Clever caching fuels Samsung 840 EVO SSD

Samsung just announced its 840 EVO solid-state drive. The initial press release covered the basics, but it was a little short on the nerdy details we tend to obsess over here at TR. We’ve since learned much more about the drive, and we’ll have a full review soon. In the meantime, we can share a few interesting tidbits about the EVO and its unique approach to caching.

Improved write performance is one of the EVO’s defining characteristics. The boost in write speeds is made possible by a feature called TurboWrite technology, which is essentially a high-speed write cache built into the NAND. Most of the EVO’s flash memory is configured as TLC NAND with three bits per cell. However, the portion dedicated to TurboWrite is addressed as one-bit SLC NAND. The single-bit flash offers higher write performance than TLC memory, but it has only one third the storage capacity per cell.

TurboWrite sounds similar to the nCache scheme employed by SanDisk’s Extreme II SSD, but there are a few key differences. The Samsung implementation is a pure write buffer; it caches all incoming host writes regardless of whether they’re sequential or random. Cached writes are only moved to main storage during idle time, though host writes may be sent directly to main storage if the TurboWrite buffer is full. The size of the write cache varies depending on the drive’s capacity. The 120 and 250GB models can cache up to 3GB of data, while the higher-capacity models have even larger buffers.

The 840 EVO has a second layer of caching enabled by Samsung’s SSD Magician software. Dubbed RAPID mode, this scheme uses system DRAM to accelerate system performance. When enabled, RAPID mode allocates up to 1GB of system memory to storing frequently accessed data. This software cache is also used to "optimize" host writes before they hit the SSD. Samsung showed RAPID mode dramatically improving the 840 EVO’s read and write performance in a couple of synthetic benchmarks, and we’re eager to try the feature ourselves.

RAPID mode isn’t just for the 840 EVO. The feature is also coming to the existing Samsung 840 Pro. You’ll need to be running Windows 7 or Win8 to use the feature with either drive.

The 840 EVO is slated to replace the standard 840 Series, and it will be priced accordingly. Expect to pay $110, $190, $370, $530, and $650 for the 120, 250, 500, 750, and 1000GB models, respectively. Like its predecessor, the 840 EVO has a three-year warranty.

I’m already late for the next event on the agenda at Samsung’s Global SSD Summit, so I don’t have time to discuss the EVO further. I can, however, provide a couple of morsels about the NAND. The 128Gb chips are fabbed on a 19-nm process, which means they’re likely larger than the 16-nm flash Micron announced earlier this week.

Samsung hasn’t released an official endurance specification for the chips, but it says they’re surviving about 3,700 write-erase cycles under current testing. That sounds pretty high for TLC NAND fabbed on a sub-20-nm process. Samsung makes a lot of NAND, though, and it selects only the best chips for use in its SSDs. This SSD-grade flash is claimed to have 20X fewer bad blocks than the firm’s "normal" flash.

Comments closed
    • Glycerin
    • 6 years ago

    I have a 128GB 840 PRO and I love it.

    • NatashaSpence00
    • 6 years ago
    • MadManOriginal
    • 6 years ago

    This trickery will definitely require better testing methods to explore how it works and when and how it impacts performance. I expect high quality sites like TR and Anandtech to go in-depth, but I worry that all the scrub websites out there who review with a handful of synthetic and/or way too basic ‘real’ tests will be post some non-realistic review information.

    • sparkman
    • 6 years ago

    > The 840 EVO is slated to replace the standard 840 Series

    Replace the plain 840 or replace the 840 Pro? Big difference.

      • nico1982
      • 6 years ago

      Did you actually read the article? 😀

        • Growler
        • 6 years ago

        Who has the time for that? 😉

        • sparkman
        • 6 years ago

        Obviously yes, since I quoted the article and asked a question about it…

        But thanks for wasting everyone’s time by insulting me pointlessly.

      • sparkman
      • 6 years ago

      For those who didn’t notice.

      What the articles don’t really discuss is how this fancy new write cache compares to the pro model.

      Is the pro model obsolete now, lacking a write cache? Apparently not, but why not?

