Page 2:Technical Specifications
Page 3:A Closer Look
Page 4:Sequential Read
Page 5:Sequential Write
Page 6:Random Read
Page 7:Random Write
Page 8:Mixed Workloads
Page 9:128KB Sequential Mixed-Workload Steady State Performance
Page 10:4KB Random Write Steady State Performance
Page 11:PCMark 8 Real-World Software Performance
Page 12:Total Storage Bandwidth
Page 13:PCMark 8 Advanced Workload Performance
Page 14:Latency Tests
Page 15:Notebook Battery Life
JMicron sent over three SSD configurations that use the new JMF670H controller and are paired up with brand new flash memory.
The road leading to Computex is paved with engineering samples. Over the next few days, you'll see us testing new controllers from JMicron and Phison, both with fresh flash configurations that we haven't seen used on retail products yet.
First up is JMicron, which sent over three parts in configurations we're expecting to see from partners at Computex. The JMF670H controller was first demoed a year ago in Taipei, but is now coming out of its development stage. This four-channel processor will compete in the low-cost mainstream category against Silicon Motion's SM2246EN and upcoming SM2256 controllers (the latter previewed on Tom's Hardware just a few months ago). Phison also has alternative in this segment, such as the S8 and S9 controllers that are already shipping.
We've talked about the race to the bottom. The mainstream space is where the large volume is. NAND flash makes up the bulk of an SSD's bill of materials, so in order to hit lower prices, the flash has to cost less. NAND comes from wafers, the price of which is fixed. If you want cheaper flash, you need to harvest more dies per wafer. This creates a big (or in this case small) catch-22. The smaller you shrink the manufacturing process, the less insulation there is between flash cells, causing problems with endurance and data retention.
Overcoming those issues requires more powerful controllers able to cope with beefier error correction. So now we have this issue where the cost of flash is going down, necessitating sophisticated (and increasingly expensive) processors. This adds cost, but ultimately shifts the burden of maintaining reasonable pricing to the controller vendor, and not the giant flash fabs.
SSD controllers have always played a large part in our coverage of solid-state storage. But their importance will increase not only as it pertains to performance, but also endurance. In the years leading up to now, keeping flash healthy was relatively easy. We're on the final steps of 2D planar flash though, with program/erase cycles dipping down to 500. NAND flash manufacturers are combating this two ways. The first is with 3D cell structures that pack more bits into the same amount of space. Samsung is already shipping 3D V-NAND technology in its 850 Pro and 850 EVO. IMFT (Intel Micron Flash Technology) announced a 3D strategy but retail products will not come to market for some time. Flash Forward, the partnership between Toshiba and SanDisk, recently announced BiCS, its 3D NAND product.
With 3D cell structures still a way out for most flash makers, steps to extend the life of 2D planar NAND were taken. Pseudo SLC (pSLC) has played a large role in mainstream SSDs for some time now. Samsung introduced TurboWrite with the 840 EVO in 2013. SanDisk released similar technology to TurboWrite, calling it nCache.
Just a few weeks back, IMFT announced FortisFlash, a product name for what was once simply referred to as MLC+. FortisFlash is a component of L95B, the same 16nm flash found in Crucial's MX200 and Micron's M600. Micron used Dynamic Write Acceleration to describe the M600's pSLC functionality.
We normally associate pSLC write modes with performance increases, but the technology also helps extend endurance as well. In the latter stages of 2D planar NAND, this is imperative. The short of it is that in order to take advantage of modern flash technologies, companies need new controllers with more horsepower.
- Compliant with Serial ATA International Organization: Serial ATA Revision 3.1.
