Getting new information from a major vendor isn't always easy. Once you get through the marketing pages and white papers, it can feel like everything has all been said before. And maybe it has. But by getting information straight from the horse's mouth, in a candid, informal conversation, sometimes new facts emerge or get said in new ways that make them more intelligible or pertinent.
This particular conversation between Tom’s Hardware and Kingston grew from the memory vendor’s ongoing evangelism meetings, called SSD Acceleration Programs, in which the company goes out and presents the technology to IT professionals to explain what it is and the benefits it offers. Attendees cover the whole spectrum of technical knowledge, but several questions seem to always emerge from audiences. We used those as the basis of the following conversation with Kingston senior technology manager Louis Kaneshiro and lead engineer Tony Chen.
Tom's Hardware: Hi, guys. I know the technical press, including Tom’s, has done a good job of describing the ins and outs of SSD technology. I’d like to see if we can tackle some of the big issues within SSD, like endurance and TRIM, in a little different way.
Louis Kaneshiro: That sounds good. I should throw in here, William, we launched the V+ drive just a few days ago, and that does support TRIM.
TH: Great—now I’ll have to upgrade my V-series. Thanks. Any more product placements you want to add?
LK: No, I'm good. [laughs] Just so you know, I'm more of a user first and a geek second. I know the geek route has been traveled many times on this subject. So I'm going to try to take it from a little bit more of a user perspective. Let's move out of the benchmarks and into the real world.
TH: OK, let's start with endurance. We know that early SSDs were at risk of failure. What's the reality today?
LK: Tony should jump in here, but it’s true. They do wear out—the NAND that's in your USB drive, the microSD card in your BlackBerry, your camera, all of that. It has a limited amount of program/erase cycles—write cycles—and sooner or later you are going to reach the end of that. When you do, essentially that drive will stop working.
Tony Chen: More precisely, when you reach the mean time of a drive—its data endurance—it’s not really a fail. It's not like the data is no longer there or the format has vanished. It just becomes read-only at the end of its life. You still can recover your data to other media. Most of our Kingston SSD drives carry at least one million hours MTBF, their average life span. We also cover at least three years in the drive warranty. That data endurance compares well against traditional hard drives. Actually, SSD has a better life span in the media and in the interface. And SSD has 1500 G of operational shock tolerance. You’re not going to find that on a hard drive. So however you want to look at it, SSD has a better life endurance.
TH: So SSD failure isn’t like an HDD head crash where I'd just lose everything. I don’t have to panic, even if the drive goes “bad.”
TC: Yes.
TH: So the risk is than that I won't be able to write to it. But there’s a linear relationship between endurance and capacity, right? As you double the capacity, you double its expected lifespan?
TC: Yes. Data endurance numbers doubled for the same about data read and written per day.
TH: Given that, how many years are we expecting from the current crop of drives?
LK: Well, that's the hard part. You almost have a sliding scale if we’re talking about client usage models to server usage models. They’re very different. The worst kind of writes that you can apply to an SSD are random. You will wear a drive out quicker that way. If all the writes are sequential, that's the best case scenario for an SSD. A typical client workload is probably a mixture of those—not all random, not all sequential.
TC: For example, with our new V+ SSD, the proper life is like this: usually, we say the effective read/write duty is about 20% of the power-on hours. With this, normal operation is 8,760 hours per year, and this allows you to read and write 20 gigabyte per day in operation. With these numbers, we know that an SSD’s expected product life is actually much better than a traditional hard drive.
TH: Let’s circle back to that difference in the effect on endurance of sequential versus random writes. If the drive controller dictates how every bit gets written to the drive and wears the memory evenly, why there is a difference between sequential and random?
