Part 2: DRAM goes cryogenic
This entry was posted on Tuesday, July 18th, 2017.
In part one of this series, Rambus Chief Scientist Craig Hampel told Semiconductor Engineering’s Ed Sperling that cryogenic DRAM (below minus−180 °C or 93.15 kelvin) offers numerous power and performance advantages. These include increased transistor performance, the elimination of leakage, wires that super conduct, the ability of DRAM to operate as a non-volatile device and the potential to more easily design higher-capacity memory systems.
In part two of this series, Hampel describes what future cryogenic-based systems might look like.
“[Ultimately], we do think we’ll see cryogenic computing happening, meaning people [will] use superconducting technologies. One example is the technology known as Josephson Junction, [which] could show four or five orders of magnitude more efficiency than CMOS,” he explained.
“[Secondly], quantum machines are going to need a fairly conventional infrastructure to make them usable and programmable [and] we consider the memory sub-system to be part of that.”
Rambus’ approach, says the chief scientist, is to take fairly conventional DRAM architectures, designs and process and optimize them for cryogenic operation.
“Temperature is [currently] one of the reasons we don’t have the kind of reliability that we’d want and it’s an impediment to the reliability of the data center,” he stated. “By moving down the energy and thermal curves we immediately get better system reliability.”
With regards to architectural specifics, Hampel sketched out his vision of what a cryogenic memory configuration could look like in the context of a larger system, with the UI, network interfaces, and storage on top of the hierarchy at conventional temperatures.
“Below is DRAM at 77 Kelvin that will interface with computing elements – some CMOS and some other technologies that take advantage of superconducting interconnects,” he explained. “At less than 77 Kelvin, we will start to see CPUs, GPUs [and] TPUs, [along with] a number of conventional compute technologies. These also serve as the interface for quantum accelerators – so we think we’ll see quantum machines [below the above-mentioned technologies]. There are a number of companies doing quantum in the millikelvin range – and that’s kind of the way we see the computer hierarchy playing out in a cryogenically optimized environment.”
Hampel also discussed the lifespan of DRAM operating at cryogenic temperatures.
“There are really two things that drive the lifespan of DRAM. The first is Dennard Scaling, meaning, you can buy more DRAM [that is] more energy efficient. Yet at some point, simply maintaining the power profile of that older DRAM just isn’t worth it. As scaling slows, that’s going to happen less rapidly,” he added. “There is also a thermal lifetime of DRAM, with energy and stress induced by [constant] writing and sensing and all the required voltage. So, from a reliability standpoint, we do think [cryogenics] is going to [offer] significant improvements in the reliability of the operation of semiconductor devices at these temperatures – which should increase their lifecycle.”