Multi-band RF interconnect speeds network-on-chip

Bringing a new approach to on-chip communications in multi-processor systems on chip (SoCs), researchers at the University of California at Los Angeles (UCLA) have developed a multi-band RF interconnect technology. As well as boosting data transmission speeds and lowering latency compared to existing network-on-chip (NoC) approaches, the technology provides a reconfigurable on-chip interconnect architecture, researchers say.

 

The research is described in a paper to be given at the High Performance Computer Architecture (HPCA '08) Symposium to be held Feb. 16-20 in Salt Lake City, Utah. Principal investigators include Jason Cong, chairman of UCLA's computer science department; Glenn Reinman, assistant professor of computer science at UCLA; and Frank Chang, professor of electrical engineering at UCLA.

 

"We are very bullish on this," Cong said. "I think this will open a new wave of on-chip communications. The latency and power advantages alone are attractive, and the reconfigurability opens a lot of new possibilities."

 

Instead of charging and discharging signals over traditional RC [resistance-capacitance] interconnect wires, the RF interconnect technology transmits waves over transmission lines, Cong noted. Data transmission is therefore at the speed of light. And bandwidth is greatly magnified because it's possible to assign multiple frequency channels, or "bands," that transmit data streams over the same piece of wire.

 

According to the HPCA paper, researchers were able to demonstrate an average 13 percent boost in application performance and an average 22 percent reduction in packet latency compared to a traditional NoC mesh, with only 0.13 percent area overhead. The multi-band RF interconnect also claims lower power consumption compared to traditional global interconnect.

 

However, because it requires extra circuitry and silicon area, it's not practical to use RF interconnect for the entire chip. Thus, the HPCA paper describes a two-layer hybrid NoC scheme called Mesh Overlaid with RF Interconnect (MORFIC), where a conventional NoC mesh provides short-distance communications and the RF interconnect handles long-distance communications between processors on a multicore IC. "We can think about the RF interconnects being the superhighways, and the mesh being the local surface streets," Cong said. "Not everyone needs access to the highways."

 

UCLA has done previous work with off-chip RF interconnect, but the HPCA paper is the first to describe on-chip RF interconnect in a multicore scenario, Cong said. The paper describes a year-long pilot project funded by DARPA and the Gigascale Systems Research Center (GSRC). Cong said that Intel developers have participated in some of the planning meetings of the UCLA RF interconnect research group.

"Multi-band RF interconnect provides an alternative way to look at the multicore high performance architecture and effectiveness of the network interconnect structure," said George Cai, principal architect at Intel Corp. "If it can simplify the network interconnect and the transaction scheduling in multiple processors, it could bring additional performance and power benefits. The multiple band RF transmitter/receiver and interconnect controllers of the UCLA RF interconnect research are in the detailed design phase. We need to check the actual measurement results against the expectation."

 

What's driving the research is the prospect of multicore ICs with dozens or hundreds of processors, and the need for an efficient on-chip communications mechanism. "You can see that the whole industry has made the move from traditional frequency scaling of single microprocessors to multicore applications, and there's a huge push to see what we can use for on-chip communications," Cong said.

 

While processors are becoming increasingly capable, current baseband on-chip interconnect is "sort of stuck" in terms of clock rate and limited to about 5 Gbits/second, Cong said. "We do not need to extend the baseband further," he said. "We offer an additional band called the RF band. We propose to use multiple bands so that none of the device capabilities will be wasted."

 

Conventional RC interconnect, according to Cong, uses only a small fraction of the available bandwidth on a multicore IC. The HPCA paper notes that in 90 nm CMOS technology, the typical repeater signal runs at around 4 Gbits/second, requiring it to occupy only about 4 GHz of bandwidth. But since the f_T (frequency at unity current gain) of 90 nm CMOS transistors is around 120 GHz, only a small percentage of the total available bandwidth is used.

 

Because RF interconnect uses electromagnetic waves, it can transmit data at the speed of electromagnetic waves (effectively, that of light) in the transmission line material. This cuts data latency by a factor of 7 in 22 nm technology compared to RC lines, according to the paper.

 

One of the most significant benefits of RF interconnect is the ability to modulate data signals into multiple carrier frequencies. For example, one channel or frequency band could run between 10 and 15 GHz, another between 20 and 25 GHz, and so on. The paper gives a hypothetical example of a ten-carrier RF interconnect line that uses frequencies ranging from 20 GHz to 200 GHz, where each band transmits a 10 Gbit/second data stream. The total aggregate data rate in this example is thus 100 Gbits/second for one transmission line.

 

"The beauty of RF interconnect is that you can have 10 or 12 channels on it," Cong said. "If I put 10 cores on the same interconnect bus line and each one has a transceiver, I can create a virtual crossbar where any two cores can talk just by selecting a common frequency. You can have multiple conversations going on that do not interfere with each other." It's analogous, Cong said, to the way many different cell phone users can talk to each other over different frequencies.

 

With a 5 GHz separation between signals, it may be possible to have 10 to 30 channels on a single wire "comfortably" at 90 nm, Cong said. If individual channels don't need to reserve 10 GHz, then 30 to 100 channels is possible, he said.

 

One challenge posed by multi-band interconnect is the assignment of frequencies. According to Reinman, this can be handled through "coarse grained mapping" of frequencies to access points.  In this approach, hardware counters monitor dataflow communications and then adapt the frequency to avoid latency, he said. This approach, he said, "shows you what the best frequency assignment would actually be."