TC: Let's put it this way. Most of the time with your hard drive, you need an operating system. The random reads and sequential reads have a major effect on the system’s behavior. When you boot up your computer, you are doing a sequential read. Same with hibernating and application loads. But a lot of times, you also need to access your information, the user data, and that's a random read because your data is spread out everywhere. Now, with a hard drive, the arm has to move. And with this SSD drive, there are no moving parts, so no chance of mechanical failure. So, compared to the hard drive, SSD would provide better performance.
TH: Does multi-bit MLC enter into this discussion? Is having three—or later, four—bits per cell going to change the endurance dynamic, particularly when weighed against SLC?
TC: Three bits per cell is already in the market, but you also have to look at density—32nm or 25nm. That provides the density for more stack. The use of two or three bits per cell is just a current trend. The NAND semiconductor industry is more interested in how it can provide maximum data per square inch. Typically, we are looking for more cell density than bits on this point.
TH: But will 3-bit have an impact on endurance? With more electrons being pushed through each floating gate, does that erode the oxide layer more quickly?
TC: No, actually, because right now all NAND has ECC correction. You’re talking about losing all data endurance, so you cannot recover it. You’re talking about the voltage converting in the cell.
LT: But Tony, if we were to implement today's 3-level cell NAND on an SSD, would the endurance of that be less compared to a current MLC product?
TC: I would say that would be partially true. With data endurance there are two different issues. One is how long it will last, and one is how to correct errors if they appear.
TH: OK, for people who might have missed it up until now, what is TRIM?
LT: OK. A characteristic specific to NAND is that you cannot write over a NAND cell that has previously stored data. For example, on your notebook hard drive, let's say you deleted your 2GB iTunes folder. All you’re doing when you delete is just deleting the pointer to that file, right? The data is still physically there on the drive, which is why you can do recovery and things like that later on. But let's say you've got a new batch of data that you're copying. The hard drive is going to look for a space, see that 2GB of free space, and just copy right over the top of it. On an SSD, that same thing couldn't happen—not quite that way. When you delete that same 2GB iTunes folder from your SSD, you’re still deleting the pointer, but when you try to write back over that area, the SSD won't do it. The NAND cells have to be empty before it will do that. As you probably know, this process is called garbage collection. Now, with TRIM, if you were to delete that same 2GB iTunes folder, Windows 7 would basically tell the drive that there's just been a deletion, so go ahead and clean that 2GB area up. Do it ASAP. Don't wait until I need to make a write. That way, the drive will never have to stumble over invalid data. Is that about right, Tony?
TC: Yes. And actually, TRIM was an improvement even before Windows 7. When you want to write data to a single page, you need to erase the whole block and move the data in that block page to another area. That's why SSD has this built-in write latency that inhibits greater performance. Many previous generations of SSDs, when they first come out, would hang for several seconds when you’d write to them. So even before TRIM was available, a lot of SSD controllers had to implement garbage collection in the firmware. Garbage collection means that when older data is deleted, it creates an indirect table to notify the controller and say, “OK, when you have free time, you can start to move those various data out, erase the block, and make it free for the next time the system wants to write to it.” The TRIM command really just helps the SSD do better garbage collection. It essentially provides that fresh, new drive performance without any degradation.
TH: What is the role of spare drive space in all this?
LK: Pre-TRIM, spare space was something people desired. Just five or six months ago, at the Intel Developer Forum, Intel was advocating leaving 10% to 20% spare NAND when you configured the drive. So if you purchase a 64GB SSD, it formats to 59.9GB. If you were to follow the Intel advice, you’d actually partition to like 52 for all that spare NAND. Then they said, once TRIM came out and became supportable on Windows 7, it wasn’t as big of a factor.
TH: A lot of drives now in use predate TRIM, and common wisdom always said that the more spare space you can leave, the better the drive will perform. So 20% is better than 10%, etc. But if all I’m doing is small writes, that shouldn’t involve a lot of data shuffling. Why do we need so much room?
