Neural Networks Meet 5G with Ceva IP

Article By : Junko Yoshida, EE Times

PentaG platform to enable AI on the go, MWC set to be hotbed for 5G use cases

PARIS — Participants at the Mobile World Congress in Barcelona next week will be plunging hip-deep into 5G use cases to discuss the myriad challenges involved, ranging from massive machine communications (NarrowBand-IoT) to ultra-low latency-required communication for connected cars. But now that the first phase of 5G standardization is complete, with the publication of 3GPP-Release 15, the biggest deal in the mobile industry is how best and how fast it can bring to customers’ equipment 5G New Radio, very-wide-bandwidth radio channels supporting data-access speeds of up to 10 Gbit per second.

In short, so-called “enhanced Mobile BroadBand (eMBB)” is everyone’s main chance right now, and that’s where a fierce 5G race will unfold in sunny Catalonia.

Among tech suppliers, Ceva Inc. will be in Barcelona, riding high with its comprehensive 5G New Radio enhanced Mobile Broadband IP Platform. Ceva devised its platform for designers of 5G smartphones, fixed wireless access and embedded devices.

Gideon Wertheizer
Gideon Wertheizer

Gideon Wertheizer, Ceva’s CEO, told EE Times, “Many companies — whether they plan to use 5G for the fixed wireless or embedded systems — are looking for technology they can bootstrap themselves to reach the entry point of the 5G market” as soon as possible. Calling 5G “extremely complicated technology,” Wertheizer said, “They are telling us they need something now.”

With a 5G IP platform called PentaG, Ceva is offering “a 5G modem subsystem,” Emmanuel Gresset, business development director of the wireless business unit at CEVA, told us.

The PentaG, based on a flexible architecture, has several key IP components. These include Ceva’s enhanced XC4500 DSP with 5G ISA extensions, a PentaG AI processor and vector MAC unit co-processor, a cluster of Ceva-X2 DSPs and a polar and LDPC encoder and decoder accelerator. They also come with necessary software to go with them.

(Source: Ceva)
(Source: Ceva)

Ceva brags about the flexibility of the PnetaG platform. It allows potential customers to “mix and match” any of PentaG IPs with their own design, explained Ceva CEO. “The PentaG architecture does not include a CPU,” he noted. This is deliberate, he added, “so that customers who want to use their own, legacy modem solutions can continue to do so,” while they can bring in IP blocks from PentaG to enhance their 5G offerings.

Neural networks meet 5G
Perhaps most notable in the PentaG platform is its built-in AI processor. “We are the first to use neural networks for communication,” claimed Ceva’s Gresset.

Wertheizer explained, “We found machine learning works really well when we have to figure out channel behavior in communication.” He added, “Neural networks work best when you need something you can’t predict by programming.”

Ceva has already trained the communication-related dataset on its own, offering customers patent-pending algorithms that go with the AI processor.


To maximize data throughput in 5G, a smartphone must send accurate channel-state information (CSI) to base stations. The CSI gives instructions to base stations so that in sending data, the stations can adapt transmissions to current channel conditions and certain parameters specified by smartphones. Constant closed-loop contact between smartphone and base station enables reliable communication with high-data rates in multiantenna systems, explained Gresset.

Asked about the AI processor inside Ceva’s 5G IP platform, Mike Demler, senior analyst at the Linley Group, told us, “As far as I know the use of a neural-network engine in PentaG is unique. It’s a very clever way to address the complexity and variability of 5G communication channels.” He added, “Neural network algorithms are good at identifying patterns in whatever data they are analyzing, and in PentaG-based modems designers can train the network to produce an optimum set of operating parameters that adapt to dynamically varying conditions.”

Asked why the neural network engine for 5G, Demler explained, “It’s all a matter of optimization.” He said, “Back in the analog radio days you would tune the knob to pull in a weak signal. Back in the day, I spent a lot of time ‘DXing’ on my old Transoceanic shortwave radio. Because of the 5G complexity, it would be more difficult to have a processor perform that optimization in real time, especially given the low-latency requirements. So, pretraining a neural-network inference engine to automatically adapt to varying link conditions is a better solution.”

