For some niche applications, the new memory technologies are already becoming established. Over the next 12 to 18 months there will be some very different types of devices emerging with embedded non-volatile storage that will provide the designer with very interesting options. Nick Flaherty reports.

The world of semiconductor memories is about to get turned upside down. After long years in the wilderness, new memory technologies are starting to become commercially realistic with the latest manufacturing processes to offer the promise of fast, non-volatile storage.
Part of this comes from recent developments in technology in other areas such as hard disk drives and high density DVD video disks. But part of the move to the new technologies is that the existing non-volatile memory - flash EPROM - is set to face some major challenges over the next few years.
Flash has been the mainstay of silicon-based non-volatile storage for the past 15 years, used in most applications from industrial to mobile phones. The problem is that as silicon technology advance, the problem of programming and reliability of flash increases.
Flash is based on electrons tunnelling through a barrier, and as voltages get lower, the programming time gets longer, and it is already relatively high at several milliseconds. Reducing the thickness of the barrier helps, but that increases problems with breakdown that can reduce reliability. All this is coupled with the fact that flash never had a very good lifetime anyway - typically around 100000 read and write cycles.
So new, truly non volatile technologies are coming through with write times equivalent to today's volatile memories such as SRAM, and lifetime of a million times more cycles.
"In the last 20 years we have seen only pure evolution of memory technology from EPROM to flash which are all based on the same technology," said Phillipe Berger, product marketing manager for memories at ST Microelectronics. "The new technologies are all revolutionary, using new silicon techniques, so it will take years before they have real commercial value."
With the latest silicon manufacturing processes, the memories - the magnetic RAM (MRAM), the ferroelectric RAM (FRAM) and the ovonic memory - are starting to have enough capacity to be useful.
MRAM is the highest profile at the moment, with a 4Mbit prototype shown in December (2003) based around a magnetic tunnel junction (MTJ) where magnetic material stores the data bit.
"We are very excited about this," said Saied Tehrani, director of MPEM technology at Motorola Semiconductor division. "It really brings two parts of the industry together - semiconductor and magnetics. We have taken the thin film technology and integrated that with the silicon transistor technology and used the magnetic polarisation to store the data and the silicon transistor for reading and writing the information."
"A lot of the expertise comes form the rapid developments in the hard disk drive and the technology that we used to store the bits is very close to what is used in the magnetic head of a disk drive," he said. "Obviously there has been some changes to make it applicable but a lot of the knowledge is common.
It is only in the last few years that this has become possible. "The figure of merit we use is the change in the resistance and that has improved by more than an order of magnitude since 1995," said Tehrani. "That increase is translated into faster access and means it can now compete with semiconductor memories. That means we are getting the best of both worlds - the speed of semiconductor and the reliability of magnetics."
The 4Mbit test part has a read write time of 25ns, significantly better than flash. The MRAM cell also hasn't yet seen a failure mode in the programming cycle, and that means it can have 10E15 or 10E16 cycles. The parts have already been cycled 10E12 times and haven't seen any changes. That is backed up by the long experience with the magnetic materials properties from the disk drive business. "History shows that there is no fundamental fatigue mechanism for the material," said Tehrani.
However this is not the whole story as there are other potential failure mechanisms, such as the breakdown of the magnetic thin film as the electric field across it can be high. "With the magnetic tunnel junction we have worked really hard to identify a number of failure modes for that and that's applicable on read states with electrical fields and temperature and we can get a good prediction of lifetimes. We are looking at parts pre million failures, where one 1bit fails in 4Mbits, and the mean time between failure for that is over 10 years."
The 4Mbit part is built in 0.18µm technology but for commercial products Motorola is looking at jumping a process generation and going straight to the current leading edge 90nm. This is being developed at a plant in Crolles, France, in a joint development with Franco-Italian chip maker ST Microelectronics and Dutch electronics giant Philips.
This would allow 64Mbits or even 128Mbit stand alone devices to be built, but that is not the aim. "Motorola is not in the stand alone memory business memory business any more, that's not the focus for us," said Tehrani. "There are compromises that you have to make for embedded memory over stand alone. Making standalone memories you reduce the size of the cell, increasing the complexity of the process to do so. For embedded memory you can't have a complex process."
So Motorola is planning to embed MRAM into devices alongside other functions. Tehrani would not comment on what these devices would be, except to say that they would be out on the market in 2005 and Motorola would announce the roadmap later this year. "Microcontrollers are a definite possibility for this technology," he said.
IBM and Infineon have also been working on MRAM technology, although they have only shown a 128Kbit part so far. But they have transferred the technology to a French-based joint venture, Altis Semiconductor, to develop commercial memories.
However another non-volatile technology - ferroelectric RAM - has been around for as long as MRAM and is already commercially viable.
"From a high level MRAM and FRAM are quite similar," said Mike Alwais, VP FRAM products at Ramtron, the pioneering company that has been developing the technology for the last 15 years. "They are both an attempt to create a true non volatile random access memory. But the underlying technologies are totally different. We use charge, which has no magnetic properties at all."
