The SSD Flash — Summer 2010

Welcome to the Summer 2010 issue of the SSD Flash. Due to scheduling of releases of new products that we wanted to cover, the issue has been delayed somewhat but includes some recent market entries that are very interesting.

A discussion of the SSD landscape
With more than half of 2010 now history, it is a good time to calibrate on some of the developments in each of corner of the envelope of SSD products. To keep things organized, we define the corners of the envelope as Foundry Owners, Drive Format SSD, PCIe SSD and Appliances.

Foundry Owners
The SSD market cannot be taken out of context of the importance of the supply side at the raw Flash component level. This ‘first corner’ must be healthy to continue the investment in the foundry facilities to keep the flow of higher densities coming online. Intel-Micron is now shipping its 25nm MLC (two bits per cell) in volume and has begun sampling its ‘3LC’ components (three bits per cell). Micron is shipping 34nm ‘E-MLC’ that is, somewhat premium priced three-bit per cell MLC spec’d at 30,000 program/erase cycles. In addition, Micron will shortly be in volume production of an enterprise-class drive that uses 34nm single level cell components. Samsung and Seagate recently announced a cross licensing and partnering agreement. Toshiba has invested heavily in appliance manufacturer Violin Memory. Overall, business is good in the Flash business. Apple, alone, is expected to consume upwards of 20% of worldwide flash foundry production this year for its iPAD and iPhone products.

Drive-format SSD
Industry leader STEC has been joined by a number of new entries in the enterprise drive-format corner of the envelope. Pliant, OCZ, Unigen and Anobit, to name a few, are particularly focused on the enterprise market. The SandForce SF1500 enterprise controller has enabled a number of integrators to enter the market, while STEC, Pliant and Anobit point out that they have developed their own controller technology and have differentiation from that standpoint. The drive format products have a great advantage of being able to plug and play into an existing array. This does come at a cost over other implementations. Drive format products need to pretend that they are really a disk drive. This means that an industry standard interface, be it Fibre Channel, Serial ATA or Serial Attached SCSI needs to be built into each drive. Besides being able to plug-and-play into a number of existing products (drive bays, JBODs, storage area networks), the drive format products provide reliability benefits in terms of no single point of failure and compatibility with in-use protection and self-healing strategies.

PCIe format SSD
While taking advantage of the full bandwidth of the CPU bus, PCIe format SSD products have some interesting trade-offs. Once the exclusive domain of Fusion-io, products in this class have been announced by Network Appliance, Texas Memory Systems, OCZ, LSI and new entrant Virident with its TachIOn product. These devices provide extreme performance at premium price points. They certainly keep to the rule of taking the highest speed devices closest to the server CPU. In many situations this direct-attach deployment makes perfect sense. However, as in all things storage, there are tradeoffs. A big one, especially when you are dedicating a single piece of expensive hardware (these PCIe devices can run to tens of thousands of dollars) to another expensive device (a high-end server), lots of concerns about the expense of redundancy and the potential for single points of failure taking very costly components offline are associated with these devices.

Appliances
SSD appliances are growing in numbers and capabilities. Avere Systems was early to market with its FXT2700 NAS system designed to front-end other filers and acting as a supercharger in the filer environment. Another early entrant, Whiptail Technologies, has focused on block storage and specifically the IO requirements of Virtual Desktop images with its 2U package capable of supporting up to 6TB. Violin Memory adds large scale aggregation and modular, hot-swappable SIMMs, that is, Flash boards that can be populated into its 3U chassis. A single 3U can support up to 10TB of high speed Flash and a single rack scales up to 140TB. NextIO adds the dimension of IO virtualization and a practical use of the PCIe SSDs with a shared server pool. This provides an answer to the single point of failure issue and the ability to share expensive resources across a pool of servers and PCIe SSDs.

Assessment
It is clear that the market can rationalize all four corners of the envelope. As foundry owners produce lower cost and more reliable components, the other product categories can move on with higher performance and lower cost per byte. In our opinion, no single corner will dominate. It is possible for a foundry owner to use its vertical integration to dominate the other three corners. While this may be possible at the disk format SSD corner, it is unlikely that PCIe and Appliances will be coming out of the foundry owners any time soon. But stay tuned, this can all change very quickly.

