Micron Demos SSD With 1GB/sec Throughput

Lucas123 writes "Micron demonstrated the culmination of numerous technology announcements this year with a solid state disk drive that is capable of 1GB/sec throughput with a PCIe slot. The SSD is based on Micron's 34nm technology and interleaving 64 NAND flash chips in parallel. While the techology, which is expected to ship over the next year, is currently aimed at high-end applications, a Micron executive said it's entirely possible that Micron's laptop and desktop SSDs could have similar performance in the near future by bypassing SATA interfaces."

No SATA, eh?

[fuzzyfuzzyfungus]

SSDs built into mini-PCIe cards aren't new, so obviously they are possible(and I remember the concept going back as far as 44pin IDE drives on special PCI cards). Historically, though, these cards have appeared, from the perspective of the computer, as ordinary IDE or SATA adapters that just happen to have storage attached.

Does anybody know if this widget from Micron is similar, or are they actually pushing some new flavor of interconnect that will require BIOS tweaks and/or special drivers?

Oblig

[PearsSoap]

64 NAND flash chips in parallel should be enough for anyone!
I'm curious, what are the applications for this kind of disk speed?

Re:Yes that's nice.

[amazeofdeath]

RAID does not actually work that way. Yes, you can get increased speeds with certain RAID configurations, but this is a whole different beast.

Re:throughput IS NOT most important parameter

[myxiplx]

Trust me, throughput is still important if you're running these in a fileserver on a fast link (10Gb ethernet link, infiniband, fibre channel, etc). The read & write speeds of standard SSD's mean you need a whole bunch in parallel to prevent them becoming a bottleneck, which makes them hard to integrate into existing servers.

In contrast, a single fast PCIe SSD can drop right in. There's definately a market for high bandwidth SSD's in high end storage devices.

Re:Oblig

[im_thatoneguy]

Uncompressed HD, 2k and 4k film playback and capture.

At work we regularly are working with dozens of layers of 2048x1024 32bit uncompressed footage at the same time.

Re:Yes that's nice.

[lysergic.acid]

that's still not quite as impressive as 1600 MB/sec throughput [micronblogs.com] using 2 drives (which can be integrated into a single-card solution).

Drive on a card....

[Amphetam1ne]

Storage on an expansion card is nothing new, my Amstrad 1512 had a 40mb hdd on an ide card.

Re:Interleave

[billcopc]

You're implying that SSDs fail as often and disastrously as fast-spinning disk platters.

They don't, which is why a beowulf cluster of SSDs is a beautiful thing, though my concern is DDR2 can deliver much faster throughput and ns-latency, while the density trails a bit behind SSD but not that bad.

With 4gb DDR2 modules hitting the mainstream, and 8gb modules in the high end, what's stopping someone from putting a bunch of them on something like Gigabyte's i-Ram (minus the stupid SATA bottleneck) and having themselves a DIY uber-SSD ? Sure, there are differences but it's nothing a battery can't fix.

Re:Yes that's nice.

[TheRaven64]

The InPhase disks are $180 and the drives are $18,000. Unlike your external disk, the disks are rated to last 50 years. Not sure how much the Optware versions cost, but they start at 1TB and go up from there.

Re:Yes that's nice.

[nbert]

AFAIK no SSD apart from Intel's newest line provides any real advantage over spinning disks. They are faster in some areas, but in others they perform very poorly (write times for example). You'll get far more realistic numbers if you specify a real file in of. Here is the difference:

Desktop nerdbert$ dd if=test.zip of=/dev/null
136476+1 records in
136476+1 records out
69876088 bytes transferred in 2.249553 secs (31062211 bytes/sec)
Desktop nerdbert$ dd if=test.zip of=Herbietest
136476+1 records in
136476+1 records out
69876088 bytes transferred in 3.291721 secs (21227829 bytes/sec)
Desktop nerdbert$ dd if=test.zip of=/dev/null
136476+1 records in
136476+1 records out
69876088 bytes transferred in 0.876336 secs (79736653 bytes/sec)
Desktop nerdbert$ dd if=test.zip of=/dev/null
136476+1 records in
136476+1 records out
69876088 bytes transferred in 0.843004 secs (82889392 bytes/sec)


The last two runs illustrate that caches speed up the process, so a real file should be even slower than in this example.

Re:Yes that's nice.

[Yvan256]

In that context, then yes I do see how that would be a huge advantage.

I was looking at this as an average /. reader, as you say.

Bottleneck removed

[w0mprat]

This would be the first time a storage device would significantly saturate system memory bandwidth.

Indeed Intels SSD has a internal NCQ like command queue system to mask latency of the host. Common storage controllers are (obviously) not up to the job.

1gb/s from a single drive, that finally brings storage speed back in line with moore's law, which only capacity has followed it seems.

Re:Yes that's nice.

[Glonoinha]

There are a few videos on youtube of guys that RAID'ed iRAM's showing just insane performance.

If it weren't for the cost of adding four of these (plus four 1G sticks of pc3200 on each) I would have already scored a similar rig - but right now I'm working on a limited R&D budget. Maybe next year.

That said - these are really, really sweet - but I have to ask whether the RAID'ed iRAM or the new Micron SSD can hold a candle to a ramdisk ( see also : http://www.ramdisk.tk/ [ramdisk.tk] ) - I figure on a machine that can actually address 8G or more of memory (likely : Windows 2003 Server based) and use a massive chunk of it as a ramdrive - which is going to come out ahead?