FRAM uses PZT - an alloy of lead, zirconium and titanium - that acts like a capacitor to store the data. "The cell design with one transistor and one capacitor is quite traditional, making it like a non-volatile DRAM," said Alwais. "That's probably why we are ahead - we are more readily adapted into a mass production environment. With our nine licensees we have shipped about 50m parts, and most of our licensees are still in the development phase. Fujitsu is our principal foundry and shipping 8bit and 16bit microcontrollers with embedded FRAM."
Samsung, now the world's largest memory chip maker, is also a licensee and has talked about doing stand alone memories, but development has been slow.
"During the downturn there has been limited investment and it's harder than it looks," said Alwais. "There has been a lot of emphasis on the 0.35µm and 0.25µm process nodes but they are too late. If they can get to 0.13µm quickly they will have a chance," he said.
US chip giant Texas Instruments has also licensed the Ramtron technology and like Motorola is looking to jump to the 90nm process technology to embed the memories in devices such as digital signal processors.
It did a comprehensive evaluation of next generation memories several years ago and came to the conclusion that FRAM was the best choice as it is easily scaleable to 90nm.
"We are still shooting for 2005 to have FRAM in a product on the market," said a TI spokesman. "There has been progress in MRAM, for sure, but we still believe FRAM is the right choice for our market space. In terms of bringing embedded MRAM to market in high volume products customers will buy, MRAM still faces significant hurdles."
Another chip giant, ST Microelectronics, is working on all three technologies, and after the same kind of analysis chose a third option - ovonic memory - as the way forward, licensing technology from a US start-up Ovonyx.
"We are investigating the three emerging technologies of ovonic, MRAM and FRAM," said Giulio Casagrande, R&D Director at ST's Memory Products Group. "As the memory product group we have done an assessment of what we can expect from the different technologies and we have chosen ovonic because it is the only one which can produce a cell size that is competitive with flash and has the ability to scale better than flash," he said.
This ability to scale is very important because the emerging technologies take a significantly long time to mature and it is important to look at the scaling and the cost competitiveness in the future, he says, and the only one that has this is ovonic. "We started working on this two and a half years ago and recently extended the scope of the license we have with Ovonyx to include embedded applications," he added.
The expanded license agreement permits ST to use the ovonic technology in both discrete devices for flash replacement and embedded applications with its microcontrollers, logic, and other products.
"We are very pleased that ST has exercised its option to expand the scope of its license for use of OUM technology, and we look forward to continuing to work with them in our joint development program," said Tyler Lowrey, president and CEO of Ovonyx.
"The market for Flash and embedded nonvolatile semiconductor memory is growing rapidly, and ST is an excellent partner to work with towards commercialising ovonic thin-film memory technology in a wide range of integrated products," he said.
The ovonic cell is based on chalcogenide, an alloy of antimony, telluride and germanium, which changes phase between the crystalline conducting state and the amorphous, non-conducting state. The storage element is made of a small area of chalcogenide which changes phase and that brings a change in resistance of over two orders of magnitude. This 100x range gives it a fast write time of 100ns and the first cells have demonstrated a cycle lifetime of 10E11 to 10E12 cycles.
The Ovonynx technology is already being used in DVD disks by the major disk makers and the basic technology has been demonstrated to be extremely long life.
The integration of the material has also been demonstrated in a full commercial process, which is important. "The material is relatively friendly, much more so than the MRAM and FRAM materials and this is something that we like in manufacturing," said Casagrande. "We are now developing megabit size devices to investigate the reliability aspects before we reach the industrial stage of mass production."
ST is following two options: for high density stand alone memories it is using bipolar selection transistors integrated under the memory element for the most compact cell. This is a fully selectable array that allows full parallel access and full granularity so that an individual bit can be changed. That is a key advantage over flash where a whole page of data has to be erased to change one bit.
For embedded applications, ST is using standard CMOS transistors as the selector. This gives a larger cell and is only viable for smaller size memory but it is a stripped down version of the process and better for embedded memory.
"We are not looking at displacing flash but looking at the future, flash will have problems with scaling and we want to be ready to replace it with a better performance technology," said Casagrande. "We have to develop something that will compete with flash which is running very fast and is very well supported. That's why we are extremely cautious and concerned about scaling."
But ovonic is not the only option that ST is following. "We are also developing MRAM with Motorola and Philips for embedded applications which goes very well with our assessment that MRAM is less viable in size and scaling to compete with flash," said Casagrande. "The motivation behind the MRAM development is different and it's a good relationship so we have a solid footing in MRAM."
ST also has a joint development with the IMEC research institute in Belgium on FRAM. "Our assessment is that it's the most mature technology," said Casagrande. "But we also have to say that while it is getting mature, it is still low volume and low density and we also understand the possibility that it is not realistic to compete with flash because of the cell size and the difficulty with scaling.
Instead FRAM can be considered a good competitor in niche applications, mainly smart cards and RF tags, he says. "We have investigated the technology and developed a module at 0.18µm but it is not obvious that the return on investment is good enough to justify bringing the technology in-house and move it to volume production."
So the battle will continue. Flash is not about to give up and die, and there are various approaches such as storing multiple bits in each cell that extend the lifetime of the flash. It is well established, well known and cost effective, and will take a lot to shake out.