3D memories 'the way forward'
Whatever non-volatile memory makes it into the next decade, it will have to be a technology that can be stacked in layers, according to the former head of Intel’s memory research unit.
Now heading up a memory startup called BeingAMC, Stefan Lai told delegates at this week’s International Electron Device Meeting (IEDM) that continued scaling of flash memory in two dimensions is in jeopardy: the technology is simply running out of electrons. “My opinion is that multilayer memory is the solution,” he said.
Although Lai said the search for a replacement for flash remains opens, his vote remains for the technology he championed at Intel: phase-change memory. “This is a technology that can have both RAM applications and storage applications,” he claimed, but he emphasised the need for any future technology to be composed of crosspoint storage elements that can be layered on top of each other. “You get away from the charge-storage transistor [of flash] altogether.”
Resistive RAM, which has appeared recently, is a possible contender, said Lai, but lacks the reliability demonstrated so far by experimental phase-change memories and even the way it works is still a matter of debate among researchers.
Lai admitted that the right material for commercial crosspoint memories based on technologies such as phase-change remains elusive. “The challenge is that you rely exclusively on material properties. If you don’t get the materials right, you don’t stand a chance. New material discovery techniques are required. And the material control techniques are not yet known. So any crosspoint memory will have a long learning curve,” he claimed.
But, the advantage of moving the memory cells into the interconnect layers above a chip will make it easier for companies to enter the market. “You have simple manufacturing: it involves only back-end steps. And the steps can be repeated, repeated, repeated,” Lai explained.
By having the non-volatile elements in the metal layers, Lai proposed a possible unified memory for portable computers. One of the problems with phase-change memory is it comparatively slow write time, although it should perform better than flash.
“The write problem can be solved by using a DRAM buffer. In most applications, you are reading from memory a lot more often than you are writing to it. You probably need 1GB of DRAM for every 10GB of phase-change memory as a buffer,” Lai explained. It should be possible, he added, to build a DRAM using conventional processes and then move to another back-end plant where the non-volatile memory layers are added on top, so that the device contains its own DRAM buffer.
“With just two chips, I have all the memory I need for a notebook computer. It is not quite a universal memory but it is very close. With one kind of memory, you can meet the challenges of building a notebook computer,” claimed Lai. “The one who is successful in this will create new business opportunities.”