TC: When you boot up, you load applications. When you surf the Internet, you store temporary files. When you use a real burst of memory, you need page caching from the hard drive. So hard drive space is really critical to the OS. It's my understanding that with Windows 7 right now, you need a minimum of 16GB just to install the OS. When you have a 40GB drive, you can figure out how much free space you’ll need. You also know how much memory you have in your system. You take those factors, put them together, and figure out how much free space the system needs for the best performance. This is a basic formula for any OS to run smoothly. Besides that, SSD wear leveling requires a certain amount of free space for use. I like that Intel came out with the tools for this even before the TRIM command. They have an optimized program that helps the data to move in the drive, and with the TRIM command, they can help the system without using additional utilities.
TH: Intel occasionally tells me that it consistently ranks at the top of the SSD field because it has the best controller and firmware. Since some of your drives are rebranded Intel models, can you explain what makes Intel’s algorithms better?
LK: The controller is the engine, whatever flash device you have. If anyone would understand I/O, it would seem to be Intel. I think it leveraged that expertise and in-depth understanding of its own NAND. I mean, not all NAND is the same. We do a lot of multi-sourcing with our USB products, and as we change from one NAND manufacturer to another, we see differences in performance. I think the Intel drive has an optimized controller designed to work specifically with its NAND.
TH: Not to be rude, but if you guys are basically putting your own sticker on Intel’s drive, what value can Kingston to bring to the product?
LK: Well, that's more of a business question. Think of how Intel traditionally gets into Company X. They come in a Dell or HP box, right? Maybe in a ThinkPad. They wanted our help to proliferate their drives, and we have a channel at the consumer, system builder, and corporate end-user levels. We have people that end up talking to the same Company X. I accompany a lot of those guys on these conversations whenever we have a new technology. But for us, Intel’s drives helped us launch with a very strong product into the SSD market. The win for Intel was that we were covering area they couldn't get to normally.
TH: And I like your mobile and desktop SSD upgrade kits. Very clean and easy. Whoever thought of putting that combo together, good job.
LK: [laughs] Actually, that was me. Not a lot of our competitors have something like that, and if they do, it's often not as well put together as ours. I remember our first conversations with an SSD partner—not Intel—and we actually asked if they had any plans to do this. They looked at me like I had two heads. Like, why would I do that? I asked, “What about customers that have bought an HP whatever and have no install CDs? It's on the drive’s recovery partition.” They just shrugged their shoulders and said that was the customer’s problem. Needless to say, they’re not a partner of ours, but things like that just convinced us that if we were going to do this, we'd have to do it on our own. The upgrade idea just makes sense for us. Corporate IT has their own standards and methods of cloning. We needed to make it easy for consumers.
TH: We obviously do a lot of benchmarking on Tom’s. What’s the official rule on how people should condition their SSDs for proper testing?
LK: For us at Kingston, it's no secret. We just take Iometer and run it through a 4K, 100%, random write test. Iometer will basically fill up the drive so every piece of NAND, essentially every cell, now has data, and then it's benchmarked at that point. That process levels off the drive very quickly. If you take a brand new SSD from Kingston, create a partition, format it, and run any kind of benchmark. You'll get a monster number. But as soon as you go back and level that off with something like Iometer, you'll see that number settle into a consistent result. And the number that we publish publically is that leveled-off number.
TC: I think we do a 5- or 10-minute Iometer test. You'll see the drive level off when you're doing that. Another way is to run a secure erase before the benchmark. From then on, the drive will be conditioned.
TH: Do competing brands use the same process when they state their numbers?
LK: Funny that you mention that. We do a lot of competitive analysis, but I don't think we do too much comparison to their marketing sheets. But based on the reviews I've seen, it seems like the numbers match up. You’d be pretty dumb not to because those numbers will drop pretty quickly.
TH: I have a cousin who borrowed my X25-M for his gaming rig at PDXLAN, thinking it was going to change his entire universe. After the event, he admitted that he really couldn't tell much of a difference, and I was like, “Dude, your game is going to be limited by your graphics card or CPU, not your drive. What did you expect?” So let’s clear this up. What are the top applications that people actually do need SSDs for?