      Anandtech has some more details here:
      [url<]http://www.anandtech.com/show/7150/samsung-launch-the-840-evo-up-to-1tb-and-faster-writes-for-120gb[/url<] The EVO and PRO models around 256 GB have the same write speeds. But my impression is that the PRO model lacks a write cache. Does this mean another, faster PRO model is on it's way featuring a write cache?

        • sparkman
        • 6 years ago

        Also the EVO is available in sizes up to 1 TB whereas the Pro tops out at 512 GB I believe. Again leading to my original question, does EVO replace the Pro? Apparently not, but why not?

          • nico1982
          • 6 years ago

          Look at the price 🙂

            • sparkman
            • 6 years ago

            wow, you fail

    • brucethemoose
    • 6 years ago

    Superfetch is similar to RAPID, and is essentially the same as fancycache. Samsung already bundles similar software on thier laptops.

    I’ve been usng facycache for awhile, and as long as you have RAM to spare, it’s great. All the writes get buffered in RAM to turn random i/o into one, long sequential write. It speeds up loading times (even on an SSD), and it’s great for taming all the bloatware that comes on laptops with slow HDDs.

    If you want a preview of rapid, just read the review of fancycache

    [url<]http://thessdreview.com/our-reviews/romex-fancycache-review-ssd-performance-at-13gbs-and-765000-iops-in-60-seconds-flat/[/url<]

    • ALiLPinkMonster
    • 6 years ago

    Can’t wait for that review. Gotta love how just a few years ago, having a 60GB SSD was like driving a Lamborghini. Now if your OS is on a disk, you’re doing it wrong.

      • danny e.
      • 6 years ago

      My entire company is doing it wrong! *sad face*

        • derFunkenstein
        • 6 years ago

        Mine too, and 5400 laptop drives to boot.

        Then again they gave me a system with 8GB of RAM and Windows 7 Pro 32-bit, so there’s a lot of stuff corporate IT is doing wrong. lol

    • Symmetry
    • 6 years ago

    RAPID mode sounds like what the OS does already (or at least does on Linux, I don’t know that much about how the Windows filesystem works).

      • Klimax
      • 6 years ago

      Actually, more limited as it might not have all info from system and is very limited in size as Windows can use any free memory (5GB or more not a problem). Responsibility for caching in Windows is distributed between memory manager, file system driver and cache manager. (For all cacheable reads and writes – can be disabled for particular access to file)

      Interesting example is muxing of video file with files size over several gigabytes, which when moved to another drive can see virtual speed over gigabytes/s, because it goes directly from memory…

      • Aliasundercover
      • 6 years ago

      Yes, modern operating systems do a good job sharing memory between program needs and disk caching. Windows has done it since NT with incremental improvements in later editions. Linux does a superb job of it.

      A read cache in system memory will only interfere. Multiple layers of cache tends to result in using up more memory to store the same data again. At least the read side is harmless beyond wasting your memory.

      A write cache in system memory is dangerous. Caching writes without having the filesystem blow up on a sudden crash or power loss is hard. It is a task which keeps talented filesystem authors busy for years. Some writes must precede others, some others can be re-ordered for efficiency. Every bit of enforced ordering costs performance but failing to enforce it can turn your data to mush. An external layer messing with this is perilous.

      Some hardware does it safely enough by being separate from the rest of the computer, separate power which lasts long enough to finish writing and separate control so a computer crash doesn’t mean a storage crash. Neither of these can happen while sharing system DRAM.

      From what this article has to say RAPID mode looks like trouble to no benefit.

    • willmore
    • 6 years ago

    I really like this idea of being able to treat a block as either MLC-3, MLC-2, or SLC. A smart drive controller might even notice bit failures (thanks to the FEC codes) and downgrade a block from MLC-3 to MLC-2 or even to SLC. You’d eat into your overprovisioning, but you could drastically increase the useable life of a drive this way.

