Memory and Memory Subsystems (MSS), long accused of being the bottleneck to higher system performance, and 'throttling' the MPU with their high latencies and addressing limitations, now finds itself also as the "bad boy of power consumption." Other server system elements have made great strides in power reduction, but at the end of the day, memory 'owns' more of the system power consumption than any other system element. Server Farms, recently a sign of 'advanced internet evolution', are now in the Stink Pen insofar as energy consumption is concerned.

And though steps are underway to reduce power of the MSS, after all is said and done, 'It's the Memory Subsystem, Stupid'.

For most memory designs, too, cost reduction (= die size reduction) remains their foremost objective, though, to be sure, power has moved closer to the top of the design objectives list over the past few years. For the 50nm node products, the main benefit appears to be productivity: moving from 65nm to 50nm, plus the evolving benefits of superior designs, give 2x the die/wafer and more performance. Ceteris paribus, however, power comes down only fractionally, unless the 65nm design was 1.5V and the 50nm was 1.35V.

The tools for reducing power consumption among memories, with some liabilities for performance and build cost, are understood, though many possibilities also lie in wait. Reducing operating voltages, as has been the mode recently, is pure goodness; the high-performance speed bins still seem within reach; the user gets 20% power saving, almost for free...x2 to get the power savings for reduced air conditioning and assorted system build costs.

But, one might ask, if this is such low hanging fruit to go from 1.5V to 1.35V, why not go one more step without a pause, to 1.2V, and get double the benefit? And, even 1.2V does not seem to be any other than an 'industry convention'. Systems with 1.0V or even 0.9V have also been demonstrated in test vehicles, and seem fully achievable...though maybe not productionable for all the end markets that DRAMs serve.

Should more finely-segmented arrays be an option that goes back on the table, to reduce power? For complex server DIMMs, with 36 to 72 chips on each of them, is there a place for 'intelligent design', and more informed and advanced control of the elements that make up DRAM power consumption...refresh, chip and array select, power-down modes?

Mobile MSS power issues/concerns: In phones, and in the larger mobile segment, one of the great forces trying to squeeze DRAMs out, comes from the power-concern side: Swapping NVM/PCM for DRAM makes for nonvolatility AND fast read-write. The MSS (meaning, in this case, LP DRAMs + NVM) constitutes up to 30% of standby power consumption in mobile phones, so it is a big hit on talk time, and makers are 'motivated' to get the power down or out.

Laptops and Desktops: Though LP DRAMs of sorts have been around for many years, we know of no laptop computer which uses them. The memory subsystem is not the power hog it is in phones, and the savings by using 'LP' DRAMs over PC DRAMs, is not enough to justify more MSS cost with higher prices DRAMs and 'LP' DIMMs. However, we have seen in just the past six months, an early and quick take-up of DDR3 DRAMs in laptops, highlighting their lower power capability. Some of this is 'green marketing' to concerned buyers, but the power savings are real, though perhaps of little consequence in terms of battery life. BUT, laptop interests, which now make up more than half of PC unit shipments, and their low-cost cousin, netbooks, may be the force that tips the DRAM voltage discussion in favor of a quick move to 1.35V and then 1.2V DRAMs. What's not to like?

For desktops, cost is paramount and they are almost all plugged into the wall. But, so as they have wagged the DRAM tail for so long, maybe they are about to lose control of the DRAM voltage or LP feature roadmap, be hoist by their own petard, and eventually have to make do with 'what the other guys define, and make'...as others so long have had to dance to their fiddle.

With netbooks being the fastest growing computer segment, by a wide margin, and something of a tabula rasa as far as legacy system design constraints, maybe the Netbook marketplace will entice new DRAM lower power technical developments which then will migrate UP into laptops and PCs. With 20M+ units a year, and +2GB per system, maybe DRAM vendors will try for power differentiation in new and innovative DRAM and MSS designs.

Servers: Server DIMMs, historically pushing the chip-density and DIMM density envelopes, pose the microelectronic version of what the server owner reads on his utility bill. Around the server industry today, 8GB and 16GB DIMMs are the new bleeding edge, and even if they use the newer DDR3 DRAMs, or better still, DDR3L, servers are still rather huge power consumers, constraining system design, mandating larger power supplies, heating and cooling, fans airflow, 'hot spots', etc.

IBM Goes from 'Bipolar' to 'CMOS' in early 1990s, dislocates people, fabs and technology roadmaps: Maybe I am reminded of an anecdote I first heard when I went to IBM in 1995. Until IBM decided to reform its chip business in the early 1990s, and let loose of its 'backward' ways, all the close-in and high-speed system caches were bipolar...lagging the merchant industry's embrace of CMOS over HS bipolar which occurred as much as a decade earlier. But, in addition to the fab process problem, letting loose of all its bipolar capacity in Fishkill, NY (vacating Dutchess County, NY, as it closed down and eventually refit its fab to run CMOS), the major complaint came from the board designers, who had gone to great lengths to 'handle the bipolar power problem'...heat sinks, fans, intricate thermal analyses, limitations of chip layout density on the boards...system cost and performance constraints.

This kind of problem the 'memory industry' faces today; whether there is a simple "bipolar-to-CMOS" solution so close at hand as there was in the early 1990s for IBM, remains to be seen, but seems not probable. More likely, the solution will be piecemeal, and made up of many evolutionary changes over a long period of time. In servers, FB DIMMs was a start three or four years ago, to attack the limited address capability of Registered DIMMs and, for sure, some things were learned along the way. But FB DIMMs had no traction, and a limited following, and were subsequently replaced by better (and denser) R DIMMs..and been wholly replaced by a separate discussion for the follow-on generation of server DIMMs, so called Load Reduced (LR) DIMMs. FB DIMMs was maybe the last 'technical challenge' that was pure performance-driven, with little regard for power, which eventually came back to bite them in the AMB.

But the power problem, so long pushed aside in favor of 'more performance', is real, it is here, and it will get worse before it gets resolved.

Fortunately for the memory industry, signs of recovery have been apparent for about the last two or three months, so there was reason for business optimism. But, no matter how much the profits and headcount of Silicon Valley's corporations have been impacted during this worst-ever downturn, new ideas and technical advances just keep on coming. This year's theme, "Beacons of Innovation", was apropos, and MemCon brought new products and technologies to the fore, and improved understanding to more mature discussions and products.

Now nearly forty years old, the 'semiconductor memory industry' is hardly maturing. The industry's future, the technical and standards roadmaps, and the issues are as hotly debated as ever they were, albeit at line geometries that are 100 or 1000 times smaller, and costs that are 100 or 1000 times larger. And, we should add here, risks that are also 1000 times larger. In chips, everything scales up, or scales down!

The full MemCon agenda can be found at www.denali.com/memcon

My favorite take-aways and impressions:

SSD talk; still talk: The topic of SSDs has been a mainstay at Denali's MemCon for about four years, often viewed as the Next Big Thing and market driver for NAND Flash. Indeed, one might surmise that SOME substantial fraction of the est. $35B invested in NAND manufacturing capacity in 2005-08 was exactly for the purpose of supplying that market demand. But with still weak take-up of SSDs from netbooks, to laptops through desktops, and onward to the enterprise market, the bloom is off the rose more than a little, to hear what most speakers were saying: the message is still positive, but manufacturers are clearly disappointed with the size and commitment to today's bona fide "SSD" business level, and probably tomorrow's SSD market outlook. NAND prices are still way too high compared to HDDs, despite declines of 50-60% per year for those same past five or so years. No sooner than one technical 'whattabout' raises it head, and is resolved, than another shows up to again retard the significant adoption of SSDs in the top-to-bottom computer line up. To the rescue, to continue the march to improved storage performance economics, come more improved caching systems (see Denali CTO Mark Gogolewski's talk), more software tools and a better understanding of system 'usage models', which get more out of less. This evolutionary process, without huge leaps into the SSD unknown, nor buy-in of fast evolving, expensive and (some say) unproven technologies, is making the transition to total solid state both more measured, but also less risky.

Still today, big unknowns still lie ahead for SSDs (or, as it were, the 'known unknowns' like "what will be the cost, and performance impact of x4 cells and sub-25nm processing?), just as the industry probably under-appreciated the issues it would face in displacing a significant fraction of those HDDs that were in computers, from the vantage point of back in 2004 and 2005.