- Supports one-port 1.5/3.0/6.0Gb/s SATA I/II/III interface
- Supports ATA-8 command set
- 32-bit embedded processor - ARM9 base instruction set
- 32KB embedded masked program ROM
- 192KB embedded system RAM with ITCM
- Support maximum 8CE’s Flash per channel
- Support Toshiba/SanDisk 32/24/19/15nm flash
- Support Intel/Micron 25/20/16nm flash
- Support legacy/Toggle 1.0/Toggle 2.0 mode flash
- Enhanced endurance by dynamic/static wear-leveling
- Supports 4K/8K/16K bytes page size
- Supports dynamic power management
- SMART (Self-Monitoring, Analysis and Reporting Technology)
- Data integrity under power-cycling
- Supports BCH 60/72-bit ECC
- Support Shift read feature of NAND flash when ECC fail
- Supports one module DDR3
- Support up to 4Gb
- Integrated SATA III port and four-channel flash controller
- LED indicator for SATA read/write access (optional)
- LED indicator for SATA PHY link up (optional)
- Provides 22 GPIO pins for customer
- Provides UART and JTAG for software debugging
- Built-in power-up self-test (BIST)
- Manual and automatic self-diagnostics
- Provides voltage low detect interrupt
- 288-ball TFBGA package
- Supports online SATA firmware update
- Support 1/2/4/8 banks selected free
- Firmware supports 1/2/4 channels selected free
The JMF670H controller uses powerful BCH ECC for MLC flash, so we suspect that this part will have a short life-cycle. JMicron plans to introduce a new processor specifically tailored for TLC NAND in the future. Hopefully we learn more about the JMF810 and JMF811 controllers at Computex, ahead of products shipping in 2016.
JMF670H improves upon the JMF667H controller currently in use by several SSD manufacturers. The JMF670H supports both Toshiba/SanDisk 15nm MLC, SK hynix 16nm MLC and Intel/Micron (IMFT) 16nm flash, as well as legacy flash. We'll take a look at two types of flash today.
IMFT's 16nm flash has been shipping for around a year now. In fact, Crucial released the BX100 and MX200 at CES last January, both with Micron 16nm NAND. The MX200 family has a few capacity sizes with pSLC, though not all models utilize the new feature. Crucial tells us that the larger drives don't require pSLC.
Micron says this about FortisFlash: "Get the performance of high-speed MLC while boosting endurance with our 20nm and 16nm FortisFlash devices, which come in standard BGA packages of one to 16 die stacks. With FortisFlash NAND, you can count on higher endurance than standard MLC without the usage limitations of Enterprise MLC (eMLC). Pair FortisFlash with advanced ECC methods, and you can exceed the standard 3000 write/erase cycles and see endurance levels over 10,000 P/E cycles in well-engineered systems. You can also improve data transfer rates with FortisFlash devices, which support ONFI’s high-speed synchronous interface.”
Flash Forward announced 15nm MLC quite some time ago. I've already tested two products with this NAND inside, one well over a year ago and one more recently. We suspect Flash Forward had issues with yields, and that's why retail products have taken so long getting to market. The first 15nm drive, SanDisk's X400, was announced a few days ago. It's designed for the embedded and low-cost HDD replacement segments.
The Toshiba/SanDisk 15nm node packs more dies per wafer than any other NAND on the market. As long as yields are good, the flash could cost less than even Micron's 16nm stuff, possibly leading to another price drop across new SSDs. We hope that advancements in flash will pave the way for $60 256GB SSDs by Black Friday.
JMicron provides firmware to its partners, as well as reference design material for component layout, power schematics and so on. Companies can quickly move through internal R&D stages and progress to production of retail products very quickly. Given the know-how to program pSLC modes, SSD vendors without the engineering know-how can get JMicron's help.
Like the other controller makers, JMIcron came up with a catchy name its pSLC mode: Write Booster.
A Closer Look
We're testing three different configurations today, all of which employ JMicron's JMF670H controller and Nanya DDR3 DRAM. As mentioned, the JMF670H is a four-channel controller that supports a single DRAM package of up to 4Gb density.
The M.2 model with 512GB of Toshiba 15nm flash is by far the most exciting of the three. Not only do we get to experience the JMF670H controller, but we also get to test it with rare flash. After Computex, we expect several existing products to move over to Toshiba 15nm NAND without a public disclosure (not something we condone), accompanied by a handful of of new models (the right way to change up an existing drive).
This is a double-sided M.2 2280 configuration that you should see in Ultrabooks toward the end of this year.
