Marvell And Cavium Forge A Datacenter Platform
November 21, 2017 Timothy Prickett Morgan
It has taken what seems like forever, but Arm server processors are starting to get some legs just as a massive consolidation wave, driven as much by the end of Moore’s Law as by the desire to always be bigger, is undertaking the semiconductor industry. All we need is a recession and a price war in the datacenter and a lot of compute, storage, and networking incumbents could be toppled. It wouldn’t be the first time, and it will not be the last.
This is why semiconductor giant Broadcom wants to pay a stunning $130 billion to acquire sometime rival Qualcomm in a not-yet-hostile but unsolicited takeover deal, and is also why Marvell is spending $6 billion to buy Cavium, which doesn’t compete very much with Marvell but which is making a lot of noise these days with its ThunderX2 Arm server processor and which has a steady business selling MIPS-based Octeon III multicore chips for networking and storage applications. The Marvell-Cavium deal makes good sense, and Matt Murphy, president and CEO at Marvell, and Syed Ali, Cavium’s co-founder and CEO, did not find might argument here among Wall Street analysts analyzing the deal earlier this week. This being Wall Street, they might be hoping for more upside that they can play, particularly with Cavium’s stock rising ahead of the deal on rumors and on the Broadcom-Qualcomm deal announced a week ago.
While a higher bidder is always possible, this is the deal that the companies have worked many months to do, according to Murphy, and it does, as many have pointed out, have “strategic logic,” a nice play on words for two semiconductor companies.
Might makes right in semiconductors because fabs cost tens of billions of dollars and chip development for those fabs costs tens to hundreds of millions of dollars, depending on the complexity of the devices and the yield curves of the fabs. Intel, which used to be the world’s biggest chip manufacturer as ranked by revenue, has been usurped by Samsung due to supply shortages and price spikes for DRAM memory and flash memory across all kinds of devices, including servers and networking and storage gear. Up until now, Intel has been the process leader, but the company has had woes with its 14 nanometer processors, and will do another tweak (at least) past the current “Skylake” Xeon SP processors that were launched in July. Qualcomm has not only beat Intel to market with a 10 nanometer chip with its “Amberwing” Centriq 2400 processor, but it has delivered a chip that can stand toe-to-toe – or rather socket to socket – with the Xeons