Performance NVM market: In fact, even though ONFI has been here for several years, and toggling for only slightly less, the high speed NAND flash market is still in its infancy. It is still viewed as having undiminished promise to deliver NAND and SSD bandwidth in multiples of what is widely available today. Ed Doller, in his kick-off keynote, showed a slide of today's NVM applications, dominated today by NAND, in which a huge fraction of the GB lodged in low-performance end of the spectrum: USB cards, DSC chips, Compact Flash, MP3 players. No wonder they can use x3 cells, with sometimes 10K or even 1K write cycles endurance.

There remain strong believers in a high-end promise of the fundamental NAND technology, which includes the vast majority of NAND-based SSDs, as well as heretofore unrecognized applications. Indeed, much of the Performance NAND effort today is truly pathfinding and technology development, and every week brings a new product announce which sets a new high standard for R/W performance, endurance or cost-performance. All in all, 'Performance NAND' is just getting started, with standards, technical capabilities, and applications all evolving lock-step to exploit what is widely felt to be NAND's huge performance and low-cost potential..

Almost from the beginning, no matter how much the talk was of HDD/vinyl/tape displacement markets, it was recognized that the "NAND-market-fitting" exercise would be comprised of developing its core technical promise, concurrently with developing applications that can project those technical benefits into a product that the market would buy. There has been a constant back-and-forth between NANDers, saying "Here's what we can offer, what we can do"...and users who were seeking to define their uses, and refine their spec and performance needs, to match those of the silicon suppliers.

The devil is in the details: Heading down that same 'spec' and 'performance capability' road, we also heard in Numonyx, EasyCo and others' talks, that software and improved system design can deliver significant improvements in NAND endurance, bandwidth and RW performance...only it is not here yet, and faces an uncertain market--more of the NAND 'Double Do' Problem: what can the silicon do and what does the application need. These are very much two moving targets, with the entire industry trying to sort it out. These several talks posited a new lever that the designer could pull for system performance improvement: delving into the exact details of how the device was to be used, or was being used...'usage models'...and applying more system and silicon knowledge to take advantage of unique features of the application or the silicon.

Repeat after me: Market Fragmentation is coming...or maybe not: If I heard the words "Fragmentation of Product Mix" once, I heard it a hundred times in the course of MemCon. The market needs a more diverse mix of DRAMs and NAND flash to best serve its technical needs and low cost requirements. For markets which place huge pressures on commoditizing their constituent products, and 'terminate with prejudice' any attempt to add bells and whistles, this is a new froce that has to be reckoned with...and we can only wait to see where it comes down. It does not appear that gravity can pull the pieces together, either for NAND or for DDR3 DRAMs. New products, and new applications of available flash and DRAM technology, will appear and be accepted or rejected by the market.

For DDR3 DRAMs it is now, PC DDR3 and DDR3L(V) for 'low voltage', which, instead of the Standard 1.5V, can be 1.35V and maybe eventually 1.2V), plus Low Power DRAMs...LP DRAM, Mobile DDR and LP DDR2, plus Graphics DRAMs (or G DRAMs). For NAND, the SLC of 2004 has become MLC and now soon x3 from some vendors; most expect an eventual x4 cell that can store four bits, with yet t.b.d. performance and cost for the 'degraded sockets'...and another shoot out the top for HS NAND flash: toggle of ONFI interfaces and roadmaps.

Efforts to pull all these contradictory and conflicting trends together, to reduce the product set to a small number of parts, are being thwarted by application trends that need either low cost or high performance, to nearly the exclusion of the other characteristic.

All of that is just my 'tip of the iceberg'. There's much more on the Denali MemCon website, www.denali.com/memcon, which has PDFs of all the presentations, actual webcasts of some presentations (with audio), and a MemCon Community website with discussion links to like-minded attendees and others interested in the discussion and topic of this year's MemCon.

Again, we'd like to thank our Platinum Sponsors, Samsung and Rambus, our Gold Sponsor, Numonyx, and our Silver Sponsors, EasyCo, SanDisk and SPMT. We'd also like to thank our track sponsor and Conference Partner, Web-feet Research for their support. We know how scarce funds have been since the fall financial meltdown, and appreciate their continued support.

What has changed since late 2008: As late as 3Q and 4Q08, DDR3 DRAMs could only be found in high-end game PCs, where they were attractive for their higher performance, and where the market could support their higher prices. In early fall, DDR3 DRAMs still cost 2x-4x DDR2, though admittedly, DDR2s were at something of an all time historical low of 60 cents. Today, DDR2 prices have come up a lot, and are now in the $1.05-$1.20 range for 1Gb ($9 or $10 for 1GB DIMM), closing the gap from the bottom. But newer DDR3 designs have also reduced the earlier die-area disadvantages, thus reducing production costs. And DRAM makers seem to have decided, 'If we are shipping $1 with every 1Gb DRAM, it might as well be a DDR3 DRAMs, which has a future...and that Taiwan's price bombers do not have yet to any appreciable degree.'

At the crossover speeds, DDR2 800/1066 and DDR3 1066, prices are very close to one another. At DDR3-1333, there is a mixed and sometimes significant premium for DDR3 compared to slower DDR2s, which, if the history of DDR1 and DDR2 pricing is any guide, will be whittled down to nothing in a few quarters' production and use. DDR3-1600 is still not found in any PCs that we have seen...but for DDR2, recall DDR2-533 was kicked off 3Q04, DDR2-667 took up 1Q06, and DDR2-800 was really a 'post-mid-2008' phenomena (and still growing, displacing -667s)

Low Voltage DDR3: Last year, the industry devised an alternate Low Voltage Roadmap for DRAMs. The standard DDR3 spec called for 1.5V operation, which had been decided nearly a decade ago. The low voltage roadmap agreed on last year calls for a 1.35V option, designated DDR3L, which saves about 20% in power (subject to many other considerations and conditions, of course). But most vendors can hit the performance spec even with this reduced voltage, so it is expected to be adopted widely; all major DRAM makers have 1.35V product and roadmaps; Samsung and Micron announced DDR3L server DIMMs and SO DIMMs a few days ago, for (almost) everyday applications...and it is only a matter of time before "L" comes to dominate DDR3 production. There is also an even more agressive 1.2V discussion and emerging DDR3 (or DDR4) spec, but it is still too remote to forecast the timing and extent of adoption...but the preponderance of heat-spreaders on DDR3 DIMMs should give some indication of where the power problems lie.

DRAM makers have consistently been able to produce the top speed grade at high yields, for DDR1 (>-400 to -600) and DDR2 (>-800/1066 to -2100), so they have wisely been reducing the operaing voltages, for 'green considerations' without compromising premium sales.

For users, this "L" roadmap is 'free' power reduction.

Off the charts, too: While the DDR3 specs are -800, -1066, -1333 and -1600, an additional bin at -2133 is under discussion...mostly aimed at PC makers. However, there is a vibrant market for gamers and overclockers, which offers many other variations...-1800, -2000, and Elpida's fastest-to-date, 2.5GHz 'technology showpiece'. These varieties may calm down in the next twelve months, as the -2133 spec moves forward for affirmation.

For most vendors, DDR3 will really take of with their nom 50nm DDR3 designs, which will be hugely productive (die/300mm wafer) and could be DRAM maker's ticket to profitability in 2010.

The Company: Unity has 42 employees, with a wide range of semiconductor pedigrees from Inmos, Micron, Fairchild, Ramtron, Simtek, and AMD. It is headed by Darrell Rinerson, formerly an Executive at Micron Technology, who with two others founded the company in 2002. Darrell is the lead patent developer listed on most of the patent plaques lining the entry hall of the companies' Sunnyvale facility.

Now seven years old in development, and with a total of $75M in three rounds of funding, Unity is confident enough in its technology and roadmap to start talking to manufacturing partners, as it concurrently moves product and technology development from 'proof of concept' to 'characterization' and something close to 'production-readiness' over the next two years.