LK: On the client side, for me, it’s all about boot times and being able to open up applications. Here in the office, we use Lotus Notes for email, which can be slow starting up. When I click on that icon to launch Lotus Notes from a hard drive, I'm waiting seconds and seconds and seconds for that password prompt to come up. With an SSD, it’s like I blink and it’s there. Or one common response from first-time users is, “Holy cow! My system’s booted already!” And we have all these little agents—anti-virus, firewalls, a bunch of stuff that the hard drive has to chew through even once it hits the desktop. For me, that's where SSD has really shined. Yes, with gaming, your levels will load quicker. If I'm playing Halo or whatever, the map will load faster from the drive, but the game play itself? I don't think that's going to make a big difference. Gaming like at PDXLAN, that's mostly about processor and GPU and RAM. From the corporate IT side, we've also heard a lot of good things on the encryption stuff. When we have corporate customers running encryption software, running it on SSDs seems to help with the transition from HDDs. In terms of the original image being created, we've heard numbers as bad as four hours to build an image and encrypt it. After switching to our SSDs, that dropped to one hour. That's huge.
TH: It almost sounds like we’re having an overclocking discussion.
LK: Yes! It’s a performance upgrade. I mean, you’re increasing the overall responsiveness of your system. It's akin to getting a processor upgrade. If you benchmark a system with PCMark Vantage and do nothing but change the HDD to an SSD, I gained over 40% overall system performance when I did this on my ThinkPad. It’s just like I overclocked the thing into orbit.
TH: I’ll second that. My main system used to take five minutes to boot. Now, with an SSD, it’s under a minute.
LK: If we can measure it in the amount of work you can get done, I think that’s a big deal. You’re not waiting for apps to open. I hang in HPC circles sometimes, and what’s the point of going with 24GB of memory and all this other good stuff when you’re still waiting on that hard drive? In the past, my solution has been to stripe four hard drives, but after one look at SSD technology, I saw that one SSD would smoke my four hard drive stripe. I just thought, man, this is the Holy Grail. And when I wanted even more performance, I found that SSDs in a RAID scaled just fine. So I don't know if there’s one killer app that I would pick for SSDs. Encryption is definitely one that we can point at easily, but I just think it's the overall system response. Like look at netbooks. That underpowered system is great for doing simple things, but to me it's almost unusable. Throw an SSD in one, though, and it gets usable really quickly.
TH: The president of OCZ told me that. He said I’d be stunned at the number of drives they sell that go into netbooks as upgrades.
LK: Yup, I wouldn't doubt it. People are doing whatever they can to upgrade to SSD. Throwing Vista or Windows 7 on a netbook is tough already on the processor. You've only got 2GB of memory. You’re making that thing work pretty hard, and SSD can really boost that, plus extend the battery life.
TH: You mentioned striping SSDs. What performance increase have you seen there? Does it scale linearly or taper off like hard drives?
LK: Oh, yeah. It gets up to about 600 or 700 MB/sec when we've messed around with an integrated Intel RAID controller. After that, adding any more doesn't really help. So technically, I guess, after the third drive. You can gain capacity after that, just not performance. But it does scale especially if you go with a discrete RAID controller. You can benchmark with Iometer and see the scores pretty much double with two drives. The bandwidth doubles. The random IOPS will pretty much double. At IDF last year, we had four of the E-class Kingston drives, and we were displaying 1,000 MB/sec sequential read speeds using a high-end Adaptec RAID controller.
TH: Are we being bottlenecked already by the SATA 3 Gb/s interface? Do we need to move SSDs onto a 6 Gb/s connection?
LK: I think you're going to need the controllers to support it. That's probably a bigger issue. Some drives already have the current SATA II spec pinned in terms of performance. Tony, how much better will SATA III make SSDS?