Our second M.2 drive uses 128GB of L95B MLC+ flash. Adata purchases finished wafers and then packages the flash in-house to save some money. This 128GB model employs just two NAND flash packages, and fits the surface-mount components on a single-sided 2280 M.2 stick.
The third drive utilizes the JMicron JMF670H paired with Nanya DDR3 DRAM and Adata-packaged 16nm MLC+ flash. This model comes in a 2.5" form factor and uses four NAND packages.
Comparing Products Used in Testing
Our test procedure starts out with a light conditioning phase, so our numbers do not always correlate with what the vendors claim in their marketing material. Specifications printed on retail packaging and webpages are sometimes called hero numbers, since the performance is measured at the best possible time, when the drive is fresh out of its box.
The three JMicron JMF670H-based products group together in the sequential read test. At first, we were surprised to see the 256GB model outperforming the 480GB version with Toshiba 15nm flash. It's possible that the lower-power M.2 interface plays some role in these results, though. This is the first time we've tested production Toshiba 15nm.
Still, as it sits in its current form, the JMF670H with L95B NAND is a bit slower at sequential reads than Phison's S8 with the same flash (as benchmarked on Patriot's Torch 240GB).
The JMF670H-based products pick up speed and become more competitive in our sequential write metric. This is an area where low-cost SSDs often struggle. Although these early JMicron samples are not capable of matching Mushkin's Reactor 512GB (Silicon Motion SM2246EN), their performance isn't too far behind. The SMI SM2246EN is fairly mature by now, and has received a handful of firmware revisions to improve its numbers.
Here we see JMicron's Write Booster technology in action while writing 128KB blocks across the entire user LBA span. Since most copy and paste operations with sequential data are usually fairly small, most users don't need a lot of bandwidth for a long period of time. If you paste a picture, you'll write it at or close to 400 MB/s. A larger file, like a 50GB Blu-ray ISO, slows down as more flash is used to hold the information.
Random performance is also a sour spot for low-cost SSDs. This is where controller clock rate and efficiency come into play. We like to use the 10,000 IOPS mark at a queue depth of one as a measuring stick to divide mainstream and enthusiast SSDs.
The JMF670H-controlled products are all within reach of that magic number, although none surpass it. Keeping in mind the controller uses only four channels, we're certainly impressed with the random read performance of all three capacities.
Random 4KB writes have always been a problem for JMicron's controller architecture. The company powered some of the first affordable consumer SSDs, and over the years it has worked hard to improve its performance in this discipline. Although we measure 4KB random writes in IOPS, this particular measurement is really about latency.
A single log file update to one program can bring your system to a crawl until the task completes. Native command queuing allows for up to 32 outstanding commands, though only one can run at a time. It it hangs or takes longer to process than expected, your experience is diminished. High IOPS performance translates to low latency.
The JMF670H is now on par with the random write performance available from several mainstream and even some premium SSDs on the market today.
Mixed workload testing is standard practice in the enterprise space, but it's just now becoming more popular for desktop storage benchmarking. In order to make this testing viable for client workloads, the number of data reads increases to 80%, leaving 20% for writes. AHCI, the underlying protocol for SATA, is a half-duplex interface, so devices can only read or write at one time, never both. This increases latency as tasks end up waiting on each other.
The JMF670H performs well in both sequential and random mixed workloads. The same rules still apply: low queue depths weigh heavier than high queue depths for desktop use. The baseline starts at QD2 though, and ramps to QD4 for multitasking.
128KB Sequential Mixed-Workload Steady State Performance
Most users shopping for low-cost SSDs don't run professional applications that write large amounts of data in short periods of time. Doing so can create a steady state condition where performance drops to very low speeds. Using 80% reads for desktops and 70% reads for workstations, we can see steady state performance in the last two charts.
Under both workloads, the JMF670H-based samples perform proportionally with their on-board flash. All of the dies share similar density. However, the amount of interleaving decreases with each step in capacity. As a result, steady state performance decreases as the SSD gets smaller.
4KB Random Write Steady State Performance
Here we see the worst possible random write condition. Again, the JMF670H's performance scales as capacity increases due to interleaving.
