Your strategy change in 2018 was not only about the new "Data Center First" initiative but also about increasing growth rates in the core business. Has the strategy worked?
Yes. We have more or less always grown in our core markets, but our typical growth rate a few years ago was in the low single digits. The goal was to bring the growth rates here into the higher single-digit range. And we have succeeded in doing that.
What are the reasons for this success?
We have significantly increased value creation. Put simply: we used to supply FPGAs, but today we supply components such as the MPSoCs, which not only contain the programmable logic fabric, but also processor subsystems and various accelerators - and we have also integrated analogue technology or a NoC. In addition, Xilinx is now also increasingly focusing on platforms. These include our Alveo family with various accelerator cards, Vitis as a development platform and, most recently, we also offer SoMs with our Kria family. None of this existed a few years ago.
These developments have paid off. If you look at the communications market, wireless and wired, our revenue has grown by 14 per cent on average since 2018. That doesn't sound like much, but in that context, it's important to note that we lost sales to Huawei because of the trade tensions between the US and China.
Initially, it was said that we were not allowed to deliver some products and then in September 2020 it was said that we were no longer allowed to deliver anything at all, which means that since then we no longer make any sales with Huawei. This is not easy for Xilinx, because Huawei is not only the largest communications company in the world, but also one of Xilinx's top customers. In addition, there are other companies that were affected by these trade problems, some directly, some indirectly. Accordingly, the 14 percent can definitely be considered a success.
And what about automotive, ISM and A&D?
The automotive segment is the second fastest growing segment for Xilinx after the data centre segment. And here, too, I can say: since 2018, we have been growing at double-digit rates on average. Last year was of course difficult, first the market fell sharply due to the pandemic, then it recovered very quickly, and now the industry is struggling with supply problems. Still, the growth rate over the last three years is pretty good. The industrial sector has also done well. And once again, we are not relying on long-established methods in this market segment either, but are quite innovative. Kria proves that. The first SoM is optimised for industrial applications. And that is only the beginning. This also shows that we are moving up the value chain and making a lot of efforts to accelerate development for our customers. This not only has advantages for our customers, but we also generate sales more quickly. And the medical sector has also grown very well in the last three years. A&D, i.e. Aerospace and Defense, has grown by more than 45 percent in the last five years.
In total, sales in all our core markets have grown by 9 per cent on average since 2018, some more, some less. But that means we have achieved our target.
The third item on your strategy list for 2018 was to continue to advance adaptive computing, how is that going?
Xilinx is the largest PLD manufacturer, so adaptive computing is nothing new for us in this sense. What is different today, however, is AI, which we are focusing on very strong, even if the market itself is only just beginning.
But Xilinx is also competing with many other semiconductor manufacturers here...
Yes, but we have a decisive advantage, even if many developers are not aware of it. An application does not only consist of AI algorithms. Even in image processing, pre-processing of the data is still necessary before the actual AI algorithm, and often post-processing as well. Therefore, it does not make sense to focus exclusively on TOPS. And the advantage of Xilinx is that we can accelerate the entire application, not just the AI algorithms. That's why we can often realise a better application experience with our FPGAs, even though we perform worse in the pure TOPS comparisons.
TOPS are nevertheless a critical parameter in AI applications...
I didn't want to say that we did badly on this parameter, but it is not the only parameter that needs to be taken into account. Our 7nm Versal components are equipped with an AI engine that is characterised by high computing power per watt and also per area and that supports both fixed and floating point data formats. With our components, application-specific data paths, memory hierarchies and data movements can be realised. Thanks to these features, the AI Engine also performs really well. For example, if you compare an UltraScale+ IC based on 16 nm with a Versal component with AI Engine, Versal achieves a factor of 20 higher computing performance in image classification. Versal also outperforms UltraScale+ by a factor of 5 for 5G and Massive MIMO systems - all with 40 percent lower power consumption.
And we can also boast higher performance against GPUs. A comparison using the Google Speech Commands dataset showed that Alveo U250 outperforms nVidia V100. If one then adds the advantages that Numenta's sparsity algorithms bring, the computing power of Alveo U250 with sparsity algorithms turns out to be higher by a factor of 100 compared to nVidia V100.
The example with Numenta shows the performance made possible by the combination of the logic fabric and the AI engine. If this combination is used correctly, enormous leaps in performance can be achieved. That's why I'm sure: adaptive computing will become even more important.
This means: the strategy change of 2018 was a complete success and we will continue to pursue this strategy in the future.