Samsung 256GB SSD is World's Fastest

i4u submitted one of many holiday weekend slow news day stories which starts "Samsung Electronics announced today the world's fastest, 2.5", 256GB multi-level cell (MLC) based solid state drive (SSD) using a SATA II interface. Performance data of the new Samsung 256GB SSD features a sequential read speed of 200 megabytes per second (MB/s) and sequential write speed of 160MB/s. The Samsung MLC-based 2.5-inch 256GB SSD is about 2.4 times faster than a typical HDD. Furthermore, the new 256 GB SSD is only 9.5 millimeters (mm) thick, and measures 100.3x69.85 mm. Samsung is expected to begin mass producing the 2.5-inch, 256GB SSD by year end, with customer samples available in September. A 256GB capacity is getting large enough to replace hard-drives for good — now just the prices just need to come down further for large capacity SSDs."

MLC, not SLC.

[Anonymous Coward]

But it's a Multi-Level Cell based Flash drive, not a Single-Level Cell based Flash drive. The cells hold 4 states, not 2.

High capacity, yes, and apparently high speed as well. Excellent... but also lower reliability. SLC Flash is extremely durable these days, but MLC Flash is not, last I checked, even one tenth as long-lasting.

How much lower? Well... ...frankly, we don't really know yet. We won't really know, as such, until they start to die - which could well be 5-10 years, and if so, that's really not bad - and you might not see the same type of can't-write-blocks failure, but rather a more conventional can't-read-blocks failure. Which would be about as bad as a hard disk crash (and we might have to develop whole new data recovery techniques).

Maybe it might last years longer than a hard drive owing to fewer moving parts. Perhaps it will slowly die, but good write levelling will largely mitigate the issue and overall it'll come out better, or about the same. Or perhaps we're looking at a flaky brick with lower reliability than a Quantum Fireball.

Early adopters, start your engines, because someone's gotta find out.

For enterprise use, it might be wiser to stick to more conservative SLC flash. Past that, all bets are off.

But we're seeing the beginning, here. Hard drives are, slowly, on the way out. It'll be a long phase-out where they are much more cost-effective for a long time... but it is coming. And I, for one, welcome our new nanosecond-seek-time overlords.

Re:A 128GB SATA drive costs $3,000

[Anonymous Coward]

You are comparing to the wrong technology. SLC drives are extremely expensive, but this is a MLC drive. You should look at these items instead to extrapolate cost:

Super Talent MLC Solid State Disks on Newegg [newegg.com]

The Samsung drive is much higher performance than these, but the fabrication costs should not be too far off. I'd guess $1500 for the 256 GB model when it comes out.

Re:Random write ops?

[Fweeky]

No, SSD's have always shined at random *reads*. Small random writes have traditionally been where they're very weak; you might manage 160MB/s writing large chunks, but if you're droping 16k blocks all over the place (as, e.g, databases are apt to do) you'll be lucky to manage 1MB/s because of the overhead each write incurrs, certainly on cheaper drives aimed at portable use.

Hence, it's a perfectly reasonable question; depending on how they've implemented it, they could be anywhere from 20-20,000 random writes/sec.

Re:Random write ops?

[GreyWolf3000]

No, SSD's have always shined at random *reads*. Small random writes have traditionally been where they're very weak; you might manage 160MB/s writing large chunks, but if you're droping 16k blocks all over the place (as, e.g, databases are apt to do) you'll be lucky to manage 1MB/s because of the overhead each write incurrs, certainly on cheaper drives aimed at portable use. Hence, it's a perfectly reasonable question; depending on how they've implemented it, they could be anywhere from 20-20,000 random writes/sec.

To expound, NAND flash chips are broken up into 128KB 'blocks' which in turn comprise 64 2KB 'pages.' You can read any page you want on the entire chip in the same amount of time (no moving parts), but to *write* any particular page, you need to perform an erase on the *entire* block. Let me repeat, to write one page to NAND flash (2 KB), you have to erase a 128 KB block. The reset operation will transition all bits on that block to a 1, and you have to go back and tell it which bits to set to zero.

Re:Random write ops?

[TheRaven64]

Unless flash has changed a lot since I last looked at it, this isn't quite true. To write a 2KB page, you first need to erase a 128KB block, but you can then write into the remaining 126KB without needing an erase first. If you're using something like LFS or ZFS on your disk, this translates to very fast writes. If you're using a filesystem that doesn't have copy-on-write semantics, then this will still be quite slow because you will rarely encounter this kind of access pattern (you will be more likely to read 128KB, modify 2KB in a buffer and then re-write 128KB, which, as you explained, is very slow since it requires a 128KB read followed by a 128KB write just to write 2KB - the same reason that RAID-5 is slow for writes).

Re:Summary

[TheRaven64]

And since it's using MLCs, expect to buy another one quite soon after. While SLC flash is up to the 100,000 to 1,000,000 rewrites level, MLC is still closer to 10,000. This, combined with the larger cell sizes on most MLC products means that it is likely to wear out much faster. This is why most flash manufacturers are only pushing MLC technology for consumer electronics type applications (e.g. MP3 players), where you want high density and low cost but don't perform many writes.

Re:Random write ops?

[GreyWolf3000]

What you said doesn't contradict what I said at all. You can definitely achieve much better performance by not scattering your writes; sequential access is very fast. For random writes, you're going to have to seriously buffer to avoid wasting erase cycles. Fortunately, the same algorithm that avoids erase cycles for performance is also the same algorithm that does wear leveling. I don't know too much about ZFS or how copy-on-write would help except on a purely abstract level.

Re:Large enough? No way.

[Firethorn]

Discounting microsoft bloat, when I look at stuff like this I think about people like my parents and grandparents.

You know, people who after having and using their machine for four years still has 80% of their 80GB HD free. Where the biggest increase in HD usage in the last year was microsoft patches.

For power users like me that DO get into games, video, and music on the computer, a 250GB SSD is enough to last quite a while. Heck, from initial build I'm likely to throw my OS and programs on the SSD and get the cheapest per GB HD(or two) for the rest of the stuff.

Going from 10GB to 100GB was 'Big', it enabled the start of consumer video. Going from 100GB to 1TB enables HD storage for the movies most people would watch in a month to a year. Current broadband speeds enables the downloading of HD streams in useful periods of time with a queuing system.

Basically, I'm saying that we've reached the point with HD storage that the majority of people don't need any more. They won't use what they already get on a bargain machine. It's like with CPUs. If you're not a power gamer*, the bargain basement machines will all run a cleaned up windows** and associated software with good speed. Or even one of the easier versions of linux and open office.

*My grandmother loved bejewelled. Mom does various solitares. They 'game' a lot, but their games aren't exactly demanding on computer systems.

Wear leveling

[tepples]

And since it's using MLCs, expect to buy another one quite soon after. While SLC flash is up to the 100,000 to 1,000,000 rewrites level, MLC is still closer to 10,000. This, combined with the larger cell sizes on most MLC products means that it is likely to wear out much faster.

MLC products use wear leveling [wikipedia.org] at the controller level to spread writes to high-traffic areas such as directories and extent maps around the physical medium. That's also why they store 256 GB and not 256 GiB: that's 7 percent for spare sectors. Even if this wear leveing is only 10 percent effective, how long does it take to write to all 256 GB of the drive 1,000 times?

Re:The serializing of the random read-writes

[Wesley Felter]

High-end SSDs have a RAM write buffer, but it complicates the controller significantly. I expect sophisticated controllers to trickle down to all SSDs eventually.