Crossbar takes on DRAM and flash storage with super fast, super long-lasting RRAM tech

Startup Crossbar emerged from stealth mode Monday to announce its version of RRAM (resistive random-access memory), a new type of memory that could be a successor to flash storage and DRAM.

The company, founded in 2010, will make and license its RRAM, a nonvolatile memory, which will be smaller, faster and more power-efficient than NAND flash and RAM, said George Minassian, CEO of Crossbar.

"It is higher density ... and the current is much lower," Minassian said, adding that the memory's physical and power attributes make it a suitable replacement for storage in smartphones, tablets, PCs and servers.

Crossbar is claiming RRAM will deliver 20 times faster write performance, 20 times less power consumption and 10 times more durability than NAND flash. The memory chips will be stacked, and a 1TB module will be roughly half the size of a NAND flash module with similar storage, Minassian said.

He could not estimate the price of a 1TB RRAM module, but said it will cheaper than NAND flash partly because RRAM is less expensive to manufacture.

Crossbar
Crossbar will also license the technology to third parties. It could be two or three years before the memory shows up in products, but that depends on demand for the product, Minassian said.

"It's a matter of what company appears at what time," Minassian said.

RRAM differs from the way NAND flash and RAM operate. Unlike NAND flash, Crossbar's technology does not use transistors or trap a charge. Instead it uses a layered approach to store data. An RRAM cell has three layers, with a switch in the middle that helps determines whether the cell is storing a 1 or a 0. The top layer has a metallic electrode, while the lower layer has a nonmetallic electrode. The top layer passes metal ions into the switching media and into the lower layer, which creates a filament to keep the electrodes connected, what Minassian called a "short wire." Applying a negative charge breaks the wire and leaves a gap between the electrodes, which leaves no resistance, changing the status of the memory cell.

"This is not a gate you use in standard NAND and NOR. This is resistive, which is where RRAM comes from," Minassian said.

RRAM uses existing material and can be made in factories. Prototypes are being made in factories of TSMC (Taiwan Semiconductor Manufacturing Co.), Minassian said.

Crossbar's RRAM doesn't contain transistors, so it is easy to make as chips become smaller, Minassian said.

Crossbar
Crossbar's technology is very interesting and could be useful as manufacturing technologies improve and chips get smaller, said Jim Handy, principal analyst at Objective Analysis, in an email.

"It's generally accepted that something is going to replace NAND and DRAM someday, since these technologies have scaling issues. When that will occur is unclear," Handy said.

It is becoming more complex to make NAND and DRAM as manufacturing technology improves at a rapid rate, Handy said. Intel, which has the world's most advanced chip factories, will soon move to the 14-nanometer process to make its chips. Manufacturing smaller chips with more features requires more attention to detail, and chips could be vulnerable to a wider class of defects.

"With that in mind, and with the advent of 3D NAND as a successor to today's planar NAND, there may be another five process generations, or even more, before alternative memory technologies have a shot at replacing NAND flash. The story's probably the same for DRAM," Handy said.

Once a scaling limit has indeed been reached, technologies like Crossbar's RRAM will rapidly take away market share from the entrenched players, Handy said.

Other RRAM designs are being researched by academics and research institutions. Alternative forms of memory to NAND and DRAM also include Everspin's MRAM (magnetoresistive RAM) and PCM (phase-change memory), a memory type being pursued by Micron and Samsung Semiconductor. Hewlett-Packard is making a memory type called memristor.

Startup PernixData virtualizes server-based flash

PernixData, a startup founded by VMware veterans, says its software can turn server-based flash storage into a resource shared across a standard VMware cluster of as many as 32 systems.

Making that cache available across a whole cluster will let enterprises scale out their storage capacity with consistently fast access to data, according to co-founder and CEO Poojan Kumar, who previously managed VMware's data products. PernixData's product, called PernixData FVP (Flash Virtualization Platform), works without modifications to the hypervisor or applications and is designed to coexist with SANs (storage area networks), he said.

Putting flash storage directly into servers is a growing solution to the need for faster access to data, but the technology's been limited to individual servers, so virtualized environments can't fully take advantage of it, according to PernixData. Clusters of virtualized servers typically share data over a SAN, which can raise performance problems, Kumar said. Multiple network hops and contention over a shared SAN link can all slow down access to the data that applications need.

PernixData's software can turn flash cards on all the servers in a virtualized cluster into a shared cache, which serves as a tier of storage above a SAN for the most-used data. It's a software-only product and can use on-server flash that customers already have deployed. Announced on Tuesday, PernixData FVP is available immediately for US$7,500 per host for an enterprise package. It's initially for VMware environments, though ultimately it will work with multiple hypervisors, including KVM and Microsoft's Hyper-V, Kumar said.

