Domain-specific architectures will rise to improve performances in target markets.
« Previously: Moore's Law's death shakes industry
John Hennessy noted that DRAMs could be among the first major components to hit a wall with the end of Moore's Law. The result would “imbalance [systems] architectures—its full employment for system architects,” quipped Doug Burger.
“Vertical NAND is the closest to a silver bullet” to solve the memory gap left by DRAM, but “you need to put your money on lots of things,” Norm Jouppi said.
Flash memory got its start as storage for digital cameras. Their interfaces have improved, but “we still don’t have the best computing interfaces for flash,” said Butler Lampson, a distinguished engineer at Microsoft and an adjunct professor at MIT.
Whatever comes next, “the new regime [post-Moore’s law] will be transformational…20 years from now this industry will be unrecognisable,” said Burger.
Figure 3: Microsoft's Doug Burger pioneered graph computing at the University of Texas at Austin. (Source: ACM)
Despite the rising costs of fabs leading to a historic consolidation among chip makers, “as long as we have three or four stable players there will be good competition,” said Jouppi, noting “lots of market pressure with demand for the iPhone.”
Burger and Jouppi both predicted a rise in “domain-specific architectures” to optimise performance for specific markets. Burger called it an “ugly” trend leading to “Franken-systems…but the industry is big enough that we can accommodate it,” he said.
“Domain-specific architectures are like works of art, the ones we’ve done so far are good at using transistors,” said Jouppi.
Hennessy wondered aloud if the computer industry will return to an era of vertical companies that design everything from silicon to apps. Apple is moving in this direction, and Google seems set to follow it.
Nevertheless, “there’s [also] a flowering of architectures, some done by silicon vendors and some by cloud vendors,” said Jouppi.
The panel was divided on whether the industry will see a flowering of new processor companies after having consolidated around a handful of instruction sets. “There are very specialised skills needed in chip design so there’s value in consolidating [processor designers] into a few companies,” said Hennessy.
The slow death of Moore’s law is already undermining what has been a strategic abstraction layer isolating software developers from changes in chips—the instruction set architecture (ISA). New, broad interfaces are needed, perhaps for big vertical markets, but it’s unclear what they should be.
With ISAs, software developers “got free improvements for relatively minor changes, but [today’s smartphones already have] maybe six ISAs on a processor and half the [SoC] area is an accelerator with no durable ISA at all,” said Margaret Martonosi of Princeton.
New frameworks and libraries are helping fill the gap, but they create systems that are “difficult to verify and secure…this will get worse,” she said. If market-specific systems start to define their own interfaces they will spawn needs for new design processes and engineers capable of working “up and down the stack,” she added.
Cloud computing workloads are still very diverse, said Burger of Microsoft’s cloud division. Cloud providers have millions of customers and “even big properties run on 1-2% of our servers and can change weekly or monthly—that’s too fast for even FPGAs,” Burger said, noting the so-called programmable chips “are still too hard to program.”
The decline of Moore’s Law is already spawning a much-needed drive for more efficient software and general-purpose processors, panellists said.
“I’m hopeful we can refine software and hardware, we’ve gotten a bit sloppy in both sides and have room to refine efficiency for a long time,” said Jouppi.
“It won’t be as good as Moore’s law, but there’s room in [applications], algorithms and hardware for improvements,” said Lampson, best known for being one of the inventors of the Alto, a seminal system in the development of the PC.