Move over solid state drives. NOR flash memory is poised to take on its NAND flash counterpart in the server arena. Spansion, a flash memory company, has debuted its NOR-based EcoRAM memory with the help of its two OEM partners: Virident Systems and Appro. According to Spansion, EcoRAM-equipped servers offer up to 512 GB of DRAM-like read performance without increasing power consumption or compromising reliability.
Founded in 1993 as a joint venture between Advanced Micro Devices and Fujitsu and later acquired solely by AMD, Spansion is focused on NOR and NOR/NAND hybrid products for the flash memory market. AMD jettisoned the company in 2005, and Spansion has been struggling to turn a profit ever since. In March the company filed for bankruptcy protection, which was preceded by some rather contentious downsizing that earned it the enmity of many of its former employees.
Amid this turmoil, Spansion is moving forward with its flash-based server business. EcoRAM is Spansion’s entry into that market and represents the first NOR flash solution for servers. NAND flash solutions from Intel, Texas Memory Systems, Fusion-io, and others have been making their way into this market over the past two years and are starting to be deployed to ease the performance bottleneck of hard disk drives. The bottleneck is being choked by the growing size of data sets and the increasing computational horsepower of multicore processors.
In the Internet application area, search engine index files are now multiple petabytes in size, and even single database instances used in social networking scale to a half a terabyte. Of course in HPC, a range of applications — seismic analysis, genomics and others — have terabyte-sized data sets. The latest multicore chips are better able to handle these big databases; the problem is that main memory capacity isn’t keeping up. On top of that, the gap between disk speed and memory speed is enormous. Disk storage latencies are on the order of a million times slower than DRAM. That’s why keeping data in main memory can have such a big impact on application performance.
So along came SSDs and PCI-connected flash memory to fill this performance gap. These products made use of the same NAND flash technology as found in cameras and thumb drives, but with better wear cycles. NAND is high density, low cost, but doesn’t have great reliability, and is not as fast as DRAM — microsecond latencies for NAND flash versus nanosecond latencies for DRAM.
On the other hand, NOR flash memory is more expensive than NAND (by a factor of two or three), not quite as high density (although more so than DRAM), but nearly as fast as DRAM for reading (although more like NAND for writing). Perhaps, more importantly, NOR flash does not have the wear degradation issues during read access that you have with NAND technology. (Both unfortunately have write degradation problems.) Typically NOR flash is used where reliability and read performance is paramount, like storing code to be executed and accessing safety-critical data. The market is not as big as NAND, but the value is there for certain applications.
Spansion’s breakthrough was to put NOR into a DIMM form factor. Even though it’s more expensive than NAND flash, it has the advantage of DRAM-like read speeds and DRAM-like reliability. Compared to DRAM DIMMs, Spansion’s EcoRAM DIMMs offer eight times the capacity, but use no more power than its low-capacity brethren. Spansion’s initial EcoRAM DIMM comes in a 32 GB size and draws the same 10 watts of power as a 4 GB DRAM DIMM. Also, on a per byte base, Spansion claims the reliability is actually better than DRAM. According to the company, a 32 GB EcoRAM DIMM won’t fail any more often than a 4 GB DRAM DIMM.
Spansion is positioning EcoRAM for applications that are read dominant, that is, where read operations are significantly more frequent than write operations. As it turns out, this is a fairly common way that large databases are used. Many types of visualization and data analytics applications fall under this model, including social networking analysis, Web site searching and indexing, seismic analysis and reservoir visualization, genomics, and data warehousing/mining. Spansion believes its EcoRAM offering can tap into a $5 billion market where these large read-dominant data sets predominate. That doesn’t mean you can’t use EcoRAM for writing data. It’s just that as more write operations start to creep in, performance will degrade accordingly. At some point, the economics become unfavorable, and in these cases, it makes sense to stick with external storage or SSDs.
Unfortunately, because reading and writing to these NOR devices are so different, EcoRAM is not just a drop-in replacement for DRAM DIMMs. Spansion has come up with an “accelerator” ASIC that manages the EcoRAM modules and makes the NOR flash look like regular memory to the application. For each set of eight EcoRAM DIMMs, you need to sacrifice a CPU socket for the accelerator (which only chews up a handful of watts). Currently, Spansion only supports EcoRAM on two-socket and four-socket Opteron boards using the Torrenza socket interface. For a maxed out four-socket server, you get two Opterons, two EcoRAM ASICs, 512 GB of EcoRAM, and whatever DRAM is plugged in on the CPU side. For a two-socket system, cut everything in half.
EcoRAM is already up and running on Appro servers running oil & gas codes. The Stanford Exploration Project, an academic consortium funded by the big energy companies, demonstrated a 60-fold speedup on a data transpose operation for seismic analysis using EcoRAM-accelerated Appro gear. Although the sample data set was only 80 GB — so within the reach of existing high-memory servers — many of these seismic data sets are actually much larger. The server was also used to do fly-through visualization of an oil reservoir with the entire data set in EcoRAM flash. The general idea here is to be able to do analytics and visualization work in real time, thus speeding up the workflow, while at the same time reducing operational costs.
Appro is the first HPC vendor to integrate EcoRAM (Virident is Spansion’s system partner for the Internet market) and is targeting oil & gas, life sciences, and visualization customers for this solution. The company is not expecting to sell entire clusters stuffed with EcoRAM, just some number of server nodes that contain the extra flash memory. Appro has qualified the solution for two- and four-socket Barcelona-based systems, but pricing information is not publicly available. Of course, interested customers can always contact the company for quotes.
According to Jan Silverman, Spansion’s VP of the server unit, all the big system vendors are currently evaluating the EcoRAM technology. “Every major server manufacturer has got this up and running on their box,” he told me, adding that in the HPC space, a variety of academic and government institutions are currently testing EcoRAM on Appro gear.
Silverman said the technology is also well suited to Hadoop clusters. Hadoop is an open source framework for distributing extremely large data sets across memory in a cluster of cheap hardware. The framework was originally developed at Google, as part of the company’s MapReduce programming model, but is now is available as open source. Its original intention was to optimize Web site search and indexing applications. But the software is much more generalized, so a number of companies are now using Hadoop on large-scale clusters.
“It’s getting very, very popular, said Silverman. “In my opinion, it has just as much benefit in the HPC space as the Internet space, because it’s [specifically] designed to help you distribute your application across a cluster,” he explained. And, he added, since it optimizes memory writes into larger chunks, it’s a good fit for block-oriented memory writes like you have with EcoRAM. Spansion has gone to the trouble to integrate EcoRAM smarts into the Hadoop source so that applications can take advantage of the hardware. The opportunity here is to consolidate these clusters by a factor of eight by rebalancing the compute-memory ratio with Spansion flash. Silverman says some OEMs are looking into integrating EcoRAM to build value-added Hadoop systems.
Since the company is still in Chapter 11, Silverman has to deal with the inevitable questions from system vendors about his company’s viability. “It certainly has not made my life easier,” he admitted. But according to him, it hasn’t affected the company’s commitment to product, nor the value proposition it offers the system vendor. “The fact that we’re launching this product is an indication of how the company views the importance of EcoRAM,” he said.