Should this technology pan out in a way close to what early analyses and tests have indicated is possible, at equivalent process nodes, CMOx can be expected to have 4x the density of NAND flash, 5x the write speed of NAND, and require little in the way of special processing flows. Compared to MLC NAND's 2F2 cell size, CMOx's is merely 0.5F2, and it has no transistor to limit extreme scaling. Though still two years away from production-readiness' CMOx memories are being positioned as something of 'The Mother of All Storage Class Memories' that are taking aim at sockets where conventional silicon cannot reach for price or performance reasons. Conceivably, CMOx could even replace HDDs, and expand markets into heretofore unidentified markets, applications and products.

These are 'Great Expectations', to be sure, and a long road to travel between today's 64Mb test vehicle, now undergoing characterization in Unity's labs, and its 64Gb device, scheduled to be available in 1H11.

Novel manufacturing process flow: With the Unity CMOx storage cell technology innovation, other important business and manufacturing decisions can change. For one, for the first several masking layers, Unity's wafers can effectively use a standard bulk CMOS process, at a back generation process node, for the FEOL. They can buy common 90nm processes wafers for their CMOx memory feedstock. These 'pre-uncommitted memory arrays' are then transferred to a more state-of-the-art BEOL, for the metal oxide and metal layers, and deposition and building of the actual storage features.

'Go to market'...how, exactly? Furthermore, in the midst of today's memory meltdown, Unity is very interested in its own go-to-market model, and intent on devising a method of taking its product to market without creating the overly-commoditized bloodbath we have seen for more than a year in DRAMs and NAND flash. For this, their 'plan of record' today is to join forces with an existing memory manufacturing partner, and have limited production sufficient to service that market that can use the superior performance of CMOx (at a price, to be sure). Of course, they will have to weigh carefully the option of reducing prices and broadening the market...but this decision will have to wait until the product itself is fully characterized and the existing and potential 'high value niches' better understood.

Unity is convinced that broad licensing of CMOx would collapse the market into a food-fight, and leave everyone with nothing, as it has, periodically, for traditional DRAM and NAND flash licensing programs almost since the industry's inception.

Licensing, adoption, production: evidence and options:

To add some perspective to this problem, specifically that of the owner of a potentially valuable "Memory IP", it is worthwhile to see other ways of 'going to market'. These days, Pure IP Developers are denigrated in some quarters as being 'patent trolls', leaches and worse. They 'make' nothing, and are viewed as riding on the backs of those who do make products. Historically, IP developers made products that demonstrated the value of their IP, and got paid in product profits. Sometimes 'secret sauce' of proprietary but undisclosed and unpatented technolgies, remained out of view for years from competitors. But, especially among the industry's early founders, broad cross-licensing took place, and no one was denied access to any other technical developments in their field, though license and royalties flowed continuously. That started to change in the mid-1980s with a huge TI patent offensive against DRAM makers, and was turned on its head entirely with the founding of Rambus in 1990s, who was the first pure-play IP company in the industry.

The conundrum is that however much the IP costs, the cost use it in a DRAM/memory/semiconductor product, and take it to market, is far greater, with today's $3B memory fabs and established channels to customers. A standalone IP developer has to find a manufacturing partner to bring his ideas to life in the marketplace, and to profitabliity. And, even if one has an allegedly superior capability, it is no guarantee of a big money stream, and there is never any guarantee that it is 'forever.'.

Rambus DRAMs have demonstrated superior performance to standard DDR1-2-3 for more than a decade now, but have not enjoyed significant take-up in the market, and have thrust Rambus themselves into a never-ending sea of lawsuits...which pariah status Unity seeks to avoid.

There are, of course, instances of sustained significant technical leadership and profitability from time to time in the semiconductor memory business, but, as all the players have the same basic tool box (or, sandbox, since its silicon), eventually Intel loses its HS 1K and 4K SRAM monopoly to Hitachi, Inmos and AMD; Samsung loses its 8M WRAM monopoly to DDR1 and wide IO, Standard G DRAMs and IBM loses its UHS SRAM sales leverage with customers Sun and HP, (sustained at $140/8M unit for years) to Samsung, and then GSI (though other factors were at work there, too.) System manufacturers are highly adept at working around costly components in their systems (see "History of RL DRAM"), as are the thousands of engineers working on new materials and methods at working around what appear to be an insurmountable patent position. Wang Labs x9 SIMM patents, earning them a cool $90M in the early '90s, became worthless with the onset of x32/x36 DIMMs. CDMA technology from Qualcomm is a non-memory example, with which they have generated a significant monopoly profit stream AND widespread adoption for many years, though not without a battle at every juncture.

Rambus, Wang Labs and sometimes Qualcomm (and formerly, TI) have engendered some intense bad feelings in the industry by the way they 'exploited' their IP...which they feel is just getting paid for their technical contributions, "and if you don't like it...'

Balancing what some would call 'IP greed' with market needs and 'ability to pay' is a hard act to manage, as we have seen many times. Broad licensing and cross licensing of IP has been the common industry practice since its earliest days; good for the consumer, but maybe not for the IP holders or other intermediaries, who cannot differentiate their products.

Unity's manufacturing partnership plan is their way of limiting access to their technology, bringing production only to a level of a manageable market presence that shows good profitability in those applications that can appreciate CMOx unique capabilities. Once the show hits the road in 2011, we will see how well this works. (Saifun faced a similar quandary when they developed NROM technology...be an IP company or build a fab? This was discussed in an early DMR article.).

The Challenges Ahead: This looks like a promising technology worth watching. Only more time and more development and characterizations, and more critical evaluation in real-life applications, will tell us if CMOx is as good as it now looks. Many other technologies looked good for a while, but ran into cost, scalability, or performance challenges that could not be successfully overcome. MRAM was all the talk 3-4 years ago, but has given way to Phase Change Memories (PCMs) today as the most talked-about heir apparent to the extensible silicon roadmap on scaled NAND flash. With technical roots, and claims of Universal Memory Domination dating back at least to the early 1980s, Numonyx informs us in a press release from last December that they shipped the industry's first PCM memory for revenue in 4Q08.

In addition, the Status Quo is not standing still; silicon NAND marches on, downward in cost, towards and beyond 30nm processing by the time the first CMOx 64Gb device is shipped in 2011.

But if these claims come anywhere close to standing up under two more years of development and scrutiny, CMOx's domain will extend far beyond where we see silicon NVMs applied today...'cheaper and faster' is a hard to beat combination.

Abstract: Taiwan wants to be a big player in DRAMs, which we believe is a losing proposition and a strategic mistake, not only for Taiwan but for all DRAM makers. They should embrace their strengths and domination of the worldwide foundry business, repurpose some of those DRAM fabs to make next generation logic, and expand their design and logic foundry franchises. They have much more to lose in foundry than they have to gain in DRAMs...in fact, DRAMs is a sinkhole for money and has been for more than a decade. In the current climate of industry consolidation, megafabs and hegemony for large markets, and the move to fablessness as a superior business model, it is not even an open question which path is better for Taiwan: Foundry logic trumps DRAMs, now and forever.

Taiwan Wants to Make DRAMs After a very tough year-long DRAM market that began in the waning days of 2007, we are now well into the next phase...a more than nine-month agony emanating from Taiwan as is searches for DRAM relief in a variety of proposed consolidation arrangements among the island's DRAM makers. The "Powers That Be" in Taiwan, want to own their own DRAM IP and designs, to consolidate operations to reduce redundancies in design and development, to gain scale, and be large enough and tough enough to compete (mainly) with the Koreans. DRAM market leader Samsung and side-kick and rival Hynix, which own about 30% and 18% of the DRAM business, respectively, each makes more than ALL of Taiwan's DRAM makers combined: Powerchip, Nanya, ProMOS, Winbond, Rexchip and Inotera, plus or minus some allowance for Taiwan-based DRAM production which is owned by and built for non-Taiwanese DRAM makers (namely, Micron (JV Inotera) and Elpida (Powerchip and JV Rexchip), now that Qimonda is out of play). Taiwan has set for itself a formibable goal, from a weak position today, in a very tough and unforgiving market.

DRAM makers, worldwide, lost about $20B in the 2001-03 aftermath of the dot.com bubble bursting, and have lost another $30B this cycle, from late 2007 though the middle of 2009.

