IOP/s Like You’ve Never Seen
This is exciting stuff. I took some time–and a handy-dandy flipcam–on my recent trip out to the MAST Center to film Joe Jeddeloh’s demonstration of our enterprise PCI-e prototype. It’s really cool to prove (on hardware) what was thought-to-be-possible on paper. Come on into the lab and see the kind of IOP/s we’re hitting with this thing for yourself.
55 Comments to “IOP/s Like You’ve Never Seen”
Leave a Reply
Congrats. Guys for such thing …
there will be enormous demand for it but will you deliver? I mean that much supply as it is needed (still in some acceptable price 500-800$).
?
Secondly … maaan when can I put this into my laptop
Seriously , is it SLC/MLC …and what about wear level of such device ?!
Christopher, we have plans to deliver this product to selected customers in 2009, with wide availability in 2010. Unfortunately this specific product is not targeted at your laptop (unless you have a very big laptop!) but this is a good indication on where this technology can take us in the future. It is within the realm of possibility to deliver this type of performance in a laptop if the architecture would support it. Currently available buses in laptops would not keep up. In fact, desktop buses don’t keep up, either.
This particular product is SLC, and will likely stay SLC. The reason is the performance. While we do employ the most advanced wear-leveling algorithms in this product, the longer programming times for MLC would bring our performance numbers down. But do stay tuned–fact is, this kind of development is exactly the charter for the MAST Center; to optimize–to push the technology, and bring that knowledge back upstream to design and manufacturing optimize the base NAND too!
Yes I know even with Sata 3 (somewhere in the future in laptops) it still will be only 600/750 MB throughput and possible need for Intel to change ICH into 45nm process (or even 32) to be both:energy efficient and deliver such transfer rates. That was only ‘wish’! But it would be nice to have option to buy such drive for desktop (of coz. drivers for XP/Linux/BSD/ etc.. are must) especially when price is comparable to 2xVelociraptor set in RAID (that’s around 600$) I assume that 256GB card would be ok. I would flood the market with this card (if there weren’t cost/production capacity issues)! Probably you’ll target it directly (and only?) to corporate/business market … and US-only but still after recent news about ‘graphite-based’ flash it seems that Flash/SSD is the only way for nearest few decades.
Regards
Chris
I have seen something like this (please correct me, if its not basicly the same) on fusionIO.com. They have, just the way you did, equipped a board with chips, attached an pcie connector onto it and have it work as a hard drive.
I think they are selling quiete well, since prices are not available on their homepage – so they are pretty sure hitting high as long as customers dont have a bit of a choice as of today.
When will the first compeditor to the fusionIO hit the market?
Greetings from Germany
Peter
[...] Read the review here! [...]
[...] Read the full Article Here! [...]
I noticed that the card was a 4x physical form, why not go full on 16x if, as said by Dean, “In fact, desktop buses don’t keep up, either. ”
This reminds me of the fusionIO, with both the high bandwith 600-800 MBs, and the 100k iops range. As I recall fusionIO was only able to hit 100k iops, but they were using 4k as a minimum block.
I was wondering how this device will pull away from the fusionIO as a distinctly different product? Are these detected as drives via bios, or is it just a drive that gets mounted post boot via a special device driver? or is it just a pure IO device which is accessed via software only?
While I see this as a must have for many database systems. And as a luxury toy for high eng gamers. What do you suspect the killer apps that will make this device indespensible?
[...] introduced a SSD in a YouTube video that supposedly hit a data transfer rate around 800 MB/s and this can later be increased to 1GB/s. [...]
[...] good on paper, we’ll have to actually see it in action to believe it. There is a short demo video demonstrating the technology, though it’s hard to take much from [...]
Hey guys–great questions/comments. The X16 PCIe slots in desktop systems are generally reserved for graphics usage. So while we could use that slot, it would likely result in limited market acceptance. X8 slots are much more common in the targeted server platforms, and the board held up by Joe is a X8 board.
One of the nice features of this board is that it is recognized by standard O/S drivers as storage. No additional drivers are needed. Of course this is not to say that additional drivers won’t bring additional performance, and drivers are certainly one area of research at the MAST Center, but we feel that it is also strength of the current board.
Right-on Chris! Improvements in the speed of SATA will arrive soon. Still, being limited to one channel and half-duplex, it won’t keep up with PCIe, which is going to Gen 2 now. BTW, this card does run with out-of-the-box drivers. No promises on pricing, but with NAND pricing going only lower, SSD solutions will definitely be looking more attractive.
