Sun Microsystems cofounder and networking guru Andy Bechtholsheim predicted that networking chips — which determine how quickly you can surf the Internet — will keep following the path of progress that it has for decades. Moore’s Law, the prediction in 1965 by Intel’s Gordon Moore that the number of transistors on a chip will double every two years, is still holding up.
In the next 20 years, Bechtolsheim expects an improvement of 1,000 times in chip performance. We should all greet that with relief, since the $1 trillion-plus electronics economy depends on the continuous efficiencies gained from making chips smaller, faster, and cheaper.
“We are in the Golden Age of networking, driven by Moore’s Law,” said Bechtolsheim in a keynote speech at the Linley Tech Processor conference in San Jose, Calif.
Bechtolsheim is worth listening to. He is the founder, chairman, and chief development officer at networking hardware firm Arista Networks, which builds cloud-based appliances for large data centers. He was the chief system architect at Sun and became famous as the angel who funded Google.
He also started Granite Systems, which Cisco acquired in 1996. He developed a series of switches at Cisco and also founded Kealia, which Sun acquired in 2004.
Bechtolsheim talked a lot about leaf switches and buffers and spines and other stuff that was way over my head. But he closed his talk with a series of predictions about the future of Moore’s Law and its relevance to the future of networking, which depends on data centers with tons of servers, each with lots of chips, powered by multiple computing brains (or cores). In data centers, keeping the flow of data moving as fast as possible between the outside world, through memory, into processors and into long-term storage is of paramount concern. It’s a realm in which nanoseconds matter.
Today’s networking chips transfer data at rates of 10 gigabits a second, 40 gigabits a second, or 100 gigabits a second. Part of that depends on chips, but it also depends on optical components that transfer data using laser components, which are harder to improve compared to silicon chip technology.
“Optics, unfortunately, is not on Moore’s Law,” said Bechtolsheim. But he remained optimistic about progress in the future.
1. Moore’s Law is alive and well. By doubling the number of transistors per chip every two years, chip designers will be able to keep feeding faster and cheaper chips to networking-gear designers. In the next 12 years, the path ahead is clear, Bechtolsheim said. That will give us almost 100 times more transistors — the basic on-off switches of digital computers — on every chip.
2. The economics of chips are changing. Each generation of chip design is getting more expensive as it takes more engineers to craft designs from the greater number of transistors available. Designing a new complex switch chip for networks can cost $20 million. Making chips that sell in low volumes is no longer viable. Chip startups often can’t afford to do this anymore. And in-house designs make less sense.
3. Merchant-network silicon vendors will gain market share. Those who design chips that many different system companies use will likely prevail over those who design in-house chips for just one vendor. Moreover, the differentiation now happens in the software that runs on the chip, not in the hardware. And internal design teams often can’t keep up with advances in silicon design tools on the merchant market.
4. Custom designs lead the way. Custom designs can get more bang for the buck out of the available transistors. So even the merchant silicon vendors will have to modify solutions for each customer.
5. Using the best available silicon manufacturing technology is the key. With each new manufacturing generation, chips become faster, smaller, and cheaper. Today’s silicon chip designs have to be built in 28-nanometer technology or better. Those designs must use less power, access more memory, and perform faster.
“No one wants to roll the clock back, and the silicon march is relentless,” Bechtolsheim said.
6. Product life cycles are shorter. Each new silicon chip has a shorter life, but it can ship in higher volumes. The days of 10-year product life cycles are gone and will never come back. Chip designers and system makers can count on frequent upgrade cycles, but they’re face more competition.
7. Architecture matters. Having a faster internal engine makes a car run faster. That’s also true for a chip. With better design at the component level, the overall chip and system run better. This requires rethinking approaches that worked in the past for a more modern technology. Keeping the data flowing within the chip is critical.
8. Flexibility matters. Chips are becoming more versatile and programmable. They can support a variety of protocols and usage cases. Flexibility allows for reuse over generations and expansion to new markets.
9. Building blocks matter. In the age of multitasking, multiple components matter. Replicating cores, or brains, across a chip is the way to faster, more reliable, and lower-power chips. Every component is reusable.
10. The system is the chip. In the future, with more efficient manufacturing technology, future switch chips will be single-chip designs. That requires close communication between makers of systems, software designers, and chip vendors. Anyone who tries to lock out any of the other parties will likely be doomed.
“In conclusion, Moore’s Law is alive and well,” Bechtolsheim said.