To increase the throughput of wireless networks, companies are expanding into higher spectrum bands than ever before to connect everything from smartphones to cars to the internet. The problem is that these millimeter waves can be blocked by buildings and trees and absorbed by oxygen over long distances.
The solution is to use beamforming to steer antenna beams around obstacles. The technology measures channel state information like phase and gain in specific slices of spectrum. But instead of using traditional software to adapt transmissions to the channel conditions, several companies are looking into machine learning.
Among them is Ceva, a chip designer that licenses digital signal processors that convert everything from voices to wireless signals into digital data. The company recently said that it had designed a custom accelerator that can be embedded in 5G modems to enable advanced beamforming and link adaptation.
The custom processor is one of the building blocks of the company’s new PentaG platform. The blocks can be mixed and matched with chips from other companies in devices like smartphones, industrial sensors, or fixed wireless access equipment using millimeter waves to replace fiber optic cables connecting homes and businesses.
The accelerator was built to cut the latency and lower the power consumption to meet the requirements of the 5G standard, the first version of which was published by the 3G Partnership Project late last year. The modem can support downloads of up to 10 gigabits per second – ten times faster than the fastest 4G modems.
That throughput will not be available without more development. On Tuesday, Qualcomm said that it had demonstrated median downloads of 490 megabits per second with the Non-Standalone 5G standard, which was finished in December as an intermediary step toward the pending Standalone 5G standard.
“The 5G standard poses immense challenges for wireless semiconductor designers to build modems that are powerful, yet efficient enough to meet the strict power budget constraints of enhanced mobile broadband,” said Michael Boukaya, vice president and general manager of Ceva’s wireless business unit, in a statement.
“We designed PentaG to overcome the limitations of conventional DSP-based LTE/LTE-Advanced architectures, which simply do not scale to support 5G,” he added.
Ceva wants to use neural networks to make sense of the channel state information that smartphones and other devices continuously send to cellular base stations. That way, these base stations can tailor transmissions to preserve the throughput and latency of 5G. They can also protect signals from scattering or absorption.
Ceva said that the modular architecture of its modem could scale from Non-Standalone 5G, which only requires a software update to 4G infrastructure, to the Standalone 5G standard. It could also work with both millimeter waves and more traditional bands under six gigahertz used in communications today.
In addition to the neural network processor, the PentaG architecture includes an advanced digital signal processor with an extension for 5G physical layer control. It also features a programmable processor with 64 MAC units that handle channel estimation measurements for 5G, which requires five-times lower latency than 4G.
The company said that it would start licensing PentaG to early customers in the second quarter, with general availability in the third quarter. The building blocks of the platform can be licensed together or separately – combined with a customer's own silicon to speed time-to-market. Ceva's chips have shipped in more than eight billion phones worldwide.
Ceva has been trying to shift into other fields of artificial intelligence. Last month, it announced a new line of designs for classifying objects in photographs, processing human speech, and completing other tasks. The most advanced NeuPro design delivers up to 12.5 trillion operations per second on machine learning workloads.