Slashdot Mirror
Intel's Braidwood Could Crush SSD Market
Lucas123 writes "Intel is planning to launch its native flash memory module, code named Braidwood, in the first or second quarter of 2010. The inexpensive NAND flash will reside directly on a computer's motherboard as cache for all I/O and it will offer performance increases and other benefits similar to that of adding a solid-state disk drive to the system. A new report states that by achieving SSD performance without the high cost, Braidwood will essentially erode the SSD market, which, ironically, includes Intel's two popular SSD models. 'Intel has got a very good [SSD] product. But, they view additional layers of NAND technology in PCs as inevitable. They don't think SSDs are likely to take over 100% of the PC market, but they do think Braidwood could find itself in 100% of PCs,' the report's author said."
When given similar performance but a slightly higher price, i would prefer the SSD. I can't take the flash to the next PC as i can do with the SSD. Hard disks have a highe life expectancy than mainboards (i usually find some good use for old HDs, i never did for old mainboards). Unless the SSD will cost 2-3 times as much as the flash on the mainboard, i believe SSDs will still be used. But maybe this will lead to lower SSD prices.
Sooner or later, no moving parts beats moving parts. The magnetic disk makers have done an amazing job so far, but eventually they're going to lose out to solid-state.
-jcr
The only title of honor that a tyrant can grant is "Enemy of the State."
Now only if they could start following the server side folks and place an internal USB connector inside and then MS and others could give us the OS on its own usb drive (read only) and we could use the hard drive for updates and programs we could enhance the security as well...
Life is but a Beta test...
What does it do, scream "Nooooooooo!" and throw itself underneath the hard drive in slow motion?
If you were blocking sigs, you wouldn't have to read this.
Sounds like a good plan. Throw cheap battery backed memory, 4-16Gb onboard to act as a transparent buffer between harddrive(s) and system.
Fast IO is ensured as most operations happen in memory, and dataloss isn't an issue as the memory is battery backed.
RAID cards have done this for ages, but it's becoming real option for desktops as memory price keeps declining.
16Gb might be overkill for most purposes, you could get away with 2 if the system is used only for low-power tasks like surfing and email.
There are no atheists when recovering from tape backup.
RTFA. It's cheaper than DRAM.
Random I/O is essentially uncacheable.
I'm sure that would come as a great surprise to anyone who ever implemented a virtual memory system.
-jcr
The only title of honor that a tyrant can grant is "Enemy of the State."
There have also been rumors, however, that Braidwood has been canceled, at least in the near term:
http://www.dvhardware.net/article37368.html
If the onboard flash is a cache, that means it will be used frequently do it will wear faster. Won't that mean you're more likely to corrupt your data, even if your HD is still good ?
First of all, DDR RAM is not cheap (at least, not compared to NAND RAM). It costs significantly more per gigabyte than even the most expensive of Intel's offerings for SSD's. While it should provide more theoretical throughput than any SSD, benchmarks at various places (http://techreport.com/articles.x/16255/1) haven't shown that to be significant yet, at least from the end user perspective (some synthetic benchmarks show that the RAM based disks can be faster than SSD's, but translating that to real world usage scenarios by consumers doesn't quite show any tangible benefits).
DDR RAM uses up a very large amount of power per stick compared to SSD's do. I remember seeing the power consumption of one of the DDR2 based "volatile hard drives", and it was higher than spinning drives (at least at idle), and wasn't particularly faster than the best of intel's SSD's.
So sounds like DDR RAM on board is expensive, power hungry, and doesn't provide that much of a tangible benefit to consumers. Tell me again why it's a good idea?
Random I/O is essentially uncacheable.
I'm sure that would come as a great surprise to anyone who ever implemented a virtual memory system.
His flaw lies in assuming- or implying- that most I/O *is* random.
"Slashdot - News and Chat Sites Deviant". (Click "homepage" link above for details).
Funny - this very thing was being discussed around 1985 (I think), but using battery-backed RAM as a way to reduce boot time. The thinking was people wouldn't put up with a computer that took 30 seconds to start, and if we didn't have a 2-5 second boot time (equal to a TV), the personal computer would never fly. But since it took from 1985 (80386 chip) to 1995 (Windows 95) for a 32-bit OS to become popular, maybe 25 years is reasonable.
Or not. Man, this industry moves at a snails pace in a lot of areas. Why do we still live with the x86 instruction set. Is "the year of UNIX" here yet?