[...] Micron?s Advanced Storage Blog, Joe Jeddeloh from the company demonstrated the potential of its advanced SLC (single level cell) [...]
[...] Micron’s blog, a video showing its ultra fast 1GB/s Solid State Drive (SSD). Unlike most SSD, Micron’s [...]
[...] im Blog veröffentlichte Video stammt von Micron-Mitarbeiter Dean Klein, welcher seinen Kollegen Joe [...]
[...] mit SSD f?r PCI-Express Micron hat in seinem "Advanced Storage Blog" ein Video ver?ffentlicht , dass die Leistung eines SSD-Prototypen zeigt, der ?ber PCI-Express angeschlossen wird. Die [...]
Did I hear 200,000 IOP/s. WOW.
[...] I must quote Tg Daily for it describes Micron SSD best: ‘IOP’S Like You Have Never Seen’ [...]
GTFO, 200k IOPS random!!! That is incredible. INteresting that SATA3 is holding back that much. Any idea what the MTBF will be for these devices?
Very interesting stuff and I am glad to see you have a blog. I recognized the name in the story from /. and I think this is good news for everybody. Can you imagine a Beowulf cluster of …
That PCIe card you have there is huge! The fusionio.com solution is much smaller and is already available on the market.
Also your IO benchmarks are worthless since your only testing 2K to 2MB file transfers. Anyone who knows something about IO, knows when doing such benchmarks the file size must be a minimum of two times the size of RAM in the system so it can’t cache the file during the test.
This is impressive. What would be much more impressive would be to build in data redundancy and in time failure reporting so that no one need ever loose data to equipment failure EVER AGAIN!!!
[...] SSD-Festplatten mit PCIe-Anschluss und 1GB/s hier nen kleines Video: Micron Advanced Storage Blog IOP/s Like You’ve Never Seen ich denke die sind nur f?r Pcie x8 weil sata II zu langsam ist =D naja sind schon geil die Teile, [...]
Can a card like this one boot an OS like windows without any other more conventional hard drive connected?
[...] diretor Joe Jeddeloh anunciou no blog da empresa um vídeo afirmando e demonstrando que a tecnologia atual de usar processadores com núcleos múltiplos como o [...]
If you guys ship this expect me to buy a few dozen at least
Any word on price per GB?
Great to see a competitor for the ioDrive. I genuinly hope you are able to deliver a product as good or better as Fusion-IO, for all us users and yourselves.
It’s great your card has OS (boot?) support out of the box, but if it still doesn’t hit the consumer market for a year or more, i fear you will be outrun by Fusion-IO, wich may already be launching a 2. gen pcie card by then.
Question: was the demonstration completely system-uncached (only internal card cache)? And what was the MB length of the read/write tests? I trust 2kb-2mb was the block sizes and not the test lenght.
Since the video is too low quality to read the numbers from the benchmark, could you post a screenshot of the test results? I am aware that these are not official or guaranteed numbers, but it would be fun to see.
What about drivers for Windows OS?
Does Windows see the card as one physical drive with logical partions v.s. multiple physical drives?
Reason: Even with multiple logical drives with their seperate read/write queus Windows still creates one I/O queu for what it “knows” to be one physical drive.
A high bandwidth app for Movies would see more benifit from 1Gb/s jump. Most databases need high I/O for random reads. Hence the need for many physical drives and drive letters.
Here a 20, 50, 100 or 256G drive would be great – IMO.
Does block size make a difference if they are using multiple banks for higher IO?
Cost? Ok, Santa ain’t bringing me one this year but…..
BTW check out the specs for USB 3.0. Can you imagine one of these in a thumb drive?
Kobie
We have X58 Motherboards carrying 3 PCI-E ver. 2.0 16X slots as well as 1-4x. Would like to try out your 4x, 8x cards (Beta or otherwise). Who may we contact? Do the boards carry 128 GB or greater?
Prickett: I am not certain what you mean by your system RAM comment. This is a standard benchmarking procedure, and indeed is specified by our customers. The benchmark we are running, IOMETER, does not depend on system RAM. If you are talking about transfers being larger than the SSD (or HDD) DRAM, you have a valid point, except that the address space for the file transfers is larger than the cache and RAM. 2K to 2MB refers to the IO request size. The tests were not running from cache or RAM. Again, back to the point: This is a demonstration of what the right NAND can do in the right conditions. ONFI 2.0 NAND, plenty of channels, a very efficient controller and a fast bus.
