"NOR" Flash for Server Memory, not DRAMs?: Spansion’s recent announcement, with Virident (a local SilVal start-up), of a server NOR-based MirrorBit flash memory as a low power alternative to DRAM memory in data servers, was a many-faceted launch of a new direction for flash and a clear response to the emerging concerns that server farms are quick becoming power-out-of-control. The formal presentation, which can be viewed (approx 1:15 hrs) at the Spansion website, marks a new direction for servers, for flash in server memories, for NROM/MirrorBit technology, and provides many datapoints about the extent of the Data Center memory power problem and economics which EcoRAM is designed to address.

Though almost all LP DRAMs go into cell phones and PDAs today, servers and data centers, where costs and power are more visible and equally as critical, are driving a large interest, too. Several presentations at Denali MemCon in the past few years have focused on the Data Center Power Problem.

But it is not ready quite yet.

Those of you who have watched Spansion develop this NROM/MirrorBit technology and bring it to market in the majority of their NOR products (and some custom extensions of pure NOR, such as ORNAND and MirrorBit Quad Eclipse), to become the market leader in the cell phone NVM space…but have simultaneously been concerned about “where they go from here”, will probably not rest much easier with this new ‘technology announcement’. Spansion’s shares are not down 80% in the past 12 months for nothing, though profits (losses) are running about the same over that period.

It seems, no one really knows what happens next at Spansion, though they continue to be one of the largest NVM suppliers to the phone business, even with the entry of Numonyx earlier this year.

It will take another year or so to separate the wheat from the chaff insofar as how much innovation, market acceptance, potential for market dominance, and dare we ask, profits, Spansion can bring with this unique technology co-developed with Saifun, which they bought in its entirety more than a year ago.

EcoRAM is a custom flash memory design, built on Spansion’s 65nm-to-be-45nm process, which is essentially a NOR flash, but with the design customized so its operation insofar as writes and reads, makes it look like a DRAM in systems, only denser by 2x, far lower power (1/4 per MB), and equal or ahead in process ground rules compared to DRAM. There is no public data sheet for EcoRAM, and product is not available to customers, though enough have been built to achieve ‘proof of concept’ in a true-life server application, which was demonstrated during the presentation.

EcoRAM’s endurance is 10,000 cycles, but with proper software (Virident), bad blocks and bits can be foreseen and guarded against ahead of failures. Like DRAM and NOR, but unlike NAND flash, EcoRAM has XIP, bit-read and write access, and truly random access capability. No refresh current (power) and non-volatility are two important distinguishing flash features, compared to DRAMs used in server applications.

Background on Server Issues and Trends: As backgrounder for this announce, Spansion and its partners provided many dimensions about server farm power consumption which are useful.

Google has about 1M servers in operation today and are adding 300-400K/year; Microsoft has 400K and is doubling every year.

One watt of server operating power takes about another watt of air conditioning power to get rid of the waste heat; so, reducing server power utilization saves two watts of power for the enterprise owner. The clear message in the slide above, which lays IT DRAM power consumption (demand) against the installed base of US Solar Power's supply shows the stakes, and potential gain, from shifting from DRAM to EcoRAM in the serever universe. EcoRAM, if applied universally to all Internet Servers in the US, would reduce power demand more than the power contribution of the existing installed base of solar energy today.

Today in servers, MPUs use 25-50% of the total operating power, but the recent moves to dual- and quad-core MPUs, 64-bit architectures, integration of the memory controller onto the MPU, virtualization and dynamic clocking, have shifted a larger fraction of the server power to the memory, which ratio today is about 2GB/MPU core. Since servers have huge memory configurations to support so many cores, memory power has gone up in proportion and in absolute terms…total memory power consumption went up while MPU power consumption, despite far more processing performance, has remained close to constant.

The mad rush to redefine MPU Performance (= goodness, = high price, = value) from “GHz” to ‘instructions/watt’…almost in an instant… has had no such similar effect in memories, though the Spansion EcoRAM comes close to setting the bar at “Data transfers/watt”, or “GB/watt”. Clearly, the MPU trends of the past five years have put added pressure on memories to avoid being the ‘server power bad guys’, and so far, they have not responded. EcoRAM is one solution; perhaps LP DDR2 DRAMs, which should be in high volume in 2009, are another. LP DDR3 DRAMs, still many years away but whose definition is being discussed today, will certainly pay major attention to the power needs of the server space, where the costs are clearly evident, and significant.

Unfortunately, the technology outlook for the NROM MirrorBit, in terms of its true cost of production, scalability, read and write speeds and endurance…and long-term market acceptance… are all uncertain at this point. No one knows if some other memory technology or memory system approach is percolating in the background which is better and will come to market before EcoRAM gains traction. No one seems to have mastered the MirrorBit technology except Spansion…all its technology licensees have truly failed to bring the product and technology into stable, high-yielding mass production.

There are strong, critical voices of Spansion’s technology roadmap, among their competition; SanDisk and Intel suggest that the basic technology is doomed to be self-limiting or run into insurmountable technical roadblocks. Spansion has provided little public information to assuage its critics, though their strong and continuous presence in the phone NVM business is an important voice in their favor…someone keeps coming back for more!

Further, MirrorBit, for all its laudable attributes, has failed to be Spansion’s corporate salvation, as they are oozing red ink every quarter in an admittedly-tough NOR flash business environment. The advertised 40% fewer critical manufacturing steps when compared to floating gate NOR approaches has to help…but now that Intel and ST have become hidden financials as ‘Numonyx’, accurate cost comparisons have become more difficult.

On the other hand, Spansion was successful in bringing up SP1, its 2000 ws/week 300mm line in Japan, to 65nm over the past nine months, and will have 45nm running by year-end 2008. They have 32nm MirrorBit NOR pilot product running in the labs. They have made business adjustments of ‘make or buy’ to better focus their capital and internal development effort. They have internalized some NOR production and outsourced some non-critical and legacy parts to qualified third parties. And not a peep has been heard from Saifun since their acquisition by Spansion more than a year ago…no new technology licensees, no new income stream, no technology breakthroughs. With a full plate, a tough competitor and tough market, and limited cash, they are stressed to the max, or close.

But, with this announce, there is now more upside potential for Spansion, but timing being what it is, we will wait some more until some of these bold technology initiatives bear some real financial fruit.