying out IO file(s) (1 file(s) / 1024MB)
Jobs: 8 (f=8): [R(8)] [100.0% done] [367.8MB/0KB/0KB /s] [94.2K/0/0 iops] [eta 00m:00s]
journal-test: (groupid=0, jobs=8): err= 0: pid=25184: Wed Jun 1 17:32:44 2016
read : io=22265MB, bw=379986KB/s, iops=94996, runt= 60001msec
clat (usec): min=15, max=19595, avg=83.59, stdev=102.01
lat (usec): min=15, max=19595, avg=83.65, stdev=102.02
clat percentiles (usec):
| 1.00th=[ 20], 5.00th=[ 65], 10.00th=[ 66], 20.00th=[ 68],
| 30.00th=[ 75], 40.00th=[ 79], 50.00th=[ 80], 60.00th=[ 82],
| 70.00th=[ 84], 80.00th=[ 90], 90.00th=[ 97], 95.00th=[ 108],
| 99.00th=[ 141], 99.50th=[ 173], 99.90th=[ 900], 99.95th=[ 1704],
| 99.99th=[ 4768]
bw (KB /s): min=43704, max=53976, per=12.51%, avg=47524.23, stdev=1225.35
lat (usec) : 20=0.74%, 50=1.42%, 100=89.28%, 250=8.32%, 500=0.09%
lat (usec) : 750=0.03%, 1000=0.03%
lat (msec) : 2=0.06%, 4=0.02%, 10=0.01%, 20=0.01%
cpu : usr=1.38%, sys=7.92%, ctx=5737418, majf=0, minf=89
IO depths : 1=100.0%, 2=0.0%, 4=0.0%, 8=0.0%, 16=0.0%, 32=0.0%, >=64=0.0%
submit : 0=0.0%, 4=100.0%, 8=0.0%, 16=0.0%, 32=0.0%, 64=0.0%, >=64=0.0%
complete : 0=0.0%, 4=100.0%, 8=0.0%, 16=0.0%, 32=0.0%, 64=0.0%, >=64=0.0%
issued : total=r=5699889/w=0/d=0, short=r=0/w=0/d=0, drop=r=0/w=0/d=0
latency : target=0, window=0, percentile=100.00%, depth=32
Run status group 0 (all jobs):
READ: io=22265MB, aggrb=379986KB/s, minb=379986KB/s, maxb=379986KB/s, mint=60001msec, maxt=60001msecRead is a lot (LOT) faster than writes. Enough to make me thing I might have something misconfigurated there(?), although I'm not familiar enough with this benchmark to know for sure.
Ironically, we thought that Windows 7 would have had better support - but, as it turns out, for Circa 2H2015+ hardware, it's probably not a great platform. Here's hoping that Windows 10 stabilizes for a few years, and all the vendors agree to support that platform....
Use this guide http://forum.notebookreview.com/threads/guide-installing-win... It worked for me. Also this :) http://www.samsung.com/semiconductor/global/file/insight/201...
I have a windows 2008R2 box at home running on a 950pro.
With MBR the system would still take 30-60s to boot, but GPT reduced that to less than 10s.
You may need to copy the install data to a USB drive to get it to work. Initially I burned the iso image to a DVD and but then that DVD drive wasn't bootable under UEFI.
Windows 7 doesn't have M.2 drivers by default. See http://www.tomshardware.com/answers/id-2817439/install-windo... and good luck! I believe Windows 10 is fine though.
"Most SSDs benchmark at less than 600 MB/s[2]."
of course they do, that is the theoretical limit of non pcie drives.
Ditto - I built mine at the beginning of the year. It's very fast, but I learned two things that weren't widely reported at the time.
1) The chips on the drive get _very_ hot under use (thermal throttling kicks in at 75C). Even with minimal use they seem to have a running temp > 50C. I'm a little nervous about the long-term reliability of the drive based on the heat alone.
