A Battle in Wireless Networking Semis

This year’s MWC did not have a lot of semiconductor news. In the past the show was the hot bed of some of the most important developments in semis, but that center of gravity has shifted. That being said, there is still one important semiconductor battle taking place in the wireless world – networking chips for 5G base stations.

Modern base stations now come in multiple pieces. To greatly over-simplify there is an analog element and a a digital element, both of which can be sub-divided into separate boards if not separate systems. The semis for the analog side are supplied by the traditional analog players notably Texas Instruments and Analog Devices. The digital side is also split into multiple pieces (and some of these can be virtualized, so this rabbit hole runs deep). Again, greatly oversimplifying there is a Distributed Unit (DU) and a Radio Unit (RU). The RU is effectively the last first point of contact between a signal from a phone and the network, usually sitting on top of a tower or outside a building, while the DU is a sort of switch that can ‘manage’ a few RUs. These systems handle networking functions, encoding and decoding wireless signals, in a role analogous to the baseband in the phone. And like basebands in phones they are critical for operation of the network as they determine total system capacity. (Feel free to put corrections in the comments, we can already see them flooding in.)

There are four companies supplying the chips for the RU and DU:

  • Intel
  • Marvell
  • Qualcomm
  • Picocomm (a UK based start-up).

Intel has worked for years to build its position in this market. They have one major customer today – Ericsson. Intel’s solution is essentially a mid-server range CPU. At MWC, Intel unveiled its long-awaited latest version which embeds some critical networking functionality into that CPU. Essentially a telecom-customized Xeon. We wrote about this product a few months back. Or more precisely we wrote about what it said about how Intel views reality. From what we could tell from the show there is not a lot of enthusiasm for this product. A full-blown CPU is overkill for most of these products. So even if it has better performance than some of the competition from a total cost of ownership perspective this is not likely to be the favored approach. And to be clear, it is not clear it has better performance. Intel’s reliance on Ericsson is also likely to be a concern, especially as Ericsson seeks to expand its nascent portfolio.

Marvell has in turn worked closely with Nokia, and others, and sometimes thier chip may come with someone else’s name on the cap. If anyone counted up market share for this market, Marvell is probably in the lead, but it is still very early days. We spent a lot of time in the Marvell booth talking about this chip, and we would sum up Marvell’s approach as Flexible. They are not necessarily going to build a custom chip for every customer, but they are not necessarily not going to. We have pointed out Marvell’s very solid positioning for the future of semi-custom compute, and this is just one more example of that.

Qualcomm also has a solid line-up for this market. We could never pin anyone down on this, but our guess is that Qualcomm’s chip for DU units is the smallest on the market and so likely the best performing. They unveiled their RU chip a while back, but MWC is the first real-world sighting of their DU chip. As part of the unveiling they announced a few partners/customers for the chip, including a working RU developed by Vietel, an operator in Vietnam, who actually had the unit on display in their booth.

Vietel DU reference design featuring Qualcomm silicon.

Qualcomm is fairly well positioned both in terms of their deep understanding of the space as well as what looks like a fairly solid product. That being said, we have a concern. Qualcomm has a pattern of dipping its toe in the market for infrastructure silicon. Every time a G rolls around, Qualcomm enters the market, supplying a chip (which they used to call CSM to differentiate from the MSM chips for phones), but then they lose interest and quietly exit the market and kill the product. Put simply, this market is too small for Qualcomm. They are really good at producing chips for high volume markets measured in hundreds of millions of units a year, but the infrastructure market is measured in hundreds of thousands of units. So the risk is that Qualcomm repeats this pattern and does not stick around. Admittedly, this time may be different. The demand for base stations, especially smaller units like RUs and ‘small cells’ is potentially much larger than the traditional base station market, and may approach a volume big enough to merit Qualcomm’s continued attention. On the other hand, Qualcomm is not as ‘flexible’ as companies like Marvell. Our sense is that Qualcomm would strongly prefer selling into a market which requires catalog parts, not the semi-custom approach that seems to prevail currently. For that to happen, we would need to see a stronger, more pervasive form of Open RAN take hold, and as we noted this does not seem likely.

Finally, there is Picocom a UK-based start-up with a founding team that has been in this market for a long time. This is a rough market for start-ups historically. A small number of very large customers, selling into a very conservative end-market can be punishing for a small company. Still, this is likely to be a big enough market to support multiple niches, many of which may be too small for the giants to pursue but big enough to provide a solid base for a start-up.

Overall, this looks like an interesting market. But it also is a good microcosm of broader trends in the semis industry. The growing importance of tailored solutions, the interplay between high volume general compute predicated on a common software standard versus a more custom market.

Leave a Reply