HP wants to rebuild computers with photonic networks and massive memristor storage, – in the next five years
HP has set out an ambitious plan to redefine computing architectures, using photonic networks and memristors to replace entire data centres with single racks of machines, that operate with processing distributed across a memory pool.
The new architecture, referred to simply as “The Machine”, throws out existing practices, and simplifies around emerging technologies, and HP reckons it can deliver it in the next five years, with help from hardware partners, and the open source community to build the hardware and software the project will require.
Electrons, photons and ions
“What would you do if a database could handle 168 GUPS (Giga-updates per second?” asked HP Labs’ chief architect Kirk Resniker, holding a 3D printed mockup of one of The Machine’s computing nodes at HP Discover in Las Vegas.
Current machines including smartphones are essential miniaturisations of the computers built in the 50s and 60s, said John Sontag, vice president of HP Systems Research.
“Computers spend 90 percent of their time shuffling data from cache to store,” he said, which is a waste of capacity and energy. Instead, HP proposes special purpose processing cores, connected to a massive pool of memristor-based memory, using photonic networks.
“We want to use electrons for processing, photons for communication and ions for storage,” said Sontag. Electrons add delays and energy costs in communications, he said, while optical networks can boost the speed, capacity and distance of communications.
“We can’t get rid of copper, though it’s been called an energy sucking devil,” he said. Instead, the electronics can be condensed into special purpose compute nodes,
Ions are used in the futuristic memristors which HP has backed for some years as a high-speed high-capacity replacement for Flash memory.
Partners needed on the way
The HP Labs team showed an approximate timeline, with memristor DIMMs arriving in 2016, and the “distributed mesh computer” coming in 2020, but several steps remain before that. HP will be working with partners on productising System on a chip (SoC) packaging, and delivering usable memristor memory, and with the open source community on completely new operating systems for the machines.
Whether HP buys or partners with other players, the company is clear it will need help. “It’s not HP’s intention to build the entire ecosystem ourselves,” Sontag told TechWeekEuroe. “We are pretty certain there’s no one organisaton that can do all this. We want to bring the others all along.”
But there’s one rival that HP singles out. Comparing the potential of The Machine with IBM’s Blue Gene Q, Sontag said it would delive the same amount of operations on a tiny fraction of the energy. And with a hint of a comparison to IBM’s Watson, he said HP’s aim is “not just building a machine towin a game show. It’s something designed for a 21st century enterprise, to help engineers solve problems of education and employment, elevating the human dignity of eight to ten billion people on the planet.