Qualcomm has been claiming some time that its strength in depth spanning key technological areas will help extend its dominant position across wireless as a whole as demand for mobile connectivity proliferates further with more extensive IoT deployment.
At its recent 5G Summit Masterclass in mid-August 2022, the company cited low-power CPUs, wireless chips spanning all the key protocols including cellular, WiFi and Bluetooth, and AI capabilities, as three critical areas where it was a world leader.
Low-power CPUs are vital for many devices including smartphones, while coverage of the key protocols makes it a strong contender in the IoT, which embraces numerous applications and vertical sectors. The AI capabilities are increasingly required for onboard processing and analysis of data to enable actions around the network edge in as near to real time as possible.
Although not directly related, Qualcomm does seem to have been emboldened lately by antitrust successes against regulators that have sought to clip its wings in various ways. Its most comprehensive victory has just come over the EU, helped by a series of mistakes made by the latter during antitrust action going back several years.
In fact, the story began over a decade ago in 2011 when Qualcomm had become Apple’s dominant supplier of chips for its then-emerging range of LTE based iPhones. During the ensuing five years until 2016 Qualcomm made exclusivity payments to Apple that the EU then said were illegal under its antitrust laws because they restricted competition. This led the EU in January 2018 to impose a fine a whisker short of €1bn ($1.02bn).
At the time, the EU was confident this fine would be upheld, and even issued a statement to the effect this would warn other big tech players against similar alleged abuse of market dominance. In the end it was the EU itself that emerged chastened with implications for future antitrust actions in which it might engage.
Ironically, it was another dominant US company, Intel, that the EU had cited as having suffered the greatest harm by being effectively shut out of the LTE chipset market, given that Apple was such a huge customer. Another irony was that Apple has been seeking to wean itself from dependence on Qualcomm for modem chips in the 5G era, even though those efforts have yet to come to fruition.
Irrespective of whether Qualcomm continues making iPhone 5G chips, the European General Court sided with Qualcomm against the EU for several reasons, citing various procedural irregularities that affected the case. The most decisive irregularity was the Commission’s failure, in the Court’s eyes, to retain and then make available to Qualcomm detailed records of meetings and calls it had held with third parties. The Court also noted that Apple had had little choice from the outset in using Qualcomm chips if it was to be competitive, and that this was not the result of incentive payments.
This led the Court to quash the fine in June 2022, leaving the EU with right of appeal. It finally became clear this week at the end of August 2022 that the EU had decided not to appeal, leaving Qualcomm to score a major victory.
The case had echoes of an earlier one in the USA, again a long-running saga where the Federal Trade Commission (FTC) had begun investigating Qualcomm in 2014 for charging what it considered “unreasonably high royalty rates” for its intellectual property, subsequently issuing an order to renegotiate licensing deals. Qualcomm was accused of exerting pressure on vendors to buy at least 85% of their chips from it, greatly damaging competition.
But an appeals court finally overturned this ruling in August 2020, concluding that while “anti-competitive behavior is illegal under federal antitrust law, hypercompetitive behavior is not”.
Nonetheless, if Qualcomm’s latest pronouncements about technical superiority in all key departments are true it could well run into antitrust action again before long, both in the EU and USA, where in both cases the appeals court verdicts have elicited considerable resentment.
In all three areas of presumed excellence, Qualcomm has been reliant on key partners. For the advanced silicon it has been juggling between two foundries, Samsung and TSMC, which have each picked up significant orders. Samsung Foundry won the contract to manufacture the Snapdragon 8 Generation 1 8-core chipset announced in December 2021 using a 4-nanometer process technology.
But amid rumors Samsung had been unable to deliver stable supplies of these and other chips, Qualcomm then went to TSMC for the upgraded version Snapdragon 8+ Generation 1 announced a few months later. The specification was very similar with just a slight increase in performance, but it appears that the switch in foundries to TSMC has improved thermal performance, which had been a concern. Both chips are the highest performing for smartphones yet made, but the later Snapdragon 8+ Generation 1 now looks like being more widely used.
