Future memories not so future any more
There are not many memory technologies with a 50-year history. Phase-change memory (PCRAM) is particularly unusual in that it has only become a volume player in the past few years. Even then, manufacturers Intel and Micro are coy about the nature of their Optane memory.
Intel’s Gordon Moore experimented with PCRAM at the beginning of the 1970s, working on the ‘ovonic’ chalcogenide materials promoted by the late inventor Stanford Ovshinsky. Magnetic memory (MRAM) arguably has almost as long a history if you include mainframe core storage. Compared to PCRAM though, it’s a whippersnapper: appearing just a few decades ago.
Today, things look to be changing fast for these long-term experimental memories. The rise of the normally-off processor in IoT SoCs has focused attention on how much use memory gets while the system is operating. But it is not just those that can benefit from fast non-volatile memories likes MRAM. At the VLSI Technology Symposium, Toshiba’s Shinobu Fujita pointed to work by his team. They developed a platform to measure real-time CPU and last-level cache usage in applications processors. They are surprisingly inactive, with long gaps in accesses. By moving to non-volatile memories tuned for normally off behavior, experiments showed a two-thirds energy saving.
MRAM demonstrator
At DAC earlier in June, Arm senior solution marketing manager Phil Morris was demonstrating on Samsung’s booth an IoT SoC, designed together with Cadence Design Systems and Sondrel, that uses MRAM as a way of saving energy. As well as integrating MRAM, the SoC takes advantage of the 28nm FD-SOI process offered by Samsung to implement forward body bias techniques to support higher clock speeds when needed. Most of the 4x4mm die is memory, though a significant fraction of that is SRAM used to help demonstrate the non-volatility and power savings offered by the on-chip MRAM array.
The MRAM on the Musca S1 is able to run at 25MHz and is smaller in terms of die area than the SRAM array it is designed to replace. Although manufacturability has been a problem for MRAM, Arm sees the situation changing rapidly. Morris says the company expects to provide evaluation kits based on the Musca S1 later this year. “For wearables, it’s a no-brainer,” Morris claims.
That is likely to be followed by Arm providing rapid adoption kits in conjunction with Cadence to support teams who want to implement their own SoC concepts that employ MRAM.
Manufacturing moves
As a company providing equipment to the chipmaking business, Applied Materials also sees a shift from experimental devices to commercial bulk memories and SoCs. Over the summer, the company launched two production tools designed to automate much of the job of putting MRAM and other novel memory technologies onto silicon. At the same time, Applied is working with startups such as Crossbar and Spin Memory to help implement their IP in captive fabs and foundries.
Kevin Moraes, vice president for materials deposition products at Applied, says MRAM – once made manufacturable – has distinct advantages over flash for SoC makers. “Flash processing needs high temperature and separate mask steps. MRAM can be built directly on top of the logic,” he claims. “And with a penalty of just two or three mask layers.”
Although it does not require much in the way of masks, MRAM does need a lot of steps. “There can be thirty-plus layers. Some are very thin. Tiny differences in thickness can change the properties of the memory. So there have been some serious manufacturing challenges that needed to be overcome.”
The Clover tool that Applied is offering has multiple chambers fed by a robotic system that take care of the layer deposition. “It includes onboard metrology for the first time in the industry: it monitors the films as they are manufactured,” Moraes claims.
Many experiments used atomic layer deposition (ALD) because of its ability to lay down extremely thin films. But Applied has turned to a form of the more conventional physical vapor deposition (PVD) process. The problem with ALD is that it places restrictions on how layers are formed. “You get discrete steps,” Moraes says, adding that PVD provides greater levels of control. “With PVD, we’ve been able to reduce deposition rates to sub-angstrom per second.”
Although the wafers are shuffled around the tool for different steps, Applied designed the Clover to allow multiple layers to be put down in one chamber. “There are different cutouts for different metals: up to five in a single chamber. You have to rotate the wafer. But as long as you rotate the wafer and have the correct angle for deposition you end up with sub-angstrom uniformity.”
Endurance control
Temperature control is another key addition. As well as rapid thermal annealing, a number of the steps take place at cryogenic temperatures to prevent the films from mixing in unwanted ways as the atoms adsorb onto the surface of the wafer. “Cryogenic cooling is used in implant already. So we borrowed that from our sister division. And we have one of the leading metrology providers and added from there as well.”
A further change lies in the sputtering materials, which are ceramic based. “This is a lot more challenging. But you get improvements in signal-to-noise and a hundred-fold improvement in endurance,” Moraes claims. Endurance has been a critical stumbling for MRAM until recent years.
Applied is tackling resistive memory (ReRAM) and PCRAM using another tool: a version of the Endura PVD system. Although they do not rely on extremely thin layers, composition is critical. “The layers that they have are very complex and rely on small amounts of impurities. Any variation in these dopants affects the performance of the memory,” Moraes notes.
“Some of these phase-change materials involve toxic materials and they change on exposure to air. So you want to process them fully integrated and perform metrology in vacuum.”
So far, Applied has shipped tools to five customers for MRAM production and eight for ReRAM and PCRAM. Moraes sees a mix of customers and an expansion from the days when memory was the purview of bulk suppliers. “Standalone memory will still be driven by the high-volume players. But embedded memories will be done by foundries and logic manufacturers: a lot of value can be provided by embedding memory on top of logic. Each foundry will probably have their own types of memory that they will offer, with diversity in speed, power, and cost,” Moraes explains.
This is where the IP providers will play a role. “Crossbar, for example, have their own process technology. They will license their IP to a foundry partner and recommend our equipment to the foundry partner.
“We think it’s the right time to bring this capability to market,” Moraes concludes.