Virident joins the PCIe SSD fray
Virident announced a new entry in the PCIe SSD market with its tachIOn enterprise-class product. The company, with facilities in Milpitas, Calif., and Bangalore, India has been developing Flash-based devices for about three years, but tachIOn is its first broadly launched product. Virident joins Fusion-io and Texas Memory Systems who have PCIe products that use device drivers that run on the host system to perform the management functions associated with wear-leveling and other housekeeping functions. Other entries, including LSI and OCZ, use on-board controllers and interface to the host a Serial Attached SCSI (SAS) or Serial ATA (SATA) interfaces which do not require proprietary drivers.

The tachIOn is built from 50GB SLC Flash modules and comes in either 200GB or 400GB capacities. Unique to tachIOn, the Flash modules can be added in the field and replaced for repair or upgrade purposes. The company has qualified flash modules from Samsung, Toshiba and Micron. While currently available only in SLC, the architecture will support MLC and even has the capability for hybrid SLC and MLC modules.

Performance is very good, as is expected from PCIe and SLC, with 320,000 Input-Outputs Per Second (IOPs) for 4KB random reads, and sustained random read/write (75%/25%) of 200,000 IOPs. The wear-leveling algorithm supports a 24-year life assuming 5TB in writes per day, the company claims. At product release, drivers are available for Linux, with Windows drivers coming soon. Initially aimed at an OEM market, tachIOn is targeted at a $6,200 price for the 200MB version.

Netapp has PCIe accelerator, but at a price
In June, Netapp announced the relaunch of its PCIe cache product. With a new name, the product is actually the same PCIe 8 channel, 512GB device announced in late 2009. But with an MSRP of $80K for 512GB of SSD it carries a premium price that is hard for customers to swallow. $156 per GB is quite a premium, considering open products, like the recently announced Viridity product line, go out at $35 per GB, and consumer-class products like the OCZ go for $3.50 per GB, albeit with low-cost MLC Flash memories.

Storage system suppliers with proprietary control nodes have a pricing advantage over open systems products. Whether these markups can be supported long into the future, as more and more products like those from start-ups like Nimble Systems to value-priced systems from Dell, create increasing price pressure.

NVELO Launches DataPlex Product
Denali Systems, over the years, has been the unsung hero of many embedded controller firmware implementations across a broad range of manufacturers and products. After the recent acquisition of Denali by Cadence Design Systems, a number of Denali principals created the spin-off, NVELO. NVELO’s mission is to bring many of the advanced designs implemented in RAID, solid state drive (SSD), and hard drive (HDD) technology to the client market.

NVELO’s inaugural product is called DataPlex, a set of drivers which bring best practices for hybrid drive control to the PC market. The application of adaptive caching, first introduced by Adaptec’s MaxIQ is beginning to show up in a number of products that utilize the data stream to a RAID controller to identify hot data and maintain it as cache on an SSD. DataPlex takes that one step further, and provides a level of flexibility for laptops and desktops that has not been available. As an example, a mini-PCIe SSD card of 16GB can be added to a laptop and DataPlex will use it as cache. In a desktop example, a relatively small SSD, say 50GB, can be added in an open bay and DataPlex will cache there. A user setting selects whether hot data is written to both the HDD and the cache (write-through) or is only written to HDD if it migrates out of cache (write-back).

As we have discussed, a relatively small cache is all that is needed. In benchmarking examples, DataPlex delivers 95% of the performance of an SSD at a fraction of the cost per GB. DataPlex will be available from Tier 1 PC OEM’s in early 2011. The company is currently seeking partnerships in order to bring the capability to the aftermarket.

Smart Modular first to market with SAS coupled with E-MLC
Micron briefed us on their ‘Enterprise Multi-Level Cell’ Flash products earlier this year, and Smart Modular is the first to market with a 3Gbit/s Serial Attached SCSI (SAS) SSD with the E-MLC Flash modules. The E-MLC is based on 34nm lithography and has two-bits per cell. Micron is a little cagey about whether the modules are produced with a modified foundry process and/or are cherry-picked off the wafer. We suspect that both methods are used to get the chips that will support 30,000program/erase cycles, as opposed to run-of-the-mill MLC that usually tops out at 5,000 cycles.