2) Boot speed with Windows is still fairly slow unless the drive is formatted to GPT (GUID Partitiion Table) instead of the default MBR. It's possible to do this when installing Windows, but there's no easy way to convert after installing
For what it's worth, the Anandtech review [0] mentioned your first issue in their review back in October. Anandtech has missed things in the past, but I trust them a lot to find and point out possible issues with hardware.
[0] Linking to the last page 'Final Words' section as it is mentioned up front and centre there. A lot of times, I'll read or skim the first page, and jump directly to the final page conclusion for hardware I just want a quick overview of - http://www.anandtech.com/show/9702/samsung-950-pro-ssd-revie...
1. https://cdn.sparkfun.com/assets/8/5/c/7/4/52e157adce395f4759...
If not, what is the relationship between all these interface types? Are they intended to be soldered onto the motherboard?
Manufacturers may decide to solder the SSD directly to the motherboard, like Apple currently does with RAM on their laptop computers. This is beneficial to the manufacturer because it's more compact, less chance of mechanical failure, easier to assemble (pick&place instead of humans), lower price etc. In short: it's very attractive for manufacturers.
For consumers this is worrying though, as with the RAM on Apple laptops:
- You cannot replace a failed SSD, you need to replace the entire motherboard
- You cannot upgrade the SSD, so you have to accept the ridiculous price stepup when ordering the laptop (i.e. the 200 dollar price bump to get 8GB extra RAM on a macbook pro)
- Laptop is less valuable when trying to sell it on after a couple of years, because the storage size will be low compared to new SSD's.
NOPE. The Apple Store just sends you to a data recovery company instead, because they can't read it.
There's no better sign of bad customer service when you use a custom part but even you don't care to learn how to fix it.
This makes manufacturers happy, because more end-users will buy new devices rather than get a used one.
I just discovered this concern last weekend when a less than 3 year old Samsung SSD 840 died. Since this is not mechanical I don't know why they don't include extra pins (e.g. address, data) to read the memory in raw from the SSD chips if the failure is in the controller or firmware.
It's maddening that not only can I not get back whatever data wasn't backed up (thankfully not much) but I have no way of wiping it before returning it under warranty.
The people who handle the returns for Samsung say returned drives will be "formatted". Given it's uunresponsive, this seems unlikely
It puzzles me why drive manufacturers don't offer data recovery for their own products as a value add service. Nobody could get the data off this drive as cheaply as Samsung themselves.
Obviously we have our minds completely synchronized. At this point I don't care about the warranty but about the safety of the information in the disk. I contacted datarecovery.com and they gave me a rough range between $ 300 and 2k to recover the SSD. I don't think the price is crazy but I imagine that is a simple task to achieve if you have the right equipment.
Broken or failed screen, input device, storage controller, yada, -- doesn't matter. FDE means your data is that much more safer than without (and makes it safer provided you use a strong key).
Where there's a will, there is a way, and all that...
In a similar vein, you could double the RAM on the original Xbox by soldering a second memory chip to the unpopulated pads on the underside of the board.
If the controller fails, with existing "discrete" SSDs you can still desolder the flash chips directly, put them in a reader, and recover the data that way (e.g. https://news.ycombinator.com/item?id=8133450 ) With this integrated "SSD", it could be nearly impossible to read the raw flash, meaning no more data recovery beyond the logical level. It's basically like an SD card.
less chance of mechanical failure
The widespread problem with RoHS lead-free solder and cracking balls disagrees, as also evidenced by the large amount of reflow/reballing services and the fact that reliability-critical applications like aerospace still use leaded solder and more expensive solder columns instead of balls.
While reading that myself, I also happened upon solder springs, which seem like an even niftier idea: http://www.topline.tv/CCGA_MCS_Micro_Coil_Spring.html
Apparently that company makes the solder springs available also has RoHS safe versions of them, perhaps this could solve the cracking problem?