Taiwan's DRAM makers themselves have lost about $1B per quarter for the past six quarters (see table below), and hold huge amounts of debt, said to be in the vicinity of US$14B. When the 'profit' numbers for 2Q09 are tallied, another $1B will be gone forever. And, despite a long time coming in discussion, no firm plan seems to have emerged from the Taiwan Memory Corp., except to anoint Elpida with the title of "Taiwan's DRAM IP Savior", to plan to invest about US$300M for a 10% stake in Elpida to bolster its cash position, and to merge ProMOS into the larger TMC/Elpida entity. What this may mean for long-time Elpida partner Powerchip and its Rexchip JV, is not said, but likely they are automatic members of Club Elpida.

Details in the public domain are scarce and it is truly a moving target, but clearly it is much more modest in scope than that which was proposed last fall, which had a scale that was rumored to be as high as US$6B in available funding. We will not know for sure, until there is a formal announcement, and maybe not even then, as huge uncertainties are sure to exist about the arrangement.

What we do know is that it is now about nine months since conversations began, nothing has happened except Taiwan's DRAM makers have kept capacity on line, continued to make DRAMs, and lost another $3B.

The Koreans are the dominant DRAM players today. With their unique corporate culture, with huge financial resources at their disposal and a gambler's penchant to go 'all in', alone, make them formidable. Add to this their incumbent's advantages, their market position, their product set, design skills, manufacturing capacity, technology and experience, and one can see that they are way ahead of Taiwan in many important measures of market success. If history is any guide, they would probably not be dislodged without a large and expensive fight. "Last Man Standing" takes on a whole new meaning when Koreans are in the market.

Samsung's Greatness: Samsung makes twice as many DRAMs as all of Taiwan, owns most of their own DRAM technology (self-developed since Yong Park designed their 256K on Old Ironsides Drive in Santa Clara in 1984), has many DRAM designs in production, concurrently, down to the 40nm node as technology-proving vehicles. They are the driving force in JEDEC in both the NAND flash and DRAMs standards bodies, have a broad and profit-protective portfolio of LP DRAMs (50% share), Dense Server DIMMs (80% share?), and are the G DRAM market leader (>50% share). They are the original Market Creator and Market Leader for NAND flash, with a relatively stable 40% share, and a competitive #3 (belatedly) to Spansion and Numonyx in NOR flash, which remains a $5B business even in 2008.

Samsung has led and dominated every memory market they participated in, for as long as they wanted.

Furthermore, for Samsung, it would probably be a bad business decision to chase more than ~35-40% share of these huge, highly commoditized markets, susceptible to 'price dumping' by DRAM and NAND minions at home and abroad, and general production gluts (as in 'today'). Below-cost prices may hurt Powerchip a lot, but will impact Samsung, whose share is 10x larger, way more, in absolute if not relative terms. And, there has always been a "suicide bomber" in every DRAM downturn since 1979...they kill themselves, to be sure, but take as many innocent bystanders with them, who are too close to the explosion. The recent trend to "Pure Memory Players" exacerbates this potential: "We will not exit memories gracefully, if and since that is the only thing we have." Few memory makers have ever 'reinvented themselves'; most died of low prices and high losses.

With high shares in all major memory markets, Samsung may not grow its memory share more from today, but will improve its memory profitability; for growth, they will have to find greener pastures and still-unfamiliar markets.

Now, add in Hynix, which, by itself, is also about as large as all of Taiwan DRAM makers combined, and who is forever in the thrall of its bankers, anxious to see their investment pan out. Given their apparent willingness to throw good money after bad, and firm in the belief (or so it seems) that all that money invested can again turn green...if the competition gives up first. Unlike Taiwan, as well, Hynix has some good differentiated DRAMs, used in graphics and LP Memory applications...plus NAND flash. Hynix by itself, a distant second in all things memory to Samsung, is still out of reach for Taiwan.

Taiwan, on the other hand, gets hand-me down processes from their partners, six months after development (IBM, Qimonda then Micron to Nanya, Qimonda then Hynix to ProMOS, Mitsubishi then Elpida to Powerchip, Qimonda and self-developed to Winbond), are centered in the sights of commodity DRAMs (having only a small fraction of differentiated DRAMs), and minimize their development costs by paying for technology transfers ahead of time.

We believe that Taiwan, in focusing on strengthening its DRAM capabilities to fend off Korean DRAM dominance, is placing a bad bet, and the wrong bet. The world does not need another DRAM maker, it needs some substantial DRAM capacity taken off line, and restrained DRAM investment going forward. After two strong DRAM GB growth years, of 90% and 70% Y/Y in 2007 and 2008, respectively, 2009 promises to be a bust in terms of DRAM demand growth: pundits are calling it 20-30% Y/Y, the worst in a decade. Reductions in wafer starts are the only way to keep DRAM GB growth down that low for a year, as the steady and 'inertial' yield improvements, scheduled die shrinks and winnowing out back die versions naturally leading to more than 35-40% bit growth without adding a single wafer.

Competitively, as well, Samsung is just too big, too entrenched, and too far ahead for Taiwan to compete with head on. Samsung Electronics is a huge $60B/year electronics business, with high market share and profitable positions in mobile phones and LCD panels, both of which can feed their chip business with whatever it takes to dominate the business. Rightly or wrongly, 'fairly' or 'unfairly', this is an important aspect of Korean corporate culture that must be reckoned with. Hynix' 'family and sponsors' will stop at nothing to stay afloat, as we saw in the 2001-03 dot.com bubble aftershock.

If Taiwan wants to do battle with Samsung/Hynix/Korea, they should do it in foundries, logic and manufacturing logistics, and high-value add designs...where they have a head start and a huge advantage...vs. their 'fire and forget, huge commodity' mentality that permeates most memory makers' way of thinking and doing business.

Sure, it will/would be a good story if Taiwan were to win and establish a sustainable (and profitable) DRAM position. The story would be great: "'The Comeback Kid', How Taiwan's DRAM Business Rose from the Ashes to Become a World Leader"...but what if they lose?...and why not apply themselves in an area where 'Victory' and 'Success' are far more likely, and aligned more with their current strengths and market trends?

Samsung's Foundry Possibilities: Further and finally to this argument, Samsung, The Memory Guy, took 40nm Xilinx FPGA business away from UMC within the past six months, and was reported sniffing about in Taiwan's foundries for skilled employees to hire away. (see relevant footnote on Samsung DRAM offensive, ca. 1990, below).

Although Samsung's System LSI business and 'other' have languished in ignominy for a decade, they show sparks of excellence, and know that they cannot overtake Intel (or grow their sales and profits much more) by relying only on memories. They have been a part of the IBM-led Common Technology Platform roadmap development consortium for many years; they have their own developing ASIC group, and take in foundry (in addition to Xilinx, noted above.) Their System LSI Group is running about $2B a year...a pittance by Samsung memory standards, but it would easily rank in the Top 20 in world chip sales by itself.

Taiwan's Strengths, Taiwan' Exposure and Taiwan's Opportunity: Why "Taiwan" would want to challenge Samsung, or Korea, in DRAMs, is a question that has not been satisfactorily answered in the discussion so far.

But a more important way to look at the problem is to ask, "What are Taiwan's strengths and what are its opportunities?" As we discussed in an earlier BLOG, Taiwan is more of an entrepreneurial culture, against the large Chaebols ('authoritarian' family-owned businesses) of Korea. It is more democratic in decision-making, and less prone to be run by diktat, as the Korean companies are. In Taiwan, there are fewer cozy government-banking-corporate tie-ins than there are in Korea, or than there are in Japan...which tie-ins have some competitive advantages when played against smaller, and more disassociated entities such as are found in Taiwan. Anyone can play in a market with the Koreans, but they play at their own peril, and usually with huge disadvantages in the power that can be asserted in a market, seemingly without regard to profit or damage potential. Though there is a fine line between 'thinking long term' and investing foolishly, somehow Korea manages to make it work. American driveways filled with Kias and Hyundai Sonatas attest to their perseverance and skill.

To see Taiwan's strengths, one has to look no further than TSMC and UMC, the #1 and #2 pure-play foundries, with 50% and 15%, respectively, of the worldwide foundry business. TSMC had sales of more than US$10B in 2008; UMC had sales of more than US$3B in 2008.