The Smart Modular XceedIOPS SAS SSD achieves up to 26,000/20,000 Input-Outputs per second (IOPS) random read/write and 250/230 MB/s sustained read/write and is available in 50, 100, 200, and 400 GB capacities. As usual, these sizes are net of the 64, 128, 256 and 512GB actual capacities, the balance being consumed in various controller functions. In a 2.5” format, the new drive has enterprise features that include variable sector sizes, which allows host transfer sizes of 512, 520, and 528 bytes. Other features include temperature monitoring and throttling, and staggered power-on to limit surge in large arrays. The new drives are now being sampled to selected clients.

So what’s with another new acronym?
Now we have a major SSD supplier (Smart Modular) shipping the new Micron Enterprise-Multi Level Cell (E-MLC) components, we should take some time to discuss the growing jargon around these devices. And to add another wrinkle to the acronyms, MLC comes in different versions. When referring to MLC, there is another question that needs to be asked: ‘How many bits in a cell?’ MLC is usually two bits per cell, but three bits and even four bits per cell chips are out there. Generally speaking, the more bits per cell, the slower the access, the fewer program/erase cycles and the lower the cost per bit. Marketers love to invent new acronyms, and we just got a new announcement from Intel-Micron which touts their new 25nm ‘3LC’ technology. If you guessed that they have baked three bits in a cell, you are correct. These acronyms are not universal to all suppliers, so our recommendation is to always ask how many bits per cell. Again, single level cell is best (read, expensive) and four bits per cell is not to be used for anything but throw away storage. For example, if you lose a few bits in a photograph or music file, you will never miss it. Lose a few bits in an ERP record and you have a big problem. E-MLC strikes a new balance, with a premium two-bit per cell MLC at a higher price (but not double) of run-of-the-mill MLC. We are aware of some yield issues with E-MLC, and it may be another interim step (read cherry-picking off of the overall wafer).

Why 3LC, TLC are technically incorrect
When discussing multi-level cell, it is important to understand what the term ‘level’ refers. It only indirectly refers to the number of bits. It refers to the number of voltage levels operating in the system. Many discussions of three bit per cell devices refer to them as 3LC or TLC. Just for fun, I started a dialogue on the Linked-in Solid State Disk Innovators Network that suggested the term ‘nLC’ be adopted as standard jargon, where n=the number of bits per cell, since Intel-Micron had used the term in its product announcement. The replies from this group were unanimous in rejecting the term due to its technical inaccuracy. Instead, the overwhelming favorite was the term nbC, where n=number of bits per cell, and b is in lower case, to indicate bits instead of bytes. We are going to support this format as in the following example: ‘3bC MLC and 4bC MLC offer reduced costs per GB at a slower access speed and reduced number of program-erase cycles.’

Soligen aims at OEM integration
As SSD manufacturers seek differentiation, the focus is turning from drive form-factor to circuit boards. An example of this trend is the ‘SIMM’ modules in the Violin Memory 3200 appliance, and the new ‘FlashBridge’ from Soligen Corporation.\

We spoke with co-founders of this startup, CEO Mark Ayers and Sales Chief Doug Suvalle in order to understand their positioning of the FlashBridge in relationship to their conventional ‘circuit-in-a-can’ 2.5 inch SATA-II SSDs. The Soligen Spartan client drive uses SandForce SF1200 controllers, and the Elite enterprise drive uses SF1500 controllers.

The FlashBridge uses similar configurations and is aimed at the OEM market. Flexibility in use of Flash modules (both MLC and SLC models are available) and SF controllers is a primary advantage. The 32mm by 64mm boards provide significant opportunity to aggregate large memory capacities as opposed to the limitations inflicted by stuffing everything into a 2.5” or 3.5” drive format. The onboard controllers still provide the isolation of single unit failure that drive formats provide, for example in a JBOD array. OEM’s can package the FlashBridge circuits in ways that enhance ergonomics, as in slate-type devices like the iPAD, in blade servers or new Flash memory aggregations. Each Flash Bridge contains either 128GB or 256GB and can be combined into a single 512GB module.

Micron introduces Realssd P300 enterprise SSD
Micron is now sampling its P300 enterprise SSD, a fast, SLC-based 6Gbit/sec. Serial ATA drive in 50GB, 100GB and 200GB capacities. As seems to be the trend, those capacities are net of 64GB, 128GB and 256GB, respectively. The difference amounts to the area of the disk reserved for housekeeping, worn out areas and error correction. A straight-forward enhancement of its MLC-based C300 client drive, the P300 incorporates 34nm SLC to produce enterprise-class endurance and performance. Claiming the top spot in SATA performance, the P300 sports high bandwidth with read throughput up to 360MB/s and a write throughput of up to 275MB/s. Random read IOPs (Input-Output per second) of up to 44,000 and random write IOPs of 16,000 have been recorded by benchmarking firm Calypso systems.

