Are solid state drives ready for the big time?
10th Jun 2012 | 09:00
We reveal whether SSD tech has matured enough to be considered mainstream
SSD technology examined
Storage is the final frontier of PC performance; CPUs and graphics chips have long marched to the drum beat of inexorable performance improvements. Not hard drives, however.
Granted, storage capacities have grown so fast you can now fit the combined wealth of all human knowledge on a £50 drive. Well, just about. But performance has only inched forward incrementally.
At least, it did until solid state drives rocked up as consumer hard drives around five years ago. Since then, storage performance has joined storage capacity - and processor and graphics grunt, for that matter - on that relentless, Moore's Law-mandated journey of irresistible progress.
For SSDs, however, it's not always been a smooth ride. To say the earliest drives delivered patchy performance would be borderline disingenuous. In hindsight, they weren't fit for sale, so rapidly and inevitably did their performance fall off with heavy usage.
More to the point, while the headline sequential throughput achieved by SSDs is impressive, performance in more realistic random workloads is usually a lot less spectacular.
Outright reliability has been suspect, too. Not only have several different generations of SSDs suffered from instability issues leading to classic blue screen PC stalls, but outright failures aren't exactly unheard of, either. The latter is a particular problem given that one of the main advantages of solid state storage over magnetic drives is meant be their lack of moving parts and, therefore, robustness.
Put all together, the early history of solid-state storage failed to live up to the hype or potential. And yet much progress has undoubtedly been made. The introduction of the TRIM command in Windows 7 went a long way towards providing peace of mind when it came to performance degradation. While few, if any, drives maintain absolutely consistent performance levels over time, if you buy a drive with TRIM support you can be confident it won't turn into a stuttering, useless paperweight within months.
But the question still remains: Are SSDs ready for mainstream consumption? Read on to find out.
First, lets start with a few facts. Back in late 2007, a 16GB first-generation SSD, was yours for around £300. Peak sequential transfer rates were around the 30MB/s mark. Sobering stuff, isn't it?
Today, £300 buys you 250GB, or thereabouts, of the finest SSD technology available, which is capable of speeds of up to 550MB/s. That's 15 times bigger and 18 times faster in less than five years. Do the maths and the destination in five years seems spectacular: Fancy a 4TB SSD capable of 10GB/s transfer rates for £300?
As it happens, we don't think either capacities or performance will explode quite so spectacularly. As we'll see, SSD technology is approaching maturity, so the pace of those early gains isn't going to be repeated.
But the main moral here is that SSD technology has improved beyond all recognition. The latest offerings are now good enough for us to recommend them unreservedly. What's more, we reckon they've reached a point where usefully proportioned drives are now genuinely affordable.
For us, the sweet spot is the current generation of 120GB and 128GB models. Prices start at a little over £100, with peak sequential performance as high as 500MB/s. That kind of capacity is sufficient for a full Windows installation, a range of applications and a respectably proportioned library of Steam games.
On a related note, with SSDs at around 120GB at this price point, we're less convinced by the proposition of hybrid drives or SSD caching solutions. Our money-no-object preference has always been pure solid-state hardware. We're happy to say the balance between cost and capacity now makes that a more plausible proposition - spending an extra £50 on a motherboard with a 20GB cache drive doesn't make much sense when you can have the real thing for £100.
As ever, the most important narrative arc in SSD selection involves controller chipsets. Over the years, controllers have been both the heroes and villains of the solid-state story. In the early days, the mere mention of the name JMicron was enough to pump fear into the hearts and insert visions of a lingering, stuttering death into the imaginations of PC enthusiasts.
More recently, SandForce has emerged as an unlikely hero, taking on and beating the big boys. The fact that Intel, too, has now given in to SandForce and started using its controllers tells you all you need to know about both the difficulty involved in developing a really fast controller chipset and its importance in a given SSD - even Intel couldn't manage it.
The lessons from the Intel-SandForce escapade don't stop there. One of the other significant pieces of the SSD puzzle is firmware. That's essentially the code used to programme SSD controllers.
Intel could very easily have taken both SandForce's controller chipset and its off-the-shelf firmware and knocked out a perfectly decent SSD. Instead, it applied its internal validation procedure to SandForce's kit. The result was a modified firmware that might just be the most reliable among the SandForce pack and a drive delayed by around six months.
With that in mind, our rough guide to the best controllers goes something like this: In pole position is the second-generation effort from SandForce, the SF-2000. In general, it's the SF-2281 you'll find in most consumer SSDs.
A wide range of manufacturers knock off SF-2000 based drives, including Corsair's Force 3 and Force GT models, the Intel 520 and OCZ Agility 3 and Vertex 3 series, among several others.
It's a native SATA 6Gbps controller that dominates the peak performance tables with its raw sequential data rates. The SF-2281 also scores in terms of maintaining performance regardless of overall capacity, with speeds of around 500MB/s for both reads and writes. Some controllers see write speeds, in particular, fall off dramatically.
It's also very competitive when it comes to 4k random IOPS. That may sound like a theoretical metric. But it reflects real-world performance more closely than peak sequential transfer rates, especially when it comes to compressible vs incompressible data.
For starters, it's rare - in reality - to shunt around large quantities of compressible data. Most of the really big file types, such as audio, video and images, are essentially incompressible. And the SandForce SF-2281 is competitive rather than dominant when reading incompressible data and especially writing it.
On the other hand, the SF-2281 is up there with the best when it comes to 4k random numbers, which roughly approximate the daily drive churn generated by a modern PC.