This is the Glory of Taiwan, and it is right in the path of perhaps the most important trend in the industry today, 'fablessness for finer line geometries'. It is these next-generation processes that are increasingly expensive to develop by oneself, and will drive more and more companies to lean on foundries as the industry moves forward to advanced process nodes, in the coming years.

In the Industry's Consolidative Phase, Fablessness is Fabulous. As process development costs hockey stick up at 65nm and then 45nm and 32nm, as fab costs approach $3B per megafab, the drumbeat of fab-lite, fablessness and 'mere' design houses, resonates more loudly...and more profitably. While outsourcing of chip manufacturing is said to be in the vicinity of 20% of total chip sales in 2008, an increasing fraction of the industry's profits are to be found there...far from the idle fabs, rusting with depreciation as NAND and DRAM demand lulls, far from those huge capex budgets which strangle the higher-value-add design programs that form the value proposition in end systems into which they are sold, and shut off the profit stream of the whole of the silicon food chain. Fablessness also takes one far from the specialty processes that require constant attention to maintain competitiveness, against products and markets which are forever morphing, and/or wanting to morph into other markets, process requirements and functionalities. This is a huge exposure to a smaller producer...to develop a one- or two-generation process roadmap, then abandon it.

This is the accelerating wave of the coming industry recovery, as high-value logic, ASIC and SoC implementations make Intelligent Designs more production-justifiable and more valuable to everyone in the silicon food chain, than 'mere' bulk bits' (sold at a loss, to be sure). With the current industry 'pause', painful as it is, Taiwan now some slack time to rethink its mission, to take stock of its available resources and skills and reset its industry direction. Through its leading foundries, it is uniquely positioned to become even more dominant in the next stage of industry evolution. Sinking more money into DRAMs is a fool's errand, that will bring good to no one.

Footnote for Consideration: Samsung DRAM Offensive, ca. 1990 Samsung (as "TriStar"...its corporate/family logo) came on the DRAM scene in about 1983, after being the best Korean 'regional semiconductor player' for many years. With great fanfare, LG Goldstar, Hyundai (incarnate in the US as "Modern Electrosystems"), and Samsung all joined hands and jumped in to the world chip business. Not being shy, soon thereafter, Samsung declared its goal to become the #1 Chipmaker by the year 2000. Well, not quite, and not quite yet. But recall that this declaration was made when Intel was about #10 in the industry rankings, behind five Japanese (NEC, Hitachi, Mitsubishi, Fujitsu, Toshiba), TI and Motorola, and two Europeans (Philips, Siemens = Infineon).

But, barely were the Korean's feet wet when the Tsunami of 1985 hit, and forced major retreats and reconsiderations by all three companies. On the surface, they retreated to more traditional corporate names: "TriStar" reverted back to "Samsung Semiconductor", and "Modern Electrosystems" became, nce more, "Hyundai Semiconductor". And, "Goldstar" stayed the same, at least for a while.

They continued progressing on the technology front, but their big day only came when the Japanese Meltdown began in the early 1990s, leaving a huge market void...a DRAM void...to be filled with Korean DRAMs. Toshiba, Hitachi and NEC rode the end-of 1980's wave, but Samsung, with more money and more nerve, was in hot pursuit. When Toshiba showed signs of faltering in 1990, and whose 4Mb DRAM was not the market leader that their 1Mb had been, Samsung came into play with a vengeance.

Throwing down the gauntlet, Samsung hired the Top Guns from right from under Toshiba America's (TAEC) nose: Keith McDonald, TAEC's Director of Sales, Mark Ellsberry, the TAEC VP of Memory Marketing, and Mike Crawley, the TAEC Director of Distribution and Sales Reps, all came over to work for Samsung. Within a year, these three had a huge impact on improving the Samsung image to its customers, brought immense experience to reshape Samsung's DRAM business into the market leader it is today. As soon as 1992 came in, Samsung had risen to the top of the (DRAM) heap, and became the market leader, never to be headed off. Where each generation of DRAM had had a different market leader from 4K to 4M, Samsung has led every generation since.

This article continues our earlier discussion about the same topic from last week, entitled "...Part I."

It includes highlights of the most significant departures from the DRAM business over the past two decades, and describes, in broad terms, the experience of these DRAM makers in the market, historically and the events that led to their departures. In some cases, details relevant to today's 'prospective DRAM players' are noted, as well as some of the 'behavioral attributes or special situations' that make these cases noteworthy.

We have found, in preparing this discussion, that the internet provides a wealth of (undigested) fact and opinion about each of these companies; what is discussed below is believed to be accurate, but not comprehensive by any means.

There is much to be gleaned from these "DRAM exit" histories that is useful for any company, but especially useful to companies that find themselves with large exposures in commodity products, in tumultuous times and in volatile markets. The exits are not always (or ever) clear, but there is quite a range of outcomes from situations that might outwardly look much the same...managing the exit can make a world of difference in the financial impact. What constitutes 'the business' or 'the assets' varied widely from case to case, and, as one can see, not all 'exits', in retrospect, were at all optimized given the options faced.

Case History #1: UTC sells Mostek division to Thomson...to SGS Ates, which shuts it down: Mostek was founded in about 1968, out of TI, and located right down the road near Dallas. Like TI, they were a vital force in the early DRAM business during the 1970s, and made many valuable technical contributions (See subsequent Mostek IP licensing discussion, below). They were the original 'Standalone Memory' company, dominated by DRAMs, though they did have some small positions in SRAMs, Zero power SRAMs (which also spun out competitors like Benchmarq), EPROMs and Mask ROMs. They were also innovators in communications chips and were a Tier II player in MPUs. But forever and always, DRAMs was their glory. In 1979, in a fit of 'we MUST be in the semiconductor business' that swept over the industry, Mostek was bought by United Technology Corp. (UTC) for $345M, on the flimsiest of pretexts that their Carrier Air Conditioning equipment would eventually need to join the micro-electronics age. (Yes, I know, this sounded just as flaky then as it does today. But it was a good indication of UTC's naivete, which eventually factored into Mostek's demise.)

The decision looked very smart, as Mostek rode the semiconductor and DRAM up-cycle in 1979-80, making the UTC buyers look like geniuses. The 1981-82 downturn caused some significant distress, financial and emotional, and showed cracks in UTC's commitment. The next cycle, running from 1982 into 1984, restored their confidence. But 1985's crash brought huge pains to Mostek. They had been a very significant DRAM play, having about 11% share in a DRAM market whose power and market share was far more widely dispersed that today's concentration. But continued losses and incredibly low DRAM prices brought pain and suffering and showed UTC that they had placed the wrong bet, and that 'the chip business, and DRAMs in particular' was a different kind of business than their other mainstream businesses, such as jet aircraft engines, elevators and air conditioners. In mid-1985, after a major upheaval within UTC, UTC's management decided to unload their Mostek Division, which they did at a bargain basement price of about $105M to the French electronics concern, Thomson CSF. This was a painful and decisive action; the sale took place at the bottom of the industry's worst-to-date cycle when asset prices were rock bottom. (Gaining compensation for "IP" licensing was not a recognized business activity...profit center...yet, until TI pioneered in 1986-87. But Thomson was decisive, UTC took their several hundred million dollars worth of lumps, and moved on.

Mostek struggled under the management of Thomson for a while, weathering one layoff after another as the downturn dragged on for years. It was then combined with SGS-Ates in 1987, to make what is now ST Microelectronics, or STM.

STM did two things: They shut down what remained of Mostek's position in DRAMs, but saved and used its Texas facility, along with small product lines like specialty SRAMs. More importantly, someone then said, "Let's take a look at Mostek's patent portfolio", which had been kept in some kind of locked drawer for a good while, and they found valuable DRAM and other memory IP, which netted STM about $450M over the course of the next few years, in license fees from other DRAM makers and downstream players.

In several stages, from the onset of the 1985 DRAM crisis (which also took down Intel and Inmos, as well a convinced a number of small DRAM dabblers to "cease and desist."...ITT, National Semi, ATT, Fairchild and AMD...and more), UTC-Thomson-STM-Mostek went from 11% DRAM market share to 0%, putting about 10,000 workers out on the street, and costing a few hundred millions of dollars of losses to whoever owned the company at the time. Then STM made it all up with their "DRAM IP closet" discovery and licensing initiative. STM also bought Inmos in 1989, with its DRAM and SRAM designers (and Transputer MPU), and fabs in Colorado and UK, and remained competitive in SRAMs for a while longer.

