Processer I/O made of fiber, not copper, could help reconfigure data centers.
Startup AyarLabs is using a combination of high-bandwidth fiberoptics, low-cost CMOS fabrication and careful target selection to strike efficiently at the datacenter’s worst bottleneck.
“Moore’s Law only covers the processor, not how we move data in and out of it during processing or how to get the processor and memory working at the same speed,” according to Alexandra Wright-Gladstein, co-founder and CEO of Ayar Labs. “We are developing optical transceivers, but we’re very focused on chip I/O , not optics per se. We have a high-efficiency optical-electrical conversion engine with a very small footprint, so our optical devices are tiny. We build it using ordinary CMOS design and fabrication and focus on efficiency, so we end up with a product that uses 10 times less power than any commercial optical transceiver, whether silicon-based or traditional.”
The company is a spin-out from a 10-year MIT research project that developed a dual-core RISC-V architecture chip with 70 million transistors,850 million photonic components, 1MB of RAM, and a fiber I/O port theoretically capable of sending at at 550 Gb/sec and receiving 900Gb/sec.
When the team combined two chips, one serving a processor, the other as memory, they were able to pass data at 2.5Gbit/sec in each direction, expanding bandwidth as needed by adding wavelengths from the external laser that provided light to the chips.
The project was partly funded by two DARPA programs, one designed to increase data rates by combining processing with photonics, and another whose focus was on getting the highest number of FLOPS possible from every watt.
A paper describing the result was published in the journal Nature in December, 2015, complete with a video showing how it works.
The SoC approach is designed partly for efficiency and partly to bridge the gap in design and manufacturing between CMOS and photonics products, according to the paper, which also describes the “zero-change” approach to developing photonics using CMOS techniqes and materials in the integration of photonics to avoid the complication of custom fabrication of photonic components.
The 45nm chip was the first designed with an optical I/O. The commercial product is a lot different than the one described in the journal article, according to Wright-Gladstein, but the research version did more than just give co-founders Chen Sun and Mark Wade a chance to pursue their Ph.D.s.
It also brought in some startup capital after Wright-Gladstein entered the nascent company — then called Optibit — in MIT’s Clean Energy Prize competition and brought home both grand prizes, which totaled $275,000. https://www.scientificcomputing.com/news/2015/05/optical-chips-team-develops-way-integrate-fiber-optics-computer-chips
Optics give datacenters tools to retool
Intel’s 100 Gbit/sec silicon photonic network interface gave datacenter architects the opportunity to change the fundamental design of datacenters by eliminating much of the penalty of distance. Minimizing the delay in providing a Google Translation or passing on a Facebook “Like” used to mean larding each individual server with as much memory as possible, and keeping HDD storage in the rack or very close by so that data not stored in main memory didn’t have far to go.
Replacing copper running at 10Gbit/sec, 25 Gbit/sec or even 40Gbitsec with fiber running 100Gbit/sec, while using less energy and generating less heat, allows hyperscale datacenters to pack in even more hardware and delay just a little the need to build a new one. The year it delivered its first photonics product, 2016, was also the first time that cloud service providers bought more Intel chips than Intel’s traditional customers, according to a 451 Research paper about the project.
Sales of silicon photonic switches generated only $30 million during 2016, according to researchers at Yole Développement, but will make up 35% of the market for intra-datacenter communication by 2025, generating sales worth $560 million. Cloud service companies and large data centers are doing most of the spending to increase the bandwidth of rack-top switches, and increasingly to link geographically separate facilities using wave-division multiplexing. WDM sales rose 225% during 2017 compared to 2016, according to analysts at the Dell’Oro Group.
“Moore’s Law only covers the processor, not how we move data in and out of it during processing or how to get the processor and memory working at the same speed.”
Gordon Moore first postulated what would become Moore’s Law in 1965. That’s six years before the introduction of the first commercial microprocessor. Moore’s Law covers chip density, period. Regardless of what the chip does. Dennard scaling, now thoroughly dead, covered chip speed and power consumption.
The basic sentiment in this article is true. I/O speeds are a real challenge, whether you’re discussing processor-to-memory or processor-to-network I/O. There’s some amazing progress with both copper and optical interconnect taking place right now. Some real breakthrough technology.