Memory looks to a smarter future

Memory is practically everything in computing. Even using a slightly more conservative method for the calculation compared to that of market-analyst Gartner, chipmaker Micron Technology estimated more than 90 per cent of the transistors put onto chips in 2015 were for storing data rather than processing. Thanks to the onward march of solid-state disk drives, phones and tablets, the proportion is even higher today.

At the Design Automation Conference in Las Vegas earlier this summer, Micron CTO Gurtej Sandhu said technologies such as artificial intelligence will only increase the demand for memory. “Autonomous driving is more real than you think. The systems are getting better really fast and the amount of memory used really impacts the majority of the models.“

The drive to consume more data and with it higher memory consumption comes with a cost. Sandhu says: “One thing we do need to consider is at the system level feeding data into the processor from the system memory. Physically, it’s a bottleneck. We have researched into photonics, any technology you can think of to improve the situation. None of them are practical.”

If you can't get data from memory into processors fast enough, why not get the memory to take on more of the burden? This is leading to a shift in the way that memories are made. As a long-term manufacturer of bulk memories, Micron sees a shift in the way they are made right up to the point where processing elements move into the arrays of memory cells.

The first step is to combine processing with novel types of memory designed to streamline the tasks they perform and to tune logic and memory so that they work more efficiently together. This is in contrast to the past where cost concerns have meant adapting logic to whatever memory technology happens to be the fastest and cheapest in bulk. This is even having an effect on the companies that make equipment sold to chipmakers like Micron.

Kevin Moraes, vice president of material deposition products at Applied Materials, says: “Now we do a lot of work in hardware and software co-optimisation as well as manufacturing.”

Over the summer, Applied launched two machines intended to make novel forms of memory that lend themselves not just to being integrated onto processors, but which support unusual methods for processing data. 

“These new memories will become the basis for in-memory computing,” Moraes claims.

One of the target memories is resistive memory (ReRAM). This uses chemical changes caused by the way in which current flows through them to change state. As the cell is programmed, its resistance rises and falls. Sense amplifiers detect that resistance when the contents need to be read out. A similar technology is phase-change memory (PCRAM), similar to that found in the Intel/Micron Optane chips.

Moraes says: “A lot of in-memory compute will be based on ReRAM and PCRAM.”

Magnetic memory (MRAM) is another contender for future integrated memories, this is more likely to be used for low-power devices where the aim is to preserve data while the host system is powered down. Although it is possible to stretch MRAM beyond this, it is seen as a primarily digital memory technology. A lot of the concepts for in-memory processing rely on ReRAM and PCRAM’s ability to store intermediate values and so work as an analogue computer. 

It does not mean that conventional memory technologies are out of the running. Flash has a similar ability to store analogue values and is employed by startup Mythic in its proposal for in-memory computing chips for machine learning. Memory designer Surecore has a research project looking at how it can use its knowledge of SRAM to support in-memory computation. “We’re looking at dot-product type computations in memory using analogue storage,” says Paul Wells, Surecore’s CEO.

Applied and others believe ReRAM and PCRAM will win out in the longer term because it takes less energy to reprogram them with different values when switching algorithms compared to flash and they have the benefit of remembering their contents when power is removed, unlike SRAM or DRAM. 

To try to get these novel memories into production – and sell more of its gear – Applied is taking a different tack on how it makes and supports equipment. These novel memories have taken decades to get into production because it is so difficult to maintain high yields. That situation may now be changing. Over the summer, the company launched highly integrated tools made up of multiple processing chambers around which wafers are shuffled automatically.

This design should make it easier to deposit the finely tuned layers needed to fabricate working ReRAM and MRAM devices as either bulk memories or on top of SoCs. The company is also recruiting IP suppliers such as Crossbar and Spin Memory to help get the equipment into the foundries who make chips for fabless companies.

“Crossbar have their own process technology. They will licence their IP to a foundry partner and recommend our equipment to the foundry partner,” says Moraes.

One further stumbling block for new forms of memory remains on the computing side. Micron has slightly painful experience of introducing a novel computational memory that failed to gain traction. Sandhu says the novel approaches to computation are “not optimal for all problems. The barrier is not the chips. We can build those. The challenge is to write the software and firmware. You have to start from scratch with that.”