Wertheizer declined to provide details of the AI processor architecture. “Not at this point,” he said.  Ceva already announced last month at the Consumer Electronics Show what it calls NeuPro, a dedicated low-power AI processor family designed to do deep learning at the edge. The Ceva CEO called NeuPro “a general-purpose AI processor,” while describing the new processor inside the PentaG as “a special-purpose communication machine learning processor.”

Ceva was able to develop a special-purpose communication AI processor based on its experience — more than four years — building a tool chain for neural networks. Ceva used this with NeuPro, and it became handy for designing a communication-specific neural network processor.

Wertheizer stressed, “Although we happened to find the use case [of this AI processor] in CSI, this engine can be extended to other applications as well.”


PentaG block diagram (Source: Ceva)

Why is 5G New Radio so hard?
Challenges in 5G NR are many. First, it requires high performance and optimized power.

5G NR enables a multi-gigabit capacity of 8 to10 Gps, compared to the LTE data rate of 100Mbps. What makes this tough, said Gresset, is that the 5G NR baseband processor, which is now receiving so much data, must process faster. 5G’s ultra-low latency requirements shorten the time for data processing by five times, compared to LTE.

Second, 5G also demands multi-mode architecture. A 5G-NR modem requires a unified platform for sub-6GHz and mmWave Multi-Radio Access Technology, while also supporting LTE-A Evolution.

Third, flexibility is key to 5G NR optimization. 5G leverages massive MIMO algorithms, advanced beamforming techniques and complex link-adaptation schemes to sustain high-throughput.

But the crux of the issue, as far as Demler sees it, is that “the 3GPP standards are still under development, so a modem design needs to be flexible to support evolving specifications.” The phase two of the 5G standard is expected to come together at the end of this year. He noted, “Technically, the designs are challenging because of the increased frequency-band diversity/modes compared to 4G, greater degree of carrier aggregation, increased channel bandwidth, etc., which all require higher performance and more sophisticated processing.”



Competitive landscape
Thus far, Qualcomm and Intel are the only two companies who have shown prototypes of their 5G modems.

Demler said, “I’ve only seen prototypes, such as their demo at the Olympics. Qualcomm’s X50 is targeting 2019 production. Intel also announced its 5G-only XMM 8060 is scheduled for production in 2019, but it needs a separate chip to support 4G.”

Although Ceva might have picked up a few customers already using PentaG IP, Demler cautioned, “I don’t expect to see any chips until sometime next year.”

Put simply, “We have yet to see any details on the Intel and Qualcomm designs, and Ceva is the only IP vendor I’m aware of offering a comprehensive hardware/software solution for 5G terminals,” said Demler. He believes PentaG will appeal to Ceva’s existing customers, since it employs some of the same building blocks as their 4G solutions. “In the end, any chips integrating PentaG need to compete against Qualcomm and Intel, but this is critical IP enabling other companies to move up to 5G,” he said.

Ceva, however, is hoping that the company’s PentaG IP platform can disrupt traditional IP licensee dynamics. In addition to the usual suspects picking up the entire 5G-NR platform to design their own 5G baseband processors, incumbent LTE players such as Qualcomm, MediaTek or Intel who do not use Ceva’s IPs for their modem offerings could, in theory, license the PnetaG platform’s AI processor, for example.

(Source: Ceva)

(Source: Ceva)

Demler agreed. “It will be interesting to see if Ceva holds some key patents that enable them to license their link-adaptation technology separate from the other baseband components, perhaps even to a modem manufacturer like Qualcomm or Intel.”

Asked about PentaG’s potential customers, Wertheizer included everyone planning to go 5G. “We are offering a new way of looking at modem architecture.” In addition to incumbents like Qualcommm, MediaTek and Spreadtrum, Wertheizer said that semiconductor guys who want to newly enter 5G, and many more OEMs such as Samsung and Huawei are all fair game.

— Junko Yoshida, Chief International Correspondent, EE Times

Leave a comment