In April I wrote about an ambitious project by Hewlett-Packard to use an electronic device for storing data called the memristor to reinvent the basic design of computers (see “Machine Dreams”). This week HP chief technology officer Martin Fink, who started and leads the project, announced a rethink of the project amidst uncertainty over the memristor’s future.
Fink and other HP executives had previously estimated that they would have the core technologies needed for the computer they dubbed “the Machine” in testing sometime in 2016. They used the timeline at the bottom of this post to sketch out where the project was headed.
But the New York Times reported yesterday that the project has been “repositioned” to focus on delivering the Machine using less exotic memory technologies–the DRAM found in most computers today and a technology just entering production called phase change memory, which stores data by melting a special material and controlling how it cools.
With memristors out of the picture, there’s reason to doubt just how revolutionary HP’s project can be.
The main feature of the Machine’s design was to be a large collection of memristor memory chips. They would allow computers to be more powerful and energy efficient by combining the best properties of two different components of today’s machines: the speed of the DRAM that holds data while a processor uses it, and the capacity and ability to hold data without power seen in storage drives based on hard disks or flash memory.
Prototypes of the Machine built with DRAM and phase change memory in the place of memristors had always been part of the plan. But when I met Fink and others working on the project I also heard that those technologies would hobble the idea at the heart of the Machine.
Because DRAM can’t store data very densely and must always be powered on, computers built around a large block of it will require a lot of space and power. Meanwhile, phase change memory is too slow compared to DRAM to be much use for data being worked on. When I met Stan Williams, who leads HP’s work on memristors, he dismissed the idea that any other technology could be used to reinvent the basic design of computers as HP wanted. Fink did a good job in this 2014 blog post of explaining why his team believed only memristors could build the Machine.
Still, this week’s climb down is not a complete surprise. Fink used the timeline below as recently as December 2014, predicting that memristor memory would “sample” in 2015 and be “launched” in 2016. But a few months later, in February of this year, he told me that sampling was most likely in 2016–an estimate that HP’s manufacturing partner SK Hynix would not confirm. Microelectronics experts I spoke to said that it looked to be challenging to make reliable memristors in large, dense arrays as needed to make a memory chip.
HP now appears to be avoiding making any prediction for when the technology will be mature. The company has not yet responded to a request for comment.