Roger, good question. The card we demonstrated cannot boot in standard systems without a hard drive. However, once booted, the system does not need any non-standard drivers to operate. That said–there are well-understood solutions to the boot problem.
Hey Kevin, we’ll mark you down for a gross! Pricing has not been set.
GullLars: Thanks for the feedback. Yes, the demo was run un-cached. Correct, the test length is much larger than the block size. Once again, our goal was to demonstrate ONFI 2.0 High Speed NAND performance.
KObie: Yes, we will be charging for these–even for Santa–and his elves can’t build them! After all, it’s not a toy!
Seriously, Windows XP and Vista support the drive with standard drivers.
Good comments on the I/O queue issue you bring up. We’re showing worst case with this demo of just one I/O queue. Regarding USB 3.0: It’s still just one lane of I/O and about PCIe-Gen2 data rates. So it will still be a fraction of the PCIe performance. Still, we do believe ONFI 2.0 high speed NAND has a place in USB drives and memory cards!
Hey dean, thanks for the quick answer.
I am aware that the transfer size usually is many times greater than the block size, but various tests on other SSDs have shown different transferspeeds for different block sizes for 50, 100, 200, 500, and 1000MB lenghts, so even if the transfer size is 100 times greater than the block size, the speeds may be affected by doubling the transfer size. The impact of this varies from SSD to SSD, and some show no difference at all at transfer sizes above 100MB, so maybe this is due to card cache, and controller firmware recognizing patterns.
This is the reason i was interested in the transfer size. As for the benchmarking choice, i agree IOmeter is good, as this correctly reads SSD performance and not just does spot-testing throughout the medium like other benchmarking tools may do.
As for a screenshot of the benchmark results, i’m still waiting for a link or decline to post
Lastly: That new kind of NAND flash you are using, ONFI 2.0, how are the number of writecycles? And how is the read/write speed and accesstime compared to regular NAND flash?
[...] Read | Permalink | Email this | Comments Post tags: Bandwidth, Competitor, Conjunction, Cto, Doodad, Engadget, Fri, Fusion, Iops, Micron, Nand Chips, Nbsp, Pcie, Per Second, Prototype Device, Sec, Shame, Sheer Performance, Ssd, Storage Posted in: Gad?ety | | [...]
[...] Read | Permalink | Email this | Comments [...]
[...] Storage SSD on PCIX….I want it NOW! Commenta 5th December , 2008 http://uk.youtube.com/watch?v=m1GBPr…ve-never-seen/ [...]
[...] appeared on Engadget on Fri, 05 Dec 2008 09:10:00 EST. Please see our terms for use of feeds.Read | Permalink | Email [...]
[...] Read | Permalink | Email this | Comments [...]
[...] Read | Permalink | Email this | Comments [...]
GullLars: ONFI 2.0 NAND uses the same cell technology as standard NAND. In this demonstration we are using ONFI 2.0 SLC (single level cell) NAND, which has the same endurance as standard SLC NAND and thus is nominally 100K erase cycles. Data transfer to and from ONFI 2.0 NAND is >4X the rate of standard NAND. TPROG and TERASE are comparable.
Hey Dean,
Nice video. It’s good to see casual videos showing such things. Might want to use a external mic for the next video, kinda had to crank it up to hear some parts. ?
A few comments and suggestions for your upcoming product..
I suggest focusing on the enterprise market first, then migrating down to the consumer is the wrong direction. HPC, clustering, DB, etc have steadily been heading to off-the-shelf commodity hardware vs the spendy stuff. Once 10GbE+ becomes mainstream, there is little argument for the majority of medium enterprise applications to keep sitting on $200k servers – somewhat thanks to VMware of course.. I’d rather you guys sell 500k of these @ $1000/ea vs 30k @ $10,000. Especially if that means I can get a couple for my home systems.
If you release a 8x and 16x version, please give some thought to slotting out the board to be compatible with 1x and 4x. For example, I have a Dell XPS 420 – it has (1) 16x slot, 4x, and 1x. Yet I have 3 PCIe GF 8800 cards. I simply used my dremel tool to remove the end out of the PCIe slot. I only use these as extra monitors, so throughput doesn’t matter to me. – Point is, let people decide if they are ok with a performance drop by using slower slots. Commodity hardware often comes lacking enough 4x-8x slots. There are also 1x/4x to 16x riser cards, but your product would have to be low profile.