STM was the "Great Semiconductor Consolidator" of the era, and did so quite profitably. They liked memories, too, but only from a distance, and only as something that periodically depressed other company's share and asset prices, and gave ST huge benefits upon acquisition. Indeed, except for a modest position in Slow SRAMs and Mask ROMs in the early 1980's Video Game bubble, STM themselves stayed out of commodity memories almost entirely, until 1987, when they joined the EPROM fray, which product line they retained even through the morphing of EPROM business to NOR flash in the early 1990s, which they only a year ago, merged into Numonyx.

Case History #2: Intel's Painful Exit, 1985: Along with TI, Mostek, and some Japanese efforts (Hitachi and NEC, mostly) Intel did more to develop the early DRAM marketplace during the 1970s than anyone. Always strong on technology and weak on low-cost manufacturing ("Intel could not ship an empty package for $5 if they had to."), they were frequently first to the new generation of DRAM, locked up key accounts, then retreated or coasted once the higher-volume, lower-cost manufacturing set moved in. During the cyclical upturn through 1984, Intel maintained about a 4-5% DRAM market share, but they were already on the retreat when the market turned south in the waning days of 1984. Here, again, Intel had (prematurely, it turned out) tried to muscle the DRAM market over to CMOS versions, as an innovative and leadership specialty part ("I have seen the future, and it is CMOS"), in place of the historic NMOS at the 64K to 256K densities; It was a right technology direction, but one generation too soon: The 256K mainstay DRAM remained NMOS, though Vitelic and NMB made scads of money selling 256Mx1 CMOS DRAMs as parity DRAMs for x9 DIMMs populated with CMOS 1Mb DRAMs in 1988-90. (This departure episode is very well covered in Intel Founder Andy Grove's book, "Only the Paranoid Survive".

Soon, CMOS took over at the 1Mb generation with the rise of Toshiba, but without Intel.

Case History #3: Inmos Swept up by STM, vanishes from DRAMs, 1987 Inmos was something of an Anglo-American Intel near-look-alike, and certainly carried forward their memory mantle. They had strong technical teams, good DRAM designers, top flight high-speed SRAMs, and their own 'Transputer', an 'MPU' that was way ahead of its time, but which failed to gain traction in the market. The memory design team was in Colorado Springs, where the memory products were centered; the MPUs were done in Bristol, Wales, UK, and had their own fab there.

Inmos was beaten up badly in the 1985 downturn, and a new Government in England forced the sale of the Government stake as a part of its privatization program. Thorn-EMI (yes, the record company) picked up the Inmos option, and became a significant IP holder of Inmos DRAM and SRAM patents. But the patents were about all that survived the turmoil of the late 1980s, the fabs were abandoned, the staff cut loose, and eventually Inmos was sold to, and rolled up into ST Microelectronics in 1989.

Case History #4: LG Shotgun Wedding, "Married" to Hyundai, 1999 Not unlike what we have seen in Taiwan in the past decade, and emphatically, today, "Everyone wants to make DRAMs", and when the Koreans entered the chip business in a big way in the early 1980s, they were no different. Samsung was the best-established from the start, and after licensing Micron's 64K DRAM, everything thereafter was home grown. Hyundai and Lucky-Goldstar (to be LG Semi) relied heavily on licensed designs and technologies until their own in-house capabilities were adequate.

All three companies rose up in the rankings during the 1988-89 shortage and benefited from the Japanese Stall (Lost Decade) as the industry rolled into the 1990s, gaining huge market share in DRAMs in 1992-95. But the market fall-off, and the ensuing Asian Financial crisis in Korea in the late 1990s sent shudders throughout the Korean electronics business. Many industries were 'rationalized' at the point of a gun by bankers and government. Samsung gave up its nascent automobile business. A government-forced marriage combined LG Semi and Hyundai into a single DRAM company in 1999, renamed it "Hynix", and combined all technical assets and once-disparate process roadmaps to merge 18 months hence (about 2000/01) in a common technology view and product set. LG Semi disappeared; even LG's creditors had a hard time finding someone in the new Hynix organization to pay them. It was rough running for a while, and not without casualties.

But the worst was yet to come: Barely settled down, Hynix emerged right on the cusp of the 2001-03 industry downcycle, which caused more than $25B in memory industry losses. The new Hynix was a mainstream DRAM maker, but also had some positions in SRAM and ROMs. But even by 2000, virtually all memory of LG Semi had vanished into the larger and more powerful Hyundai operation.

"We are not alone!:" LG Semi was not the only DRAM casualty in the 1998 downdraft. With 1998 industry-wide DRAM sales of $14B generating a loss of $12B, TI's new management decided to pack it in, selling their operation to Micron. And Japan, after a very tough decade of receding influence and declining market share, began their prolonged march to the exit with the spin out and establishment of Elpida from NEC's and Hitachi's DRAM operations, both market leaders a decade earlier in the "K" densities, about which we'll have more to say below.

Case history #5: TI sells DRAM operations to Micron, 1998: TI was one of the most important DRAM pioneers back in the 1970s, but like all US companies, steadily gave up ground to the Japanese, one DRAM generation after another, during the 1980s. The downturn of 1985 pressed them severely, and resulted in their 'Muscular IP Defense and Licensing Strategy', launched in 1986-87, and which netted them more than $2B is licensing fees and royalties over the next decade, as their seminal DRAM patents were licensed to all comers.

But the 1998 DRAM market blitz, again driving huge TI losses, and a sudden change in TI's Top Management brought on by the sudden death of TI Icon and CEO, Jerry Junkins a year or so earlier, led to their 'Creative DRAM Escape'. Even today, this exit sets the standard for a comprehensive and non-lethal departure from a fierce and unforgiving market. This allowed TI to leave behind major risky market positions and cash requirements, and bring the New TI organization to life in other areas that had been neglected for years, an altogether too common situation in DRAM companies...feed the beast to starve the babies

In 1998, TI sold its DRAM operation to Micron for cash and stock. This was the most successful exit, from a financial point-of-view, that the DRAM market has ever experienced. TI gave up its DRAM business, gave Micron a ten-year holiday on TI DRAM IP (just now expired in early 2009). TI unloaded its DRAM fabs in the US and Italy, its share of TECH JV in Singapore, and its DRAM JV with Kawasaki Steel (KTI) in Japan, plus an empty fab in the Dallas area. TI had just signed on Samsung and Toshiba for ten-year DRAM royalty bearing license deals in 1998, too, and though they gave up a substantial flow (~5% of sales) from Micron, it was much smaller an assured income loss than what they retained from other DRAM makers.

Micron assumed the debt to the Italian government that dated back to the establishment of the Avezzano fab, and gave to TI a boatload of Micron stock (which was trading at about $20 at the time) in payment for TI's fab assets. TI then loaned Micron enough money to get a head start on moving the fabs to the Micron 0.21um process node. Micron made huge gains in share price and DRAM market share in the next two years, and easily paid back the debt. TI proceeded to liquidate the shares over the next few years, as well, at prices sometimes as high as $90+ (Micron's 2000 peak share price was $94.50...it's now selling in the $4.50-$5.00 range.)

In doing this deal, TI kept a good cash flow from the extensive patent portfolio, actually had a capital gain on the sale of about $100M, cleared huge assets and liabilities off their balance sheet (turning them into cash and stock), washed their hands of a painful 25-year voyage in DRAMs...and caught the 'mobile phone wave' with their newly-invigorated DSP and analog circuits operations.

TI never would have risen to be the #3 or #4 chipmaker they were in 2008, with DRAMs in their product portfolio. Indeed, they would more likely have followed Mostek and Inmos into the grave or consolidation with a healthy and rich suitor.