    Since each transition–MLC-3 -> MCL-2, MLC-2 -> SLC–involves a doubling of the margin between valid states, this might work. I wonder if it’s too late to pattent or if someone already has? 😉

    Edit: Spelling

    • jdaven
    • 6 years ago

    1 TB SSD for $650. Now we’re talkin’.

      • Firestarter
      • 6 years ago

      Are we waking people up yet? Something something wake me up when it hits $something.something

        • ALiLPinkMonster
        • 6 years ago

        Wake me up when top-shelf SSDs are $0.01/TB. I’ll be in cryo.

        • wierdo
        • 6 years ago

        The M500 960GB was selling for $570 already, so this milestone’s been broken months ago.

        This product may raise the performance bar for mainstream consumer class SSDs, but the prices have been pretty stagnant so far this year.

          • Firestarter
          • 6 years ago

          Yes but people would be arguing that the M500 has sub-par performance (which is totally beside the point). It’s a good thing that a performant SSD is getting down to this price range.

            • derFunkenstein
            • 6 years ago

            I would rather have a slower-than-the-best SSD than I would the very fastest mechanical drive as a boot drive. THen again I think that was also your point. :p

          • indeego
          • 6 years ago

          Did anyone actually buy this drive at this price? I never could find it there actually available.

            • wierdo
            • 6 years ago

            Sadly price is currently up to $599, but you can grab it directly from the manufacturer’s site:

            [url<]http://www.crucial.com/store/partspecs.aspx?IMODULE=CT960M500SSD1[/url<]

            • wierdo
            • 6 years ago

            Update: Just grabbed this drive for $560, saw a deal on it here:

            [url<]http://slickdeals.net/f/6180190-Crucial-M500-960gig-SSD-559-FS-NCIX-US[/url<] Not bad, around $0.59/gig, couldn't pass it up, that and the MLC class memory it uses.

        • NeelyCam
        • 6 years ago

        I think the magic threshold was supposed to be $1/GB, but people are still sleeping.

        I believe that when 160GB hits $80 or 256GB hits $100, things will start changing in a big way.

          • Firestarter
          • 6 years ago

          oh really? If so, the Samsung 840 at $175 should be changing things in a small way at the least

          • MadManOriginal
          • 6 years ago

          Bleh. Ever since OCZ decided to stop selling SSDs really cheap there haven’t been very many great deals. Now a mid-range SSD for $0.67/GB is ‘good’ whereas it used to be pretty common and $0.50/GB was good.

            • Stickmansam
            • 6 years ago

            Wished I had gotten more drives when OCZ had still been doing their deals.

            • Derfer
            • 6 years ago

            OCZ isn’t as relevant as you make them out to be. Their cheap drives have been shunned for awhile now due to reputation issues, while Crucial and Samsung have been duking it out unopposed for a long ass time on newegg and amazon with at least monthly deals. Check the top 10 selling SSDs on amazon. OCZ isn’t there, in fact they’re in 24th place. I blame poor product segmenting. Their serious drives cost too much and their cheap drives are shit.

          • Freon
          • 6 years ago

          *shrugs*. I’ve owned three SSDs now.

          • mcnabney
          • 6 years ago

          The magic threshold is when the SSD is big enough to replace a mechanical drive for boot/apps/games/work purposes and cost LESS than the CPU in the system. People that have had them before also had processors that cost around $300. Now a 480GB drive is getting ready to break the i5 barrier.

      • CaptTomato
      • 6 years ago

      Yes my friend that’s awesome, though I could do very well on a 7fiddy, but I’m struggling on my 250g 840 consumer.

    • Kougar
    • 6 years ago

    So how does the EVO handle a sudden power loss event? Will the data stored in the SLC cache be retained and written to the TLC normally after its powered back on?

    Compared to a normal SSD the EVO sounds like it now has twice the chance to lose data, since it could be writing to the SLC cache and the TLC NAND at the same time during a sudden power loss. I am assuming the EVO doesn’t offer built-in caps to finish any writes in progress like an enterprise drive.

      • Firestarter
      • 6 years ago

      [quote<]Will the data stored in the SLC cache be retained and written to the MLC normally after its powered back on?[/quote<] If it didn't, they might as well have used regular ol' DDR right? The whole reason Samsung is doing this at all is because they have confidence that these drives will be roughly as reliable as their previous offerings. Now, your second point is a very valid one, and I would love to hear what (if at all) they've done about it.

