Two silicon acquisitions have highlighted the continued tumult in the semiconductor industry, this week, at opposite ends of the scale. MaxLinear has made a CPE play, buying Silicon Motion and its SSD expertise for some $3.8 billion, while Intel has bought a mobile-focused GPU specialist to help it brace against AMD and Nvidia in both the data center and new video experiences.
For MaxLinear, this is about being able to tick another box when it replies to equipment RFPs. To some extent, it is not an exciting deal. Meanwhile, Intel’s recent run hints at two possible readings – depending on how much faith you place in the new regime.
Either it is continuing to flail, throwing cash at its myriad problems as it continues to lose market share to AMD’s resurgence in CPUs, Nvidia’s continued dominance in GPU, and now the rise of ARM-based chips in the data center; or it has finally refocused the tens-of-billions-of-dollars that it has spent in the past decade trying to pivot from its Xeon-based dominance and then shore-up data center incursions from new architectures.
As so much web traffic consists of video, and these data centers power this entire process from beginning to end, developments in the landscape are always of interest. It has looked, for a long time now, that Intel was on the verge of a major upset, but the industry has been slow to embrace new silicon architectures.
Intel is on something of an acquisition spree, as its new CEO Pat Gelsinger starts righting the perceived wrongs of his predecessors. In February, Tower Semiconductor was bought for $5.4 billion, then March saw AI-based workload optimizer Granulate bought for $650 million, and now May sees Siru Innovation’s GPU expertise bought for an undisclosed fee.
Siru Innovations is a Finnish firm that can trace its heritage back to the 1990s – inside Bitboys, a desktop GPU frontrunner that eventually moved into mobile GPUs in the early 2000s.
ATI acquired it for $44.5 million in 2006, which in turn was purchased by AMD two months later. In 2009, AMD sold its mobile assets to Qualcomm for $65 million, and so the Bitboys founders left Qualcomm in 2011 to found Siru.
As you can see, the founders have been around the block, but now find themselves inside a giant that has inexplicably struggled to create a proper desktop GPU platform for years. The Intel Arc GPU portfolio is expected to be revealed shortly, which might finally allow it to challenge Nvidia and AMD, but like many of Intel’s AI-focused silicon projects, there is a pretty decent chance that Arc will be a complete dud.
Siru’s arrival is too late to alter the fate of the first generation of Arc products, but Intel did poach AMD’s GPU head Raja Koduri in 2017, who has since been heading up the GPU project – and who will have worked with many of the Siru staff previously.
The cynical view is questioning what Siru brings to the table, as it has been mobile-focused for so long, but the optimistic one is that Intel might finally break the GPU duopoly that has plagued the x86 industry for the past two decades.
For Faultline’s interest, the Arc product is arriving in a backwards fashion. The current Intel GPU roadmap targets laptops in Q1, desktops in Q2, and then workstations in Q3. Samsung is already selling some laptops featuring the discrete Arc chips, and the first desktop graphics cards are going to be sold to OEMs and system builders first, before being released to the public.
The workstation and server cards are what could prove transformative to video workloads, and this mobile GPU interest should see Intel expand into metaverse-y AR and VR devices.
Intel has spent billions on AI-focused R&D and acquisitions, which should also be brought to bear in the enterprise markets. One such example is Habana Labs, which was acquired for $2 billion to provide silicon counters to Nvidia’s GPU-based capabilities.
Two second-generation products from Habana have been unveiled this week, with the Gaudi2 and the Greco. Designed for data center deep learning and inference jobs, Intel has proudly proclaimed that they outperform Nvidia’s A100 platform.
However, the A100 is getting on for two years old, and so is due a refresh. The H100 is expected in Q3, which is going to undermine Intel’s current claims.
Silicon arms races are expected, of course, and the different software and architectures used in each workload influence performance far more than the raw hardware and firmware capabilities.
Intel’s aforementioned Tower Semiconductor deal saw it acquire a “leading foundry for analog semiconductor solutions,” which has focused on RF, power, and electronic design automation. This analog vein brings us to MaxLinear’s $3.8 billion acquisition of Silicon Motion, which will see the analog purveyor add Silicon Motion’s range of NAND flash controllers for solid state storage to its lineup.
MaxLinear supplies many of the broadband and networking equipment vendors, as well as a whole pile of CPE, and so the move into memory controllers is not as odd as it first looks. It is always surprising to see how little attention is paid to all the auxiliary components packed alongside a CPU or GPU on a device’s PCB, and that supporting-cast world is the one that MaxLinear lives wholly inside.
MaxLinear reckons the combined entity will have an enterprise value of some $8 billion, with annual revenues of over $2 billion, and a total market opportunity of some $15 billion. Both Qualcomm and MediaTek were reportedly sniffing around Silicon Motion, and MaxLinear appears to have had to pay a premium to beat them to the punch – with the offer set at around a 48% premium over the closing share price.
With annual revenues of $922.1 million, and operating margins in the 27% range, Silicon Motion should certainly not be a burden. The listed savings from combining the two are expected to hit $100 million annually, which sweetens the deal, but again, this looks like a lot of cash to pay. The Wall Street crowd seem to agree, with Silicon Motion’s shares jumping 17% after the announcement, but MaxLinear’s falling 22% – the largest drop since March 2020.
Of course, the stock markets as a whole are taking a beating currently. MaxLinear is betting that it can save cash by in-housing its supply of memory controllers. Solid state flash storage is now commonplace among networking equipment and consumer electronics, and so MaxLinear gets to offer another slice of the pie in its negotiations with OEMs.
Currently, Faultline readers are most likely to recognize MaxLinear from its work supplying WiFi chips to CPE vendors, with CommScope, Compal, and Skyworth among its known customers.
As for operators, we have a sneaking suspicion that MaxLinear has a chance of being deployed in Comcast’s Xfinity DOCSIS 4.0 gateways, with Broadcom another contender. We know that Quantenna’s QSR10G WiFi chipsets power the Xfinity DOCSIS 3.1 gateway, while the accompanying Plume WiFi pod extenders used by Comcast run Qualcomm 2×2 WiFi chips.
Back in April 2020, MaxLinear snaffled Intel’s home gateway business for some $150 million, which looked like a bargain that had the potential to double its quarterly revenues. Since then, DOCSIS 4.0 work has generated buzz, as well as a possible resurgence in MoCA and its continued threat in the G.hn sector – after it acquired Marvel’s G.hn portfolio for a paltry $21 million back in 2017.