The software is designed to allow fast access to data while virtualizing server resources. The virtual machines on a given server mostly will get their data from the cache right on that system. But even when they need to tap into another server on the cluster, that will be faster than going out to a SAN, the company said: Delays can be measured in microseconds rather than milliseconds. And PernixData works with all features of VMware, such as movements of VMs using VMotion, making management easier, Kumar said.

Another advantage to PernixData's approach is that administrators can control which VMs can use which flash resources, giving preference to the jobs that most need the performance. It does this by taking advantage of both the VM and the storage being on the cluster, so that kind of control isn't possible with flash that's on a SAN, Kumar said.

Christopher Greater Area Rural Health Planning Corp. (CRHPC), which operates 13 clinics in southern Illinois, sees PernixData's software as key to implementing VDI (virtual desktop infrastructure) for its doctors and other employees.

VDI should cut maintenance and energy costs while helping ensure security and giving workers more flexibility, said Jason Rolla, chief technology officer. For example, doctors frequently move between their offices and examination rooms, he said. "Technically, they could carry their office everywhere they went," Rolla said.

CRHPC has an almost four-year-old Dell EqualLogic storage array in its headquarters data center that's connected to its servers via iSCSI. But when it tried to set up VDI using just the array for storage, application response times were much longer than the 3 milliseconds that CRHPC was aiming for, Rolla said.

Tests with 10 virtual desktops showed that using the cache with PernixData will cut down on response time by about three-quarters, and CRHPC now plans to implement the software in its full VDI rollout, Rolla said.

PernixData should be useful for databases and big data analytics, as well as VDI, said Gartner analyst Chris Wolf. Virtualized storage products from some other vendors, including VMware, have been hard to sell because they were too expensive for what they did, he said.

"As long as the product holds up to what it appears to do on paper, Pernix is going to put a lot of price pressure on some of the more industry heavyweights here," Wolf said.

Hitachi brings high-end flash to midrange HUS VM storage line

Enterprises are gaining the ability to turn existing storage platforms over to flash even as solid-state media remains mostly a tool for caching and for applications with special requirements.

On Wednesday, Hitachi Data Systems is bringing flash technology from its high-end storage platform to its midrange Hitachi Unified Storage Virtualization Machine (HUS VM) system and introducing software for the system that will allow it to be used as a high-performance, all-flash array. Other vendors are also entering the all-flash arena, including Hewlett-Packard, which announced the 3Par StoreServ 7450 Storage system less than a month ago.

Hitachi is taking the Hitachi Accelerated Flash Modulefrom its Virtual Storage Platform and offering it for the HUS VM, a midrange array designed for smaller enterprises with less need for scale. The new, flash-optimized software included in Wednesday’s announcement will roughly double the potential throughput of the system, and another software release later this year will double it again to 1 million IOPS (I/O operations per second), the company said.

The HUS VM, introduced last year, fits in between the VSP platform for large enterprises and the HUS 100 family, Hitachi’s entry-level line. Customers can now outfit an HUS VM entirely with SSDs (solid-state drives) or mix those components with spinning hard drives.

The flash module packs 1.6TB of capacity onto a card that goes deeper into a standard 19-inch rack unit than a typical SSD would, making more efficient use of space, according to Roberto Basilio, vice president of infrastructure platforms product management at Hitachi. The module also uses multilevel cell flash, which has higher density than the single-level cell type used in many enterprise SSDs.

But speed is at the heart of solid-state adoption and of Hitachi’s enhancements to the HUS VM. Each Hitachi Accelerated Flash module has its own multicore, multithreaded controller chip that can carry out functions such as compression and wear-leveling without affecting read and write performance, Basilio said.

To go along with the new flash module performance that this design brings, Hitachi has introduced software improvements to the HUS VM that can double its overall throughput. The HUS jumps from 250,000 IOPS (I/O operations per second) to 500,000 IOPS. Another optional software upgrade, coming in the fourth quarter, will double that yet again, to 1 million IOPS, according to Hitachi.

The rapid read and write performance of flash can help speed up enterprise applications where needed, though it will coexist with hard disk drives for the forseeable future, said Enterprise Strategy Group analyst Terri McClure.

“Judicious use of some level of solid-state storage at every layer of the stack is going to become the norm, over time,” McClure said.

The key is good data-management software to control how the two types of media work together, she said. This software has improved since the introduction of flash storage, and Hitachi’s capabilities there are strong, she said.

Enterprises will be able to fit 12 flash modules on each tray in the HUS VM, with a full eight-tray system holding more than 150TB of capacity. That will double by the end of this year, as the company plans to come out with a 3.2TB module.

The flash modules are available now for $28,427 each, and an all-flash HUS VM with 8TB of flash capacity costs about $300,000, including the Base Operating System and three years of maintenance.