        • Kougar
        • 6 years ago

        Very true, but it doesn’t hurt to make sure. So many things one would assume like TRIM support via some types of RAID or PCie drives doesn’t actually work but for weird reasons. Like Windows not sending TRIM commands to any PCIe device even though there’s no issue stopping it from doing so.

        I’ve only done a week of intensive overclocking before I set it and forget it, but even that’s enough for Windows to detect issues on the SSD during a filesystem scan. Given overclockers are a primary buying group for SSDs it makes the issue of doubling the risk of data corruption all the more important.

    • Wirko
    • 6 years ago

    Clever but this means that the same data is writtent twice. However, it’s possible that writing in SLC mode causes much less wear than writing to the same cells in TLC mode.

      • demani
      • 6 years ago

      Yes- SLC has MANY more rewrites than TLC.

        • willmore
        • 6 years ago

        What Wirko is asking is if the operation to write a cell to the voltages SLC uses causes more or less wear on the gate oxide than doing a write to the voltages that TLC needs.

        I don’t have an answer to that. It’s hard to say if that is the issue or if the amount of leakage that comes from damaging the gate oxide at ever P/E cycle just makes it harder for TLC to maintain its voltage margins for the specified retention life. Hmm, that’s confusing and I know what going on. Let’s see:

        In SLC, you have two charge states 1 and 0. There is 0.5 margin that a state can vary by before it becomes confused with the other state. With MLC-2, you have 1, 2/3, 1/3, and 0 which leaves a 1/6 margin between adjacent states. That’s 1/3 the margin of SLC. For MLC-2, you have 1, 6/7, 5/7, 4/7, 3/7, 2/7, 1/7, and 0 for a margin of 1/14. That’s 3/7 the margin of MLC-2 and 1/7 the margin of SLC (hope I got that right).

        So, if the gate leaks at rate X, then it’ll take 7 times as long for SLC to error a bit than MLC-3. Honestly, I don’t remember if the gate leakage is modeled as a resistance or a current source, so it might be a bit different, then again, they might–for that reason–not use simple ratiometric charge levels in MLC cells and that would fix things. So, let’s stick with the linear model.

        We’d also need to know more about how P/E cycles effect the gate leakage–is it just a linearly increasing thing or is there some other effect going on. Assuming it’s linear, you’d expect SLC to have 7x the lifetime of MLC-3 and 3x the lifetime of MLC-2.

    • 5150
    • 6 years ago

    It’s been a while since I’ve used this term, but Samsung is pwning the SSD market.

      • spugm1r3
      • 6 years ago

      It should be pointed out, Samsung is offering a 1TB option at $650 simply because Micron already did it.

        • 5150
        • 6 years ago

        I never said I want Samsung to be the only competitor, but they are kicking everyone’s ass in price/performance. I hope everyone else steps up their game to keep them honest with their pricing.

        • Neutronbeam
        • 6 years ago

        Considering my first Intel SSD in 2008 was $600 for 80GB I don’t care about the reason–only the result…and I’m thrilled. Too bad I can’t afford the expensive ones any more.

    • chuckula
    • 6 years ago

    [quote<]The 840 EVO has a second layer of caching enabled by Samsung's SSD Magician software. Dubbed RAPID mode, this scheme uses system DRAM to accelerate system performance. When enabled, RAPID mode allocates up to 1GB of system memory to storing frequently accessed data. This software cache is also used to "optimize" host writes before they hit the SSD.[/quote<] Oh the number of things that can go wrong with this kind of setup.... [Edit: If you want a RAMdisk, use one but understand the downsides like sudden power loss. The use of "magic" proprietary software to play games with file access might look good in a synthetic demo, but can also have lots of downsides when it interferes with how Windows/Linux/etc. already cache files in memory.]

      • madmilk
      • 6 years ago

      Not to mention any modern OS/filesystem already does all of that.

      • bcronce
      • 6 years ago

      Yeah, I don’t mind read-caches, but write caches can be very dangerous.

Pin It on Pinterest

Share This