Anyway, three competitors will emerge:
- Someone will put NAND directly on the drive, and get an instant speed improvement. All the tech sites will rave about it and it will be an instant must-have item.
- Their competitor will figure out a way to put the OS files in NAND, for fast booting, via a utility or firmware. The marketing war begins.
- The third competitor will work with Microsoft or Apple to get OS support for fast boot. Apple will get there first and you'll see a commercial on TV with the Mac guy wondering why the PC guy takes the entire commercial to wake up.
In a single drive system, the cost will be about the same. Doing it on the drive will create an instant performance boost on any machine, and well worth the estimated $10 added cost.
Place nail here >+
Exactly. I already have a disk cache. This solution is redundant. Also, this solution doesn't get me away from the mechanical spinning noisy hot slow thing which fails too often.
Fascism trolls keeping me up every night. When I starts a preachin', he HITS ME WITH HIS REICH!
Random I/O is essentially uncacheable.
I'm sure that would come as a great surprise to anyone who ever implemented a virtual memory system.
-jcr
You're both right.
The problem here is that "random I/O" can have at least two subtly different meanings. In the very old days they talked about random I/O as opposed to sequential (ie, tape) I/O. In that sense, yes, random I/O is often extremely cacheable, as you say. That's why virtual memory works, as system files, drivers, commonly-used applications, and so forth are accessed much more often than other daa.
"Random I/O" can also refer to I/O that does not follow any real pattern - ie, a 50GB database in which all records are accessed about equally as often. This kind of I/O is not really cacheable, practically speaking. Unless you can cache the entire thing.
What's the correct terminology for the second kind of random I/O? Random I/O with very low locality?
OtakuBooty.com: Smart, funny, sexy nerds.
It's more the case of hedging characteristics against each other.
1. SSDs handle random I/O extremely well compared to traditional harddisks.
2. Braidwood is essentially a small, cheap, 8-16GB flash based cache.
3. If Braidwood is transparent to the OS, it will have a hard time guessing what to put in the cache, because a lot of the I/O on a desktop/laptop is random, but the issue with caching the non-random part is that most OSs do caching themselves for frequently accessed parts of the disk. This means that for a transparent caching solution like that it is very hard to tell the difference between a frequently accessed piece of executable data and random I/O, since in both cases, it only gets accessed once per startup/shutdown cycle, for frequently accessed stuff it is already cached in memory, for random I/O, it is simply never requested again for a long time. So to make this caching work the flash thing either needs OS level support or very sophisticated statistics collection specifically tuned for keeping track of patterns across reboots and providing a caching solution for startup basically.
It takes a man to suffer ignorance and smile
Be yourself no matter what they say
> Your OS doesn't always have time to shut down properly. Don't think anyone's fond of the idea of having their last couple of saves go poof because Windows crashed.
So, what happens if my PC crashes because of some hardware failure and I have to plug in a different HDD for some reason? Or plug the HDD into a different mainboard? All the things I thought I wrote to the disk will be gone. In fact, the file system might be inconsistent if this thing doesn't honor flush requests. But if it does honor flush requests then nothing is gained, it'll still be the OS that does all the caching.
Well, it'll still be a great read cache, 4-16GB read cache is more that most people have as RAM caches, so it'll be good for something.
Today's experiment
Why a USB connector ? That causes the same problem as making SSD cards use the SATA interface - the serial interface becomes slower than the things it is connected to.
What I would like to see is a set of sockets on the motherboard, mapped into the main memory address space (not PCI), a physical switch on the board to make them read only and software in the BIOS to make them look like a bootable disk.
Four sockets with 16 or 32G in each would give you enough space to store the entire OS. I don't know how Windows would handle it, but in a Unix or Linux based system it would be fairly easy to mount the devices as read only partitions and map them into the filesystem. This would be ideal for a server system, mapping the entire OS into the main memory address space and making it read only.
In fact all the BIOS would need to do is make the first 100M visible as a boot partition, and leave the OS to handle the rest.
Stop using a typewriter to post on Slashdot.
Signed, everyone.
Well hopefully, there will be a BIOS option to disable this hardware in case a failure shows up. Better yet, have them removable much like the old COAST (Cache On A STick) modules of the first gen Pentium days.
Life is not for the lazy.