DDR3 add-on option:
I’ve long wanted to see a PCIe card that would add extra memory to a system, even if that memory could only be used as disk cache or extended memory for niche applications. I would love it if your card came out with, say, 2 to 4 DDR3 slots on board. Each compatible with your (upcoming) 16GB DIMMs. Using even basic write-thru & read only caching, it would prove to be an incredible performance bonus. Imagine a 128GB SSD with 64GB of DDR3 as a cache?
I assume you are building in either a 10GbE or Infiniband type interconnect for these. You’d certainly gain some respect from the commodity clustering fans if you offered a widely available (http://en.wikipedia.org/wiki/Distributed_shared_memory) DSM implementation that not only offered high speed SSD, but DDR3 class memory across the network. Violin does this today, but at a price that few want to touch. I’d love to see a hybrid card like this for enthusiast users, that would instantly penetrate the higher end HPC and enterprise environments.
With PCIe 2 16x, or more likely PCIe 3.0 8x-16x – it would be pretty easy to make that DDR3 memory available as system memory. Contrary to popular belief, not all IT admins want to buy servers @ the price premium that goes along with having 16-24 DIMM sockets. I’d rather buy commodity dual i7 boxes, max them out, and drop in 4 of your cards with 4*64GB(256GB) of ram. It’s likely that VMware would be happy to build in proprietary support for extended memory of this nature.
Another idea:
There is a considerable market of people looking to give their older PC’s a performance boost before sending them to PC heaven. A couple thoughts for you here. It would be a low profit product, but perhaps help you sell more memory to augment sales of the higher end parts.
A PCI version – For older PC’s.
A miniPCI version – For laptops.
An ExpressCard version – Hopefully the PCIe variant of it. But of course, be backwards compatible if possible.
Even if this had only 32GB for the main drive, a windows swap file driver, and a ReadyBoost accelerator. It would make a lot of people happy. One of my clients has a fleet of old P4 machines running Vista. They are limited to 1GB ram and they refuse to upgrade. But if I could drop a $150 PCI SSD card in each that took over both main storage and swap file / ReadyBoost IO, you’d have a sale.
Who knows, maybe you could develop a single module that adapts between the mini-pci and the full PCI version to lower costs. SSD’s on laptops don’t really offer a great bonus considering the SATA bottleneck. I feel there is a good market for those of us wanting to give even old laptops a refresh, even if they don’t have the latest bus.
mvrx,
Thanks for both of your comments! And I’ll say–a Dremel tool to the slot? You are my kind of user! We’re looking at a lot of options for this technology and appreciate your input. Slots are definitely a factor with this type of hardware. Most mobos don’t have the available high-bandwidth slots that can really show of the benefits of high-speed NAND. Most hard core gamers are going to fill up both x16 slots with graphics cards. For most mobos that means we end up in a x1 slot, which is only a bit better than SATA. You mentioned the newer I7 boards and these do have a x4 slot–which again makes life interesting. In the enterprise, x8 slots seem to be plentiful. Hanging off a network interface also seems to really limit the performance potential.
DDR3 on PCIe? Really? Really low latencies and no-holds-barred bandwidth, no write endurance limitations, no blocks or pages. Dang, sounds attractive. Oh darn, just bumped the power plug and lost it all!!!
Regarding your second comment–I really like your ideas and want to let you know that they fit right in. After all, we are in the memory business at a time when low profits would be better than what we’re getting right now!
Seriously, I think you’re expressing what a lot of really clever people are thinking–NAND isn’t just for MP3 players and digital cameras anymore and there are a lot of really cool ways to get NAND into computing. It’s exciting new NAND technologies like high-speed ONFI 2.0 NAND and our new hot Enterprise NAND that make a lot of these ideas attractive.