All told, the SF-2281 is a controller that's at least competitive in all areas and relatively dominant in some. If the SF-2000 series has a weakness, it's stability. Early drives based on the SF-2281 reportedly suffered from 'the blue screen of deathitus'. Subsequent firmware releases appear to have cleaned its act up. But the length of time Intel took to release its own version of the SF-2281 hints that quite a tidy-up job was involved.
Next up is the Marvell 9174, seen in Crucial's RealSSD M4, Corsair's Performance Pro series and a few others. By some of the headline metrics, it's a little off the pace compared to the SandForce SF-2000.
Typical peak sequential read speeds of 500MB/s or more are obviously competitive with SandForce drives. But at best you're looking at about 300MB/s for writes, and that's only for 256GB and larger drives.The 128GB versions are nearer 200MB/s, with 64GB drives down around 100MB/s.
The explanation for the Marvell controller's fall off at smaller capacities (this applies to other controllers, too) is complex and involves a technique known as interleaving. Controller chips have a certain number of channels, attached to which are the memory chips. If you have very few chips, you can't even populate each channel (typically, controllers have eight channels), and that means less bandwidth.
So far so simple, but even if you have a chip for each channel, you don't have the optimal solution. That's because each read or write request takes a certain number of controller cycles, let's say five or six cycles for argument's sake. The problem here is that the controller sits idle during those cycles - but not if you have additional chips per channel. In that scenario, you can request a write to chip one and while waiting for that to complete, put another request to chip two. In reality, it's a bit more complicated and the performance scaling isn't quite as linear. But you get the idea.
That said, the Marvell 9174 doesn't rely heavily on compression technology to achieve its peak speeds, which means it's closer to and sometimes even faster than the SandForce SF-2281 when it comes to incompressible data transfer. It's a similar story in random 4k access, with writes falling off dramatically with smaller drive capacities. The 256GB drives are rated at 50,000 write IOPS, which is excellent. However 128GB and 64GB drives only manage less spectacular 35,000 IOPS and 20,000 IOPS, respectively.
If SandForce and Marvell are the two biggest players in the customer controller market, there are a couple of competitive in-house efforts. OCZ recently acquired controller-maker Indilinx. Prior to the arrival of SandForce, Indilinx's Barefoot controller was the weapon of choice for most companies knocking up SSDs. Things went a little quiet immediately following OCZ's acquisition, but the recent launch of the OCZ Octane and Petrol drives using the new Everest controller signalled the return of Indilinx to front-line SSD battle.
The other controllers
For the most part, the Everest controller delivers similar results to the Marvell 9174. Peak read speeds are good across the board, including incompressible data. Read speeds, again, take a hit with smaller capacities. In fact, you'll need a 512GB Everest drive to achieve the full 400MB/s.
Our favoured 128GB capacity translates into 170MB/s writes. The Everest-based drives all deliver outstanding 4k random reads and competitive random writes. The latter fall off with the smaller capacity models.
Our final controller candidate is Samsung's latest, as seen in the SSD 830. Not a huge amount is known about the controller, but it's said to be based on the ARM CPU architecture and pack three cores, and it's a very solid all rounder. Like Indilinx and Marvell-based drives, sequential read performance is strong at approximately 500MB/s, but you'll need a drive with at least 256GB to get the best write performance, both sequential and random - and that means big money.
On paper, then, SandForce-powered drives offer the best solution for our £100-plus, 120GB to 128GB sweet spot, chiefly because you get nearly all the performance the controller chipset has to offer.
Find out more about PC Format magazine
In practice, things are a little more complicated. For starters, that's because there are a whole hill of SandForce drives out there from a wide range of drive makers. But it's also because the controller chipset and associated firmware aren't the only factors.
A syncing feeling
Understanding synchronous and asynchronous memory types is important, too. It's often the case that a single drive maker will offer models with both types of memory, but using the same controller. Corsair's Force 3 and Force GT drives, for example. In both cases it's 120GB's worth of SandForce SS-2281 SSD.
The first is around £120, the latter more like £150. The difference is asynchronous and synchronous flash memory, respectively. Which begs the question, is it worth paying 25 per cent extra for the latter?
According to Corsair's own performance claims, there's very, very little in it. On the other hand, if you examine the raw specifications of the two memory types, there's a massive gulf.
Asynchronous flash memory complies with the ONFi 1.0 spec and is good for 50MB/s, whereas synchronous memory is ONFi 2.0 and clocks in at 133MB/s. Confusing, eh? The thing is, you only have to work out the implications of the two specs to realise that the real-world gap can't possibly be that big.
After all, with a 133MB/s chip at the end of each memory channel, your typical eight-channel SSD is looking at 1GB/s of bandwidth, and we've yet to see anything close to that, not least because it's well beyond what SATA 6Gbps can handle.
Likewise, 50MB/s on eight channels makes for 400MB/s. Bung in some data compression and you're going to be bumping up against the SATA interface again. The reality, then, is that the difference isn't always dramatic.
The final point to ponder is reliability. For us, this is the toughest question of all. It often takes many months of heavy SSD use before serious problems emerge. Overall, what we can say is that reliability and stability have improved greatly over the last five years.
Is it perfect today? Definitely not. We wouldn't recommend, for instance, an SSD for heavy torrenting. But there have been a big improvements all round. For most people, most of the time, we reckon an SSD from any of the big brands today is likely to deliver good performance for at least three years.
That said, there are brands that have earned strong reputations for reliable drives. The stand-out three are Intel, Samsung and Crucial. You pays your money. You takes your choice.