DRAM has been the workhorse of the semiconductor memory
market for a long, long time. At the MemCon 2010 conference, I
learned that the basic DRAM cell design goes back to 1970 and has changed
little since. It may seem somewhat heretical to ask, "What Will Replace DRAM?"
but a lively panel at MemCon did just that.
Prior to the panel, as I noted in my previous MemCon 2010 posting,
analysts noted that NAND revenues are growing faster than DRAM revenues. This
is an overall reflection of where the electronics industry is going, given that
NAND flash is popular in mobile applications and DRAM is the traditional memory
for PCs and servers. Analysts also noted that current memory technologies are
falling short of system performance requirements.
So do we really need a DRAM replacement? "Absolutely," said
Bob Merritt, founding partner at Convergent Semiconductor, who noted that the
traditional DRAM cost/performance curve cannot be continued at small
lithographies. He noted that target applications are shifting from data
processing to multi-media, from desktops to mobile devices, and from limited
numbers of similar applications to large numbers of diverse applications with stringent
power and performance demands. Memory subsystems need to be concerned about
heat, power, and mobility, and they need to be non-volatile, he said.
Just as I was thinking that this focus on applications
sounds like EDA360, Merritt suddenly recommended that "everybody download and
read" the EDA360 vision paper so as
to "look at where the industry is going."
Merritt's solution for DRAM? "We don't need to replace the
DRAM cell structure. We need to replace the DRAM business model. The idea of a
commodity product that fits into all applications forces suppliers into a
Barry Hoberman, CMO at Crocus Technology, gave a strong
pitch for Magnetoresistive RAM (MRAM)
as a technology that will have "full cost competitiveness with DRAM" at the 32
nm process node. Right now, he said, it's an effective replacement for embedded
flash and is cost-competitive with embedded SRAM. But it's still a niche
technology, he observed.
Ed Doller, chief memory system architect at Micron
Technology, said his company has visibility into the next three DRAM process
generations - but he said a DRAM replacement is definitely needed by 2015 or so
and that work needs to start now. But don't hold your breath for non-volatile
DRAM. "The industry is trying to do that, but people are also trying to create
cold fusion," he said.
Doller quickly reviewed a number of emerging technologies - FeRAM, MRAM, FBRAM
(floating body RAM), T-RAM,
and PCM - and concluded
(with some dissent from Hoberman) that none offer a combination of scalability,
production availability, and drop-in compatibility with DRAM. Phase-change
memory (PCM) is perhaps the most promising, but lacks plug-in compatibility,
meaning that "we would have to look at the architecture of every computing
system known to man and figure out what to do with PCM." (An ongoing debate
about PCM is taking place in dueling articles at EE Times).
Mark Greenberg, formerly of Denali
and now product marketing manager at Cadence, gave a spirited and entertaining defense
of DRAM. He said that the same, simple cell structure (based on 19th
century technology) has been in use since 1970, volatility is "not an issue" when
there is battery storage, and DRAM is cheap. Thus, he said, "fear of the demise
of DRAM is overblown."
The panel didn't reach a clear consensus on what comes after
DRAM. Nor did they conclude that new technology will change the dynamics of the
commodity memory industry with its recurring boom/bust cycles. But my sense is
that the memory market is being impacted by the same application-driven world that
is changing EDA. Memory is no longer a "throw it in later and don't think about
it" commodity; it's a vital part of the system architecture from the beginning.
And this is why Cadence is now in the memory IP business, and is now the
sponsor of MemCon.
MemCon Question #1: Are Memory Boom-and-Bust Cycles
MemCon 2010 presentations are available on line.