DDR2 or DDR3 on PCIe (bootable and recognized from start like a normal HDD) would be a very good idea. DDRAM has enormous read/write cycles. Power isn’t such a big problem, why all drop this issue like a major impediment? You could have a UPS, or power by a capacitor like this http://www.511tactical.com/lightforlife.html or even a battery. This PCIe card could also have a SATA plug and while boot system would use this plug and after OS is loaded the card could switch to the PCIe slot (probably with a driver). We don’t need !!! big capacities! 8 GB is enough for OS (it could also be strip down with nlite or vlite). I would pay more for a 8 GB card than the price for a HDD. Almost no seek time and big speeds for read write is the key here. Normal data could stay on a partition of a normal HDD. How much could cost a PCIe card with DDR2 slots anyway? MRAM would be very nice but now it is very expensive and very low capacity. Why nobody make something like http://www.theinquirer.net/inquirer/news/581/1036581/return-ram-disk ?
Hi Dean, thanks for the reply. Always a good sign when a company not only reads their blogs but replies with some thoughts!
My reply to your comments:
Regarding Dremeling the slot;
I was suprised when doing some researching that nearly all higher slot sized products are perfectly compatible with minimum slots (at the cost of throughput of course). I read a thread of a guy on Anandtech who used a soldering iron to melt his, which scared me, so I Dremel’ed the end out of all my slots. I expected my system to not boot afterwords, but alas, there I go 3 16x cards on an OTS motherboard that only had 1 16x slot. Since the other cards are only used for multi-monitor and physix processing, I’m really not suffering any loss.. Anyway – to the point, if possible, consider building in an option to your future cards that allows them to use 1x 4x 8x 16x no matter the stock size. People can decide if the throughput loss is worth it. FYI – here is the non-hack official method, 1x to 16x riser. http://www.google.com/search?hl=en&rlz=1T5GGLL_enUS262US262&q=x1+to+x16+Slot+Extension+Adapter
Yup you’re right. Most users will use up two 16x slots for video if they have them. So – You’re on the right track – 4x and 8x versions of your card will be a winner. I just suggest you hit your engineers to see how difficult it would be to make them work in 1x-16x.
Re: DDR3 on PCIe
Yes! Really! I suggest you read up on modern caching algorithms (not saying that in patronizing way, even super-geeks aren’t familiar with the latest algorithms). There are so many ways you could use this even without any battery backup/capacitor/etc. Safe read-thru would never have an issue with power outage, performance/write cache cache mode would potentially lose some data in power outage (which is typically safe for even enthusiast users, not desirable for enterprise mission critical probably), and the highest most performance boosting mode would balance write caching with the full amount of DDR3 memory based on the user’s preference.. i.e. You need to write 30GB of data, the DDR3 accepts it – prioritizes it – writes it as flash speed allows.
Here is a good example, Vista **finally** introduced a more modern cache option for storage. The “enable advanced performance” isn’t recommended for mission critical stuff, but I’ve had it enabled on an array of servers for over a year and have yet to have any real problem. View this page and scroll down a bit http://www.vistax64.com/tutorials/193491-hard-drive-advanced-write-caching-fix-cannot-enable.html
You want the claim to highest throughput and IOPS, this will give you the crown for years, even if your reviewers have to write “With full caching, the data is less safe, but performance is 20x!!!”
For extra safety, drop a PCIe power connector on the back of the card, partner with an OEM to build a external PS for it. Market it with a minimum $50 UPS (a 350VA unit could probably keep it going for 30 minutes), and you’re set. If you wanted to be a bit special, build a good set of capacitors into the PS.. If you could even give your card 30seconds of power and a “power out” notification, it could clean it’s DDR cache instantly before shut down.
This also raises the VM style thought of your card being it’s own little virtual machine that the user could configure with preferences. If the card always had power, you could imagine such things as “Pre-cache OS even when PC is offline”, or “always cache most used files in profile x”. Or “Always pre-cache read-only flagged files”. Think about boosting Vista and every single file it typically uses for booting isn’t just ready in the SSD flash, but is available at the full speed of the given PCIe interface slot. WHAM. Finally my OS is booting at the CPU(s) absolute performance limit instead of my storage!
Another thought, aside from the option of making the DDR available as extended ram to the system, you release a API for video card OEMs, game makers, .NET/JAVA, Windows HPC, Linux clusters, etc – that makes the pool available as a “secondary memory resource”. It would be known as slower than system memory, but still an insanely fast memory “playground” for applications. Like a Windows Swap File that “doesn’t suck”. (This would also probably have ties to DSM standards)
…. and… you sell more cards.. You sell more DDR3 memory… you get more enthusiast / gamer sites freaking out and praising you…
If all the top 3DMark06 scores were made with a Micron DDR/Flash hybrid card, then all the extreme gamers are going to be buying the card.