The P300 replaces the 25GB/50GB P200. The higher density is achieved the old fashioned way: Moore’s Law. By using 34nm lithography, the SLC memories now are capable of 50GB/100GB/200GB in the same 2.5” package. Mass production of the P300 is expected in October.

MOSAID develops HLNAND for new FLASH control paradigm
Ottawa-based MOSAID announced the development of a new access path for Flash memories back in 2007, and now has developed prototypes of three levels of product, including an SSD prototype using a new version of the Indilinx Barefoot controller. MOSAID (not to be confused with the Israeli intelligence service), has been involved in the development of Metal Oxide Semiconductor intellectual property for thirty plus years. The company’s mission involves developing technology and then licensing it to manufacturers. We had the opportunity to discuss the advantages of HLNAND with Hakjune Oh, MOSAID’s director of strategic technical development prior to his journey to the Flash Summit.

For its HLNAND technology, MOSAID coined the term ‘Hyper Linked’ NAND, hence the HL. The Hyper Link in this case is a daisy-chain ring topology for accessing the NAND through a point-to-point connection rather than the traditional bus multiplexing methods. The result is dramatically faster access, and the ability to aggregate large numbers of NAND chips on the ring. The existing prototypes can run at speeds up to 266 MB/s per channel. The next generation, HLNAND2, was unveiled at the FLASH Summit conference. HLNAND2 has a bridge controller under development that will support speeds of up to 800 MB/sec. per channel. Furthermore, the topology can support full-duplex operation, (simultaneous reads and writes) to a given chip, effectively doubling the available speed in some circumstances.

When we asked who the ideal partner to bring HLNAND to market, the MOSAID team made it clear that a major Flash maker would be ideal, but integrators were also being evaluated. The speed and aggregation capabilities of HLNAND are bound to move SSD throughput into a new realm. The team expects some cost reduction due to reduced number of pins required in a given die, but the primary cost reduction will be in cost per IOP, rather than cost per GB.

Viking Modular aims for OEM market
We had the chance to review Viking Modular’s SSD strategy with Adrian Proctor, vice president of marketing. Since Viking Modular’s focus is on the OEM market, a little background is in order. A division of major contract manufacturer Sanmina SCI, Viking has substantial engineering and manufacturing facilities in Orange County, CA. Viking had just announced its SATADIMM product, in 50GB, 100GB and 200GB. The device provides a new level of integration for SSD appliances and server implementations, providing a lower cost and better densities than SSD-in-a-can drive formats. Adrian pointed out that these can drop directly into open DDR3 slots on a Nehalem processor board, and can be powered from the DDR slot and can be simply wired to the existing SATA connector.

Viking customers, typically white box suppliers, can choose from the enterprise class SandForce SF1500 controller or SF1200 controller, and a wide variety of Flash products from all NAND fab owners.

Pliant announces enterprise MLC drives
Pliant Technologies has announced two new enterprise-class flash drives (EFDs) as a way to push the entry level costs for hybrid arrays to lower levels. In a 2.5” SAS package, the LB 200M and LB 400M have 200GB and 400GB respectively, and are priced at about half the cost per GB of Pliant’s LS 300S and LS150S SLC EFDs. Using two bits per cell MLC (2bC MLC), the new drives have sustained total IOPs of greater than 20,000. While this falls short of the 120,000 IOPs in the Pliant SLC EFD, it still represents roughly 100 times the IOPs of a typical high performance HDD.

Greg Goelz, Pliant’s marketing vice president gave us some interesting insights into challenges of managing the 2bC MLC. A little known ‘feature’ of these devices it that the bits in each cell are not accessed within the same page. This is to keep the physically adjacent bits isolated from the same program erase cycle. However, the adjacent bit during a given program erase cycle needs to be accessed at a different voltage during the cycle in order to preserve its state. While this is done automatically in the silicon, special consideration needs to be given to this function during various operations. Pliant refers to this as Page Pairing Data Protection. Other advanced features of the Pliant controller include a background ‘Patrol Read,’ triple redundancy on the ECC on metadata and background memory reclamation. The drives are scheduled to ship in October.