Case History #6: Hitachi-NEC merge to make Elpida, 1999: After watching their DRAM market share peak in the late 1980s at about 70%+, Japanese DRAM makers then watched their prospects steadily wither away in the aftermath of their own stock bubble's bursting. After more than a decade of 'can't do anything wrong' and steady gains across most market segments, inside and outside chips and electronics, the Nikkei Stock Index dropped from more than 40,000 in 1989, to under 10,000 in the worldwide recession of 1990-91. Market share, investment and Japanese confidence followed in like retreats.

During the DRAM run-up in 1992-95, which it reached a still-record $40B in DRAMs, Japan's DRAM makers, with huge problems at home, excessive caution and facing the SIA's Trade Warriors at every step, gave up ground to Korea. When 1998 brought the then-worst-ever DRAM market and its losses about matching total sales, NEC and Hitachi decided to spin out and combine their respective DRAM units, which they did by the end of 1999.

The parents, properly, gave the kids a head start: DRAM technology rights, some legacy DRAM fabs and designs and capital to get running. Against a stiff headwind in 2001-2, they finally aligned their products in a uniform process, and launched their new E-300 fab. NEC and Hitachi held majority stakes in the new Elpida venture, but gradually liquidated large portions over time. Also in 2003, Mitsubishi, which had remained as a standalone Japanese DRAM maker, joined Elpida.

Recently, Elpida has been anointed the IP 'Provider of Choice' for Taiwan's Memory Company (TMC) initiative, but the fact that the venture remains stalled, and the company remains cash constrained, leaves Elpida in a still-precarious position. As Japan's sole DRAM maker today, they command some special attention with Japanese bankers and within the industry; how deep the pockets are of those 'interested parties' remains to be seen.

Case History #7, IBM Micro exits DRAMs, 1999: Despite having 'one of their boys'...Robert Dennard...invent the DRAM in 1968, from the very start at IBM there was always a 'business vs. technology' conflict about being in the DRAM business. IBM made DRAMs from the earliest DRAM days for their internal systems, which parts were not wholly compatible with Industry Standard (I.S.) DRAMs. This gave IBM systems some performance advantages, they thought, but made them price non-competitive in times when oversupply in the larger merchant market drove DRAM prices below manufacturing costs. Internal price benchmarking made the Natives look bad against 64Ks for 50 cents in 1985, and $6-8 1Mb in 1990-91.

IBM's coming out into the merchant market in 1993 leveled the playing field, and forced I.S. DRAM product designs that could be used both inside IBM, and sold to the merchant market. Some systems inside IBM still used proprietary DRAMs, custom specs or unique packages. But the Microelectronics Division now had the larger merchant market to sell to, so scale was no longer limited by internal DRAM demand.

This disagreement dipped below the surface during the initial DRAM upturn 1993-5, when even IBM's DRAM Gross Margins rose close to 70% in the waning days of 1995. But there were still currents of "Strategic Confusion of Purpose" that were happening concurrently all along. DRAMs were starved for prime development resources, they sold off a DRAM fab in Germany in 1994, they were shackled by confused cost accounting methodologies that obscured DRAMs' true profit contribution, and DRAMs were frequently used to fill fabs, and make up for the ASIC businesses' difficulty in launching its own Merchant ASIC business.

The mainstay of the DRAM operation was a Triad formed of IBM, Siemens and Toshiba in 1993, which used the trench cell to build a common 16M DRAM platform that reached production in about 1994. IBM played the lead, and engineers from all three companies led the charge at IBM's Essex Junction plant, and at the R&D and Development line in Fishkill NY. From this, each company also used its own production facilities to make and sell DRAMs to the merchant market.

IBM and Siemens shared a DRAM manufacturing operation in France, and in spring of 1996, a JV production facility was announced between IBM and Toshiba at Manassas VA. The timing could not have been worse for this venture, Dominion Semiconductor, which started DRAM production in about the middle of 1997.

The DRAM business was good until the last days of 1995, but prices dropped fast in early 1996. As the internal strategic battle went on within IBM, both the French JV and Old Dominion continued to cautiously ramp DRAMs. Neither was sufficient scale to get down the cost curve, IBM's reluctance to commit to DRAMs hindered their ability to shrink designs, build market share, or be anything but a 'Summer DRAM Maker'. Their share, which was about 7% in 1993-94, drifted steadily downward towards 2-3% as they equivocated. All systems were paused in 1998, and in 3Q99, they finally pulled the plug, wrote-off assets and withdrew from their Siemens and Toshiba JVs; France was shuttered and converted to logic, and the ownership of Dominion was transferred to Toshiba, which continued to make DRAMs. IBM's 1998 licensing of DRAM technology to Nanya was transferred to Siemens, which relationship continued until Qimonda's recent filing for Bankruptcy protection.

Overall, IBM took charges of more than $1.2B at the time, which, when added to operating losses for 1996-99, totaled more than $2.5B from the end of 1995.

Case history #8: Toshiba sells DRAM operation to Micron, 2002: Toshiba was the DRAM market leader at the 1Mb generation in the late 1980s/early 1990s. Like most Japanese DRAM makers, they had mixed-to-down results during the 1990s, hampered by cash shortages and a 'failure of confidence' in the early days of the "Lost Decade". In 1993, after falling somewhat behind the pack (notably Samsung), they formed a DRAM development venture, Triad, with IBM and Siemens with most development centered at IBM's facilities in the US. In 1996, they formed Dominion Semiconductor, a DRAM production JV with IBM located at an IBM site in Manassas VA; this was their only chip production facility outside Japan since an early 1980s Silicon Valley acquisition.

They continued to run DRAMs in their Japanese fabs and at Dominion through the 2000 upturn, and the business was good. Process technology was advanced from 0.25um to 0.21um and touched 0.18um in production in all facilities. But their extremely strong position in the emergent NAND flash market, and their persistent difficulty keeping their Dominion fab current and up to scale, led them to seek out Micron as a buyer in late 2001 (downturn). This deal was closed in early 2002; the NAND flash wafers and production, which has begun to displace DRAMs in Dominion, were moved to Flash Alliance fabs (Toshiba and SanDisk JV) in Japan. DRAMs, which used the IBM-Toshiba trench DRAM, were at first continued at Dominion for about a year, but then halted and the facility was idled while it was retooled for 300mm wafers, the Micron process flow, and a more advanced technology node. Toshiba's production of NAND continued its ramp in Japan, profitably, and Toshiba escaped relatively free after some modest write downs and one-time charges.

Micron, on the other hand, spent some considerable time and money getting the Dominion fab into production in their own technology. Of Micron's acquisitions of TI, Qimonda-Inotera and Toshiba DRAM operations, this was likely the least successful, in terms of high cost and difficulty aligning the process flow with their own product.

Toshiba, once rid of a 'foreign fab' and their DRAM business, went on, with their technology and fab partner SanDisk, to NAND greatness. Of course, once Hynix and IMFT joined the NAND hunt and ramped up production, along with Toshiba's own overly ambitious investment plans, glutted the NAND market in 2008 and made it...as bad as DRAMs!.

The week kicks off on Monday, June 22, when Denali hosts a full day of training and tutorials on their Databahn DDR3 contoller, their Industry Leading MMAV memory modeling and verification products, High speed NAND design, PCIe verification, USB 3.0 solutions and their DDR-Phy. Registration for these tutorials can be accessed from the Denali MemCon website (agenda, registration).

Two days of presentations follow on June 23-24, with industry experts sharing their visions of the future under the Conference headline, "Beacons of Innovation". The Denali-led portion, June 23-24, is free to all registrants, and features 19 presentations and four informative panel sessions discussing SSDs, NAND Flash, and LP and DDR3 DRAMs.

Also, running as a part of Denali MemCon 2009 on Wednesday, June 24, is Web-Feet Research's Annual NVM Conference, this year's focus on Storage Class Memories, which is a one-day paid event with seven sessions featuring a variety of flash memory and storage speakers, and associated panels. Web-feet's homepage can be found here, but both the agenda and registration details can also be found at the Denali MemCon link above.

Afterwards, on Thursday, June 25, the semiconductor industry Standards' Committee, JEDEC, will hold a technical workshop on their newly-released "LP DDR2 DRAM" spec, which will lead the next DRAM wave into mobile phones, PDAs, and other mobile systems. This workshop will also be in the same Hyatt Regency location, and the agenda can be found here. To register for this tutorial, go here, or through the Denali MemCon website.

