EE Times: HP Labs sees memristors in server future

[cbn]

http://www.eetimes.com/electronics-...sed-nanostore-is-CPU-for-HP-Labs?pageNumber=0

HP Labs sees ARM, Atom, memristors in server future
Rick Merritt
11/28/2012 3:01 AM EST
SAN JOSE, Calif. – HP Labs continues to make progress on its long term vision of smart memories based on its memristors as alternatives to server CPUs. The devices are one of several new categories of chips likely to seize the moment of disruptive changes in data center technologies and workloads ahead, said an HP Labs researcher.

Separately, Hewlett-Packard Co. is expected to announce within weeks the next steps in its Project Moonshot, its work on ARM- and Atom-based servers. HP is working with a broad group of companies including processor providers AMD, Applied Micro, Calxeda, Cavium and Intel on the project.

So far HP has announced an Atom-based server using Intel’s Centerton processor. It suggested it would use cartridges to flexibly upgrade a single server chassis for a wide range of ARM- and Atom-based chips in 2013 and beyond.

On a three-to-five year horizon, HP Labs is working on what it calls "nanostores." The chips combine memristors and logic that could challenge microprocessors in a new era of designs based on novel system architectures and memory hierarchies, said Parthasarathy Ranganathan, an HP Labs researcher in a keynote at the Server Design Summit here.

“We have the opportunity for new building block,” said Ranganathan. “It’s really a 3-D stack amenable to traditional workloads and even more so to new workloads, really changing the game with potentially a hundred-fold increase in performance per watt."

HP Labs continues to conduct experiments on the nanostore concept with promising results. But Ranganathan declined to provide any specifics, noting the work is still as much as three years from commercial products.

Such devices could ride a confluence of multiple waves of change. “The technology changes and workloads inflections ahead are incredibly interesting for system design,” he said.

In computing, he noted processors made a “sharp right turn” in about 2005 when performance gains for single core processors plateaued and multicore architectures took off.

In storage, disk drive capacity has outpaced data access times.[/b] DRAM capacity growth has taken a “soft right turn” from traditional levels of 60 percent a year to about 25 percent a year, he noted.

The rise of server SoCs and eventually 3-D stacks along with flash memory in server designs could help breakthrough such bottlenecks. The changes come about the same time that networking is shifting more deeply from copper to optical links.

Separately, data growth is far outpacing Moore’s Law, driving new workloads. The researcher noted a “growing complexity and dynamism of data access.” Today’s searches increasingly involve accessing multiple real-time and static databases as well as overlaid sources of personal and contextual information.

“Compared to a simple click, which once was just to a single Web server, we now have very sophisticated data analysis from multiple repositories with complex cross correlations,” he said. “It’s big data, but it’s also fast data from multiple streams with deep analytics."

In a nutshell, "memristor" is a type of non-volatile memory meant to replace NAND and mechanical drives. It is competing with other up and coming non-volatile memory techs such as Phase change memory and STT-MRAM.

Major difference (between the other non-volatile memory techs) is that Memristor can also perform logic functions in addition to memory:

http://arstechnica.com/science/2010/04/complete-logic-set-performed-using-memristors/

Now, in their latest work, the researchers have shown that it's possible to simultaneously use memristors to perform a full set of logic operations at the same time as they function as nonvolatile memory.

http://www.siliconvalleywatcher.com/mt/archives/2010/04/the_miraculous.php

We have now discovered that memristors can be used for logic -- they can be used as processors. This is very significant because instead of shuttling data to the processor and then back again, which takes time and energy, we could shuttle the processing code to the data -- which is smaller and quicker.

- I used to be worried that we are going to reach the limits of Moore's Law and the what do we do. But with memristors, we can easily lay down multiple layers of memristors, effectively extending Moore's Law by decades.

- Using memristors for processing brings other potential changes. Instead of just two states, on or off, as with transistors, memristors can represent many states. This means we can create new types of computing models, we can also create analog computers, which you don't program, but you let them learn. You can then replicate the learning to other memristor analog computers.

- We might be able to use memristors in a similar way to synapses in the human brain.

However, in the beginning (until the logic capability matures) the memristors will need to be used alongside traditional silicon logic. In this form HP calls them "nanostore" (combination of Memristors and silicon CPU).

More information on combining silicon and memristors:

http://arstechnica.com/science/2010/04/complete-logic-set-performed-using-memristors/

Still, the memristors aren't especially useful in isolation at the moment, as the authors had to rely on standard silicon circuitry to control them. They also don't expect that to change overnight, which has implications for how we might see the first deployments of memristors. "Applications of this technology will most likely require substantial parallel operations," the authors write, "in order to amortise any silicon-based driving circuitry."

http://beforeitsnews.com/science-an...aking-nanostores-using-memristors-452599.html

We coined the term nanostores as a duality of microprocessors to reflect the evolution to nanotechnology and the emphasis on data instead of compute. The key property of nanostores is the colocation of processors with nonvolatile storage, eliminating many intervening levels of the storage hierarchy. All data is stored in a single-level nonvolatile memory datastore that replaces traditional disk and DRAM layers—disk use is relegated to archival backups.

For example, a single nanostore chip consists of multiple 3D-stacked layers of dense silicon nonvolatile memories such as PCMs or memristors, with a top layer of power-efficient compute cores. Through-silicon vias are used to provide wide, low-energy datapaths between the processors and the datastores.

 

[cbn]

Deleted.

 

[Idontcare]

This seems very interesting, could become a real breakthrough if it continues to be resourced internally at HP for the long-term as needed to fully commercialize it.

With HP being little more than a train-wreck since Fiorina got hold of it, setting in motion a series of equally disastrous CEOs, what chance is there of Whitman righting HP's ship without cutting funding for this sort of fundamental basic research?