Another company has a paper out which says that you can use their solder column attach process with existing BGA chips: http://www.solderability.com/newsletter082005/ConvertingBGAC...
Could some manufacturer (while unlikely, curious if it's possible) choose to use solder springs with this SSD's BGA package to make it more mechanically reliable when installed?
Unless you only use the computer for editing text then storage requirements are growing quicker than ever.
I don't mind paying for the higher end 512GB rMBP, but the discreet graphics is a deal-breaker for me. Thus I'm left with configuring the baseline integrated graphics rMBP with 512GB flash.
1. Driver issues on Linux
2. Discrete graphics failing and taking the whole logic board with it. Most of my laptops have died this way.
3. Battery life
4. Heat
[0] (search for "thermal" on this page and through the article): http://www.anandtech.com/show/9702/samsung-950-pro-ssd-revie...
May as well go with PCIe cards then. :)
Also, which cellular provider here in the states is offering 30 GB / month at that low of a price (do they have special rural pricing or something)?
So yeah, the RasPi idea of having using those to boot might corrupt quickly, but if you use them for a camera or just transferring files between computers, you should be safe.
If not, then you're mostly out of luck, but you can try if you manage to source it. And find some datasheets. And have a driver for it. And...
On a more serious note: It would be very nice, if it was possible to design a board with components like this, but it is way to hard and needs way too much equipment and experience to do right.
What caught my eye in this SSD package is a low ball count, so it might be potentially suitable for DIY even without an expensive hardware. If it's possible to route it on 4 layers without too many small vias it might be a game changer.
On the other hand, 1,500 MB/s data and PCIe are quite serious levels of performance by most DIY project standards. :)
EDIT: actually, no, it's a single lane PCIe3, such clock rates are well beyond the FPGA range. Pity.
You seem to be familiar with PCIe 2 though, so now I'm wondering - is there a chip that could convert the 1x PCIe 3 lane into a 2x or 4x PCIe 2 set of lanes? Also, worst case, if you really wanted to use this chip, couldn't you just use the chip with PCIe 2 (albeit slower, but I think revisions of PCIe are supposed to be cross compatible)?
Never heard of such a thing, although it might have been really useful.
> couldn't you just use the chip with PCIe 2
Did not even think of this option, sounds interesting. I guess we should wait for the datasheet for this chip to be sure.
This form factor has a huge advantage for OEMs - they can just slap on a single chip that has everything built in - controller, memory management and the actual storage. Less parts = less convoluted assembly process and that is cheaper. While theoretically you can put a BGA fully integrated SSD SOC like this on a PCI-E card or M2 etc., it kinda beats the purpose.
I did my early engineering years with DIP parts, the days of when you could just desolder the memory device and place with a higher density device are long gone.
If you can get your memory density high enough there is no need for multiple chips, if there is no need for multiple memory chips then a discrete controller doesn't makes sense since it only needs to communicate with a single chip in that case the extra die costs can easily be canceled out by cheaper packaging costs and the fact that you can now sell SSD's to system integrators that could not use them before.
We're not quite at the stage of putting the controller in the chipset and just soldering Flash to the board for PCs, although that is how it's done for SoC systems.
That should be an absolute dealbreaker, though. Talk about holding your data hostage...
Also, get real, most people don't have any backups set up at all. They still shouldn't be artificially prevented from accessing their data in case of a failure...
Also known as eMMC.
Can't imagine a 512GB ROM would ever be economical given the cost to manufacture and make manufacturing changes to such a beast
I have been wondering about going back to carts. When I see a 64GB Flash Drive for $10 retail I have to think that that the price premium over pressed Blu-Ray disks starts to make sense. You can significantly cut down on loading times, avoid obnoxious installs to the local disk, and have potentially more flexibility in the future to release larger games that don't require multiple disks.
Plus, the optical drive is one of if not the most failure prone component on the system (next to only the fan), so eliminating it could cut down considerably on returns and warranty claims.