The buffer should obviously be on the hard disk. That way the data on the disk will always be in sync, even if there are writes buffered in the flash cache when the computer loses power. I can't see a good reason to put it on the motherboard instead. Especially as most consumer systems have exactly one HDD.
The article says that the flash buffer could work for "all system io". I can only think of optical disks and flash drives possibilities other than hard disks. But optical disks are interchangeable, so they have to be reread on each use anyway, and could just as well be cached in RAM. And it makes no sense to cache flash drives in flash cache...
I'd answered yes, but one doesn't control the fsync behavior of every application running on his/her system and the OS/file system can take a lot of time (even tens of seconds or more) before deciding to commit changes to the hard disk. Furthermore, a fsync may take seconds to complete and disaster can strike at any time.
There was quite a commotion about those matters when somebody filed a data loss bug against the new Linux ext4 file system in January 2009. It turned out that ext3 commits changes at least every 5 seconds and ext4 does it less often. Some applications that got lucky with Linux crashes on ext3 exposed their poor design when running on ext4. Comments #45 and #54 in the linked page are quite explanatory.
By the way that was a sloppy application coding problem (if you want your data safe on HD you fsync and wait as long as it takes to write them down) but they eventually issued some patches to the file system code to mitigate it.
Of course, because planned obsolescence has never been an issue before... especially with a corporation full of engineers.
If the onboard flash is a cache, that means it will be used frequently do it will wear faster. Won't that mean you're more likely to corrupt your data, even if your HD is still good ?
NAND flash chips are generally guaranteed for at least 100,000 erases per block. As I understand this Braidwood chip, it's a non-volatile ring buffer for data writes. Ring buffers are the easiest thing to wear-level, meaning you can just multiply the cache capacity devoted to writes (let's say 2 GB) by the longevity guarantee to get 200 TB of buffered writes before any failure occurs. And not all blocks on a flash chip fail after the same number of writes; you'll just start to lose ring buffer capacity over time.
Is this the latest FUD? That if a company brings out a successful product that's priced cheaply it'll "erode the market"?
How did the :"market" become so sacred that it must be preserved at all costs by keeping prices high? It's really funny the crap that'll come out of an MBA's mouth. He'll be all for "free markets" until someone comes along with a better product and then he'll start to squeal that the "market" is under siege.
Good for Intel.
You are welcome on my lawn.
Yes, it looks like they are planning to use a dimm-like slot.
http://www.hardware.info/en-UK/news/ymiclpqWwpyaaJY/Computex09_Intel_P55_motherboard_gallery/
Well, obviously the volatile drives aren't much faster than Intel's SSDs. Most SSDs are already starting to bump against the upper limit of what you can get out of SATAII when doing sequential reads.
The first ones I saw were for the PCI-slot and that one is limited to 133 MB/s and 266 MB/s for 64 bit PCI, both of which are lower rates than SATAII.
PCI Express of course starts at 250 MB/s per lane and tops out at 1 GB/s per lane for the latest version. Compare that to DDR3 which peaks at 12.8 GB/s per channel. To saturate a PCIe x16 lane we could settle for three DDR3 channels.
Size is another concern of course, as most of these things tends to go for sockets to plug the memory into.
So, you could try to top out a system with 160 GB of DDR3 RAM (would require 30 blocks), costing $14,099.7. And I'm not entirely sure, how you'd fit 30 blocks of RAM onto a single PCIe card, even if it's full length. This setup would obviously only be performance limited by the PCIe bus and the card's memory controller.
Now, HP StorageWorks' IO Accelerator 'only' provides about 700 MB/s depending on the workload, but only costs slightly more than half of the DDR3 solition at $7,700.
The biggest problem with the PCIe-based volatile solutions is fitting enough memory to be useful and that you're fucked if there's a bad power outage. The non-volatile PCIe solutions' biggest problemt hey're hideously expensive compared to regular SSDs and the only advantage they have to RAID-0'ed SSDs is the IO performance, as raw speed is faster if you raid a few of Intel's SSDs to a good controller.
And all the PCIe based storage mechanisms have one huge problem - non-bootable.
Just add the extra 32GB of RAM to the OS, and let it more intelligently manage the data.
I want to delete my account but Slashdot doesn't allow it.
It's that time of year. The one time every annum when I respond graciously to a request without becoming belligerent. So there you go.
Fascism trolls keeping me up every night. When I starts a preachin', he HITS ME WITH HIS REICH!