The increasing cost and complexity of making chips is driving companies to explore 3-D stacks as an alternative way to gain performance or lower power.
Samsung will describe a stack of up to eight DRAM chips that delivers bandwidth of 307 Gbits/s, more than double a 128Gbits/s stack described at ISSCC 2014. Samsung put a phase-lock loop on the 20nm chip to ease testing of the device. To lower heat, it uses “an adaptive refresh scheme considering the [chip’s] temperature distribution,” the paper description said.
Rival SK Hynix will show a DRAM stack delivering 256 Gbit/s. The chip handles “command decoding and bias generation for the memory core…[at the] logic base-layer of the stack” instead of at the memory layer as in previous designs. In addition, it uses “small swing signaling on the heavily loaded [3-D interconnect] to reduce power for driving the interconnects.”
Such dense memory chips will “enable high performance computing, accelerators and small form factor graphics cards,” said ISSCC organizers.
A separate Samsung paper will describe a 256-Gbit flash chip storing three bits per cell and using an on-die stack of 48 cell layers. Samsung has been pioneering such monolithic stacks in flash design, showing a route to denser chips without needing more aggressive process technologies.
Rival Micron aims to leapfrog Samsung’s work, describing a 768-Gbit flash chip with a 64KB page buffer in 179.2mm2 die area. It “achieves the highest density NAND flash memory by placing the peripheral circuits under the array.”
Such dense chips will help drive the market for solid-state disks which is expected to be worth more than $20 billion in 2016, said ISSCC organizers.
The CEA-LETI research institute in France also will shine a light on chip stacking technology with a paper describing a novel 4×4×2 asynchronous network-on-chip for a 3-D circuit built in a 65nm process. The chip is aimed at advanced cellular network gear and “achieves the lowest energy consumption on 3-D I/O power supply at 0.32pJ/b, and the highest data rate at 326 Mbits/s,” researchers said.
Next page: Better fingerprint recognition, machine vision