Today's Memory Market is the Toughest Ever.

The memory marketplace has always invited participation and competition since the earliest beginnings of the semiconductor business in the 1970s. Needless to say, with the memory supplier base becoming more and more concentrated, most of those once-participants no longer make memories today. Caught in the vice of below-cost selling, and an extended period of unprecendented financial losses, we are in the midst of yet another consolidation and elimination phase. Nowhere is this more evident than in the DRAM business. This industry has a ferocious ability to reduce prices and costs (to benefit consumers and expand markets), but it also has Dr. Jekyll's ability to cause immense financial damage to the greatest of companies, lay waste to reputations, mobilize the resources of huge conglomerates, banks and governments, and surface some of the most confused and spurious arguments for getting into the business, making DRAMs, investing billions of dollars in DRAM fabs and technology, continuing to make DRAMs (and invest in making them) when there is no possibility of ever making a profit, and instead taking suicidal actions to stay in a money-losing business, when those actions only propagate the damage to anyone who remains standing.

Historically, getting into the DRAM business has proven far easier than getting out of it.

Here are several observations and lessons that we have gleaned from some of the most significant DRAM departures than have happened over the past decades, which some may use to their benefit today, or ignore at their own peril. Almost all the DRAM pioneers from the 1970s, 1980s and mid-1990s are gone: Intel, TI, Mostek, Japan, Inc. (though NEC, Hitachi and Mitsubishi live on in a downsized Elpida...all other Japanese are gone to the showers), Siemens-Qimonda and Inmos. Except for Micron and Elpida (not insignificant exclusions, to be sure, but not at the center of the Main Event), the battle today is best summarized as 'Taiwan' v 'Korea', each with cash-rich stakeholders who need to decide how long to play, how much they're willing to pay, and why they are making DRAMs; these are the same questions that float up in every brutal DRAM cycle since the 1970s.

DRAM makers have lost about $20B over the last 18 months. Understandably, their share prices are a fraction of what they were in 4Q07, forcing similar losses on their shareholders. Despite the infusion of about $12B over that time period in CapEx and R&D, it has only served to make their prior investments worth less (or nothing), and put their huge long-term debt truly at risk. This is not the first time this has happened, but it is the most severe "DRAM recession", with the highest stakes and the most confounding problems facing DRAM makers, their parent companies and shareholders, their bankers and their Governments. Their huge CapEx investments in 2006 and 2007 have run smack dab into a financial-crisis led weakening of semiconductor demand today. One could easily imagine another $20B in losses in the next 18 months, if the industry does not find its way out. It is not an intractable problem, but one that requires decisiveness and some considerable attention to how the diverse stakeholders are all impacted, rewarded and punished. DRAM suppliers have only to decide what they want to do, and avail themselves of the options at their disposal.

Backgrounder on Memory Crises: Although most memory products are highly commoditized, few have caused such financial (and emotional) damage as DRAMs in terms of placing their practitioners in such compromised financial conditions that an exit was a life-and-death choice for the parent or shareholders. Not only is the DRAM business the largest, in terms of revenues, but the large center of the product line...roughly 70-80%...offers almost no room for safe haven, product differentiation, enforced 'customer (or vendor) loyalty', or margin protection. Until the advent of NAND Flash, DRAMs were the perpetual leading-edge lithography and technology memory product standard-bearer, pushing the envelope on the litho front ahead of all other products. For a long time, it was said, "You need DRAMs to drive the process technology", and that was that (at least until Intel said 'Logic drives processing", in the early 1990s.)

In this discussion below, we have described and analyzed the departures from the DRAM business for eight companies that have occurred over the past three decades. Each of these departures was driven by severe market downturns, though other factors were also at work. Each of these companies were at the time, or formerly, major forces in the DRAM business. Along the way, there were also many smaller 'DRAM Dabblers' who were squeezed out of the business, for lack of scale, skill, or a clear purpose. Here are some of the generalities that we have distilled from these DRAM Demise Experiences:

Lesson #1: Companies, owners, stakeholders and agents must have a clear understanding of why they are in the DRAM business. Real reasons, or make-believe reasons? Today's reasons, or yesterday's reasons? Are there alternative ways of accomplishing the same goal? How can we exit with minimum damage to the larger coroporation and shareholders? Can we ask and answer, 'Why are we making DRAMs?'

Lesson #2: Departing earlier is better than later. Indecision has a price, though the prospects of huge losses are sometimes needed to focus the mind on the other issues below. IBM Micro's losses ranged over 3+ years, and doubled their exit cost. Ditto for Infineon/Qimonda...and maybe ditto for Taiwan DRAM Companis...TBD

Lesson #3: DRAMs are different, in size, scale and market dynamics, not to mention a host of 'emotional factors' not based on any reality or consensus. Outsiders making DRAMs have not fared well..but even insiders have been treated to a new twist in every market cycle.

Lesson #4: Good decision-making demands a clear line of authority for making the decisions. Dictators are more decisive than Boards of Directors or shareholders, or than a herd of cats, with their different interests, though there's always need for discussion and 'consensus building'. Institutionally, companies need a clean process for responding to business crises...rights and responsibilities, pecking order for payback.

Lesson #5: Even in a down DRAM market, an enlightened exit strategy can participate in some comeback of the surviving entity; a comprehensive exit strategy can clear up a lot of problem areas and complications in the seller's existing operation, and set them on a new path. TI's sales of its DRAM business to Micron exemplified how a comprehensive package could make an exit so beneficial to the seller. By taking Micron stock back for part of the sale price was smart, since the MU shares were cheap when they took them in 1998, but cyclically rose up 10x to a record high of $94.50 in 6/00.

Lesson #6: No companies ever regreted getting out of DRAMs, even though almost always the decision was painful for all those who abandoned the business. At worst, you take a substantial loss and move on. At best, resources are freed up to 'Chase your Glory', and undo the heavy burden of DRAMs and give life to MPUs (Intel and IBM), DSPs and Analog (TI), Foundry Partner to TSMC (Vanguard)...all the other ambitions that are starved by the disproportionate burden placed on the institution by DRAMs (Japan, Inc., Infineon).

Lesson #7: The threat of government involvement, or a rich parent who is willing to fund its child's DRAM addiction indefinitely, is a good reason to be extra cautious or stay away...The market stakes are too high already, and below-cost sales too frequent, but made worse by your competitor having a large reservoir of cash at their command...cash that is often not subject to 'market rational' behavior

SRAMs, EPROMs and NOR flash (to name only the larger memory markets), as well as Mask ROMs, EEPROMs have had their own dynamic, and some of these eventual departures were painful...but nothing like what the industry has witnessed in DRAMs over the past decades. It has always proven to be easier to reduce one's position in smaller markets, especially ones that a clear transition was visible ahead...one could just stop investing and let the system wind down over a few years (Like SRAMs, moving onto the Logic/MPU/SoC beginning in the mid-1990s.) Mask ROMs was mostly this way, too, once flash gained inertia and price competitiveness, and was seen as superior and almost as cheap except in the highest volume applications, or could not keep up in density with CDs for games. Also, there are markets with only 2-3 strong players could intimidate others to leave (EPROMs: Intel and Hitachi/Japan). Departures from most of those markets were not financially painful or especially noteworthy; things just wound down (exceptions and examples of precipitous departures do exist).

This essay is merely a prolog; in another article to be placed on the Denali website on Tuesday next week (26 May 2009), we will describe the following DRAM market exits, with background and discussion about the companies' 'pre-exit' situations, and the genesis of their departure:

Case History #1: UTC sells Mostek Division to Thomson...to SGS Ates, which shuts it down, 1985+
Case History #2: Intel's Painful Exit, 1985
Case history #3: Inmos Swept up by STM, vanishes from DRAMs, 1987
Case History #4: LG Semi's Shotgun Wedding, "Married" to Hyundai/Hynix, 1999
Case History #5: TI Sells DRAM Operations to Micron, 1998
Case History #6: Hitachi-NEC merge to form Elpida (1999); Mitsubishi joins later
Case History #7: IBM Micro Exits DRAMs, record 'loss per unit of market share', 1999
Case history #8: Toshiba Sells DRAM operation, fabs to Micron, 2002