RE: The laptop accellerator
Thanks. Adding a SD card, or SATA SSD to a laptop doesn’t really do the trick for readyboost or overall thoughput yet. But 95% of laptops out there have a empty mini-PCI slot. If you could bring a <$100 mini-PCI card to market, get the word out, you’d have a big seller. We have a 133MB/s(32bit)/266MB/s(64bit) speed limit, or maybe 100MB | 200MB/sec real world, but that would beat the holy heck out of a SD slot or SATA connected flash device (IOPs being the real performance bonus). Of course, if production similarity cost savings made sense, you’d want to make the same part for PCMCIA/Expresscard/USB3 variations.
Re pdanyels:
heh ). And I’ve always wondered why this has never been seen for the PC in 13 years. Obviously Windows has always kept the market back by waiting for the chicken never giving the egg to the market (meaning, no industry standard for effectively using extended “slower” volatile add-on memory), but if a vendor (hint Micron) could come out with something that the enthusiast got (bottom up, not enterprise top down) to use, provided open source API/drivers, we’d surely see a (re)surchance in the idea of massive extra memory in PC’s boosting all sorts of performance arenas.
Hey great links, I’ve been researching product development on PCI/PCIe ram caching (with and without power backup / super capacitors) and I never saw the links you posted. I remember adding 8MB of ram to my Amiga computer in 1995 via a card (stop laughing, it was freakin cool in 1995!
One last note for you Dean (sorry, I’m honestly not trying to spam your blog m8) – If you can’t engineer adding DDR to your SSD card today, consider a secondary “pure” DDR card with an SLI/Crossfire-style interconnect to your SSD card. That could give you 9 16GB-capable slot = 144GB. Imagine your 132GB SSD with a *FULL* DDR speed cache.. Lets say by 2010 DDR3 is $20/GB. Put yourself in the shoe’s of an enterprise server admin who’s DBA’s are demanding more performance for the DB. $3k to quadruple the read performance of a DB server? Not worth the time debating the expenditure with management. Order it! Upgrade it!
As for your SATA interface idea on the card – I’ve been workin on that idea for years. Can’t post them here, but yea – it’s a long overdue feature for the consumer.
We want this http://www.acard.com.tw/english/fb01-product.jsp?prod_no=ANS-9010&type1_title=%20Solid%20State%20Drive&idno_no=270 but on PCIe and at a good price. Anybody? Hello? Somebody
Re pdanyels
Thoughts my thought. If I’ve got a PCIe SSD, and it’s pulling 800MB/sec, why can’t it include a OTC RAID 0/1/4/5/6 controller and not only act as a SSD, but an advanced cache for the larger discs? Obviously Micron would want to promote a flash only storage world, but we’re years from that. That flash (and hopefully DDR cache) would provide a huge performance bonus to the disk storage.
There have been companies like the above doing DDR Sata drives for a couple years now. They are kinda cool, and I’m impressed they’re up to 64GB now. But back to the old problem, they are on SATA, negating the true potential.
I’d like to see the same idea – but the PCIe SSD/DDR card acting as a super cache for big fat drives.. and not in an EMC style overpriced “enterprise only” approach. Start at the enthusiast market and move up!
I do like the idea of 5.25″ modules however. Would be interesting if someone came out with a PCIe 4x/8x/16x card that had multiple PCIe cable ports that you could connect to a bunch of 5.25″ modules. Imagine having that with an external cable and adding module after module to a big fat OTS external server case. Can’t really do it with 10GbE, but you sure could with PCIe cable – and boy could you scale baby….
http://en.wikipedia.org/wiki/Cabled_pci_express
Idea: PCIe card with ddr slots. On one side slots to hosts volatile ddr memory sticks, the other side slots to hosts non-volatile flash NAND type sticks. The user could decide what amount of sticks to put in this card (budget: cheap or expensive). When out of power, the auxiliary power could be used to backup volatile sticks to non-volatile sticks at a speed of 200 MBps (you do the math how much time do you need to backup). If cheaper, chips could be binded directly on PCIe card. The slowest part of a computer is the storage part and nobody do a thing about it. CPU’s speeds have grow up by the Moore’s Law, a.s.o., but the storage speed (except space) it is like it stayed on the same place.