TC: For the next generation, SATA III, we still need to go back and look at the semiconductor road map. How fast of a raw speed can we get from the NAND flash? Also, the design architecture from the SATA SSD controller interface—how many channels can be accessed at the same time? Those will determine the upgrades necessary for SATA III in the future. Everybody is concerned about how we can accommodate 6 Gb/s transfer rates and how to accomplish that. I know Micron has already announced the first one. Everybody is looking to see how this will perform in the next two years.
TH: So I shouldn't get excited about product shipping next week.
LK: Yeah, I don't think Intel is talking about it on desktop boards until 2011. But it's a big deal. It means adoption from the big PC OEMs. At this point, it's more of a DIY play.There are several motherboards already supporting 6 Gb/s via add-on controllers. It's exciting, to be sure, but I think from a consumer standpoint USB 3.0 is a lot more exciting to me right now. On the USB side—I know this is a little off-topic—we're shipping 128GB and 256GB USB drives. A guy who's buying that isn't just going to copy a couple of files over. Try moving 40GB at USB 2.0 speeds. You can go get lunch, come back, and it's still going, right?
TH: Oh, sure! I put your drive in a $25 enclosure and—bam! All that speed ran into the USB interface brick wall. Going to USB 3.0 suddenly gets us a lot more speed for our external storage.
LK: Absolutely. And at that point, could it be bootable? Could you boot to an external SSD on USB 3.0? At 4.8 Gb/s? Yeah, you probably could.
TH: Great idea. So that's your next product?
LK: [laughs] Yeah, we'll see.
TC: We should mention power consumption versus HDD.
LK: That's true. SSD power consumption is great in a client workload, but we can all agree that the big battery suck on a notebook is not the drive. It's the screen, CPU, and GPU. So an SSD might get you a little more battery life—I think I once measured about 20 more minutes on a 3-hour Dell—but it's really a server thing. There, it's like night and day. If you compare a typical 15K SAS drive to one of our SSDs, at active load you're looking at not even a watt. That SAS drive will pull 10 watts. It's about watts per IOPS. How much work can we get done with one SSD versus you having to stripe 50 or 60 15K SAS drives. So the power consumption to do actual work just mushrooms beyond belief. SSDs run cooler, so you don't have to spend as much on temperature controls, and they just consume a lot less power. On the server side, it's a very big deal. We have a lot of customers that are power-constrained. They just can't pull any more out of the wall or off the grid. So this is a way for us to help with that problem. They can stay within their power budget and still get all the work they need done.
TH: Last thing: I’ve always wondered about these G-shock ratings on SSDs. 1500 Gs? Seriously? That’s like a 27-floor drop. I have a feeling that if I drop a drive from the 26th floor, it’s not going to matter if the NAND chips survive because the casing will be in about 400 pieces. And how do you simulate that—with a baseball bat? Do you run over it with your SUV?
LK: [laughing] Have you seen our video?
TH: Video?
LK: That's why we were laughing. We actually did hit it with a baseball bat. It was pitched to me, and I hit it, and the drive still worked. We ran over it with a car and it still worked. We actually took a blowtorch to it and it still worked.
TH: Shut up!
LK: That’s why I can't believe you said that. Yeah, we even hit it with a golf club. That didn’t work. We hit it hard enough that the NAND actually decoupled from the PCB. It was flying around all over the place. We made SSD bits with that one. Anyway, we subjected them to some pretty good stuff.
TH: So an SSD will withstand a base hit but not solid nine iron.
LK: When I hit it with a driver, I basically tore off the casing. It split in two, literarily sheered in the middle. So how much force is that? I assume the golf club made contact at approximately 100 miles an hour, and we could probably measure how heavy the driver was. Dropping it off a building? Well, we're probably up a good three stories here, so tossing it off of that just didn’t seem that dramatic. We’ve done it, sure. Multiple drops, actually. But why do the obvious when SSD can do so much more?