On the AI front, Qualcomm has relied heavily on Google for incorporation of efficient model creation and automation techniques on the Snapdragon processors. The two companies have been collaborating over neural architecture search (NAS), which has emerged as an important technique for creating and then optimizing neural networks for deep leaning applications automatically rather than manually. These were available first on the Snapdragon 8 Gen 1 and are then being extended across other Qualcomm platforms. Qualcomm has claimed it was Google’s first system-on-chip (SoC) customer to incorporate the Google Cloud Vertex AI NAS in its silicon.
We can identify three key benefits for such incorporation of automated model training and tuning, at least with Google’s tools. First, it reduces the amount of data science expertise needed to bring models to the point of use, or value as Google puts it, which reduces cost and time.
Second and related to this, about 80% fewer lines of code are needed to train a model with custom libraries.
Third, there is a strong focus on explainable AI, so that the resulting models can justify deductions made rather than just issuing them as if from a black box. There has been growing interest in and demand for explainable AI, essentially a set of tools for understanding and interpreting predictions made by machine learning models. In Google’s case these are now natively integrated with some of its AI products and services. As Google pointed out, the benefit is not just to explain predictions and even satisfy regulatory requirements in future, but also to help debug and tune the models more easily than would be possible otherwise.
The other notable aspect of Qualcomm’s recent renewed marketing offensive is its emphasis on the value of millimeter wave spectrum. At best it can be said that mmWave has enjoyed a mixed reception in the industry so far, with some operators to varying degrees reining back on it. We can see from the much lower prices paid in auctions per unit bandwidth for spectrum at these higher frequencies that operators value it much less than the prime midband spectrum, or indeed than low-band spectrum. Qualcomm contends that such low valuations are misplaced and that mmWave spectrum at current prices represents a bargain.
It has recently cited various studies that appear to support this argument. One such study conducted by the Signal Research Group (SRG) and presented at Qualcomm’s recent 5G masterclass featured tests in downtown Tokyo, where performance measurements were taken on standard smartphones roaming around. A comparison was made between three options, LTE and 5G both operating at midband frequencies, and 5G at mmWave frequencies.
The study commissioned by Qualcomm tested 5G mmWave from NTT Docomo, KDDI, Softbank and Rakuten Mobile at two locations, using a combination of stationary and walk-testing. For the downlink, midband 5G achieved 2.3 times greater capacity than LTE for comparable coverage. 5G mmWave increased capacity by 10.1 times over LTE and 4.4 times over midband 5G but at 50% reduced coverage.
This led SRG to discuss the concept of the “capacity cake” where coverage is squared up with throughput. On that basis, mmWave as implemented in Japan yields a capacity cake with just over double the volume even of midband 5G, calculated by multiplying that 4.4 times increase in capacity with the halving of coverage area.
Of course, this is achieved as a result of the extra spectral bandwidth, so does not mean necessarily that operators should be paying substantially more for mmWave than they are doing. But they should consider it for these dense deployments according to SRG as well as Qualcomm, because mmWave is not just about headline speeds, impressive though those are, but providing capacity that will be required as the midband becomes saturated.
In those Tokyo tests, SRG also noted that use of mmWave improved LTE performance by offloading traffic from it. That would also apply presumably to midband 5G, suggesting that roll-out in mmWave will also yield improved performance in areas it does not yet reach and so deliver an overall dividend.
SRG also pointed out that optimization of mmWave was still work in progress with plenty more scope for further improvements in both capacity and headline performance. According to SRG, maximum sustained downlink data speeds over 5G mmWave have increased by 66% over the last three years, along with greater effective range and, now, impressive improvements in uplink throughput as evidenced by the Tokyo tests. In those tests, 300Mbps was commonly observed in the uplink and about 2Gbps over the downlink at a distance of 160 meters from the site, using 400 MHz of spectrum.
“And the 5G standard continues to offer up new features to be implemented in the mmWave range that will improve performance even more,” SRG suggested. “Four-carrier aggregation in the uplink, for instance, will further widen the performance gap between mmWave and midband 5G”.