If you're striping SSDs for a performance increase, you want to see consistency in this metric. Variability multiplies with each product added to the array. If the swings are large enough, you might as well roll a dice to see if you're getting a fast write or a slow one for each request. The three JMF670H-based drives we tested do a good job of minimizing fluctuations, while keeping the lowest write condition high.
PCMark 8 Real-World Software Performance
For details on our real-world software performance testing, please click here.
The three JMicron JFM670H-based SSDs compare well against existing low-cost rivals when we sort by capacity. In many of these tests, the difference is only a few tenths of a second, so we're not going to crown an undisputed winner. Still, the leads are clean and undeniable.
Total Storage Bandwidth
After tallying up the results by throughput, we get a clearer result that is easier to interpret. Surprisingly, even the 120GB JMF670H is faster than Patriot's Torch 240GB. That's the outcome that JMicron shot for with its new Write Booster technology. Value-oriented low-capacity SSDs no longer have to deliver HDD-like performance.
PCMark 8 Advanced Workload Performance
To learn how we test advanced workload performance, please click here.
We only look at the light use workload portion of the tests for products in this price range. Of the three JMF670H-based drives, the 240GB model performed the best. We went into this series of measurements expecting the 480GB implementation with Toshiba 15nm to score a first-place finish, but that wasn't the case. The JMF670H L95B 240GB model recovered faster and delivered the highest throughput.
Keeping with the light workload group, this time measuring service times, the JMicron JMF670H controller paired with 256GB of L95B FortisFlash delivers good performance. This is a measure of how long the entire test takes to complete, creating one of the most important charts in our article.
The two other JMF670H-controlled products also performed well, though their service times were quite a bit higher than the 256GB model.
Notebook Battery Life
For more information on how we test notebook battery life, click here.
The two M.2 products are not comparable with the 2.5" form factor drives. We test M.2 SSDs in a Lenovo X1 Carbon Gen 3 and 2.5" products in a Lenovo T440p. You can compare battery life with other M.2 SSDs toward the end of this page.
The 2.5" JMF670H-powered SSD scored low on our chart, but is average for client-oriented 2.5" SSDs currently available. This is one of the last puzzle pieces that controller vendors optimize for, so it's also possible that we'll see an improvement once the processor finds its way onto retail products.
Notebooks running on battery power reduce system bus speeds to conserve power. The CPU, PCIe and DMI all drop down to lower power states and this affects storage performance. Some devices handle the reduced power state differently. Those that do well under these conditions are considered efficient; others simply perform poorly.
The SSD industry moves quickly. Partners from all sides come together to streamline progress. Solid-state drives are no longer performance parts found only in enthusiast PCs. Most of us own products reliant on NAND without even realizing it. The number of products exposed to flash is growing, but only because new controller technology allows low-cost NAND to be viable.
The next big push will try to displace hard drives as boot devices in nearly all PCs. It's ambitious, to say the least, but projections show it's possible to sell SSDs for the same price as existing 2.5" hard drives. Even though the lowest-cost SSDs cannot keep up with high-performance models, they'r still significantly faster than their mechanical predecessors. Dramatically more random performance translates directly into a better computing experience.
The untold aspect is available capacity. You can already purchase 128GB SSDs for the same price as 2.5" HDDs. For this example, we're using a 120GB PNY CS1111 SSD and comparing it to a large number of 2.5" 5400 RPM HDDs with 250GB of space. Current pricing comes out to $49 for the SSD and $35 for the hard drive. Once 256GB SSDs start selling for somewhere around $50 to $60, you'll see them becoming even more prolific. Believe it or not, that's going to happen this year.
Controllers like JMicron's JMF670H give me reason to believe that the SSD industry will be able to hit those low price points without moving to three-bit-per-cell (TLC) flash. This is good news. MLC at low capacity points is fairly slow compared to the enthusiast-oriented SSDs we benchmark most often. But TLC in low capacities from most flash vendors is like using a thumb drive for your operating system. Sadly, in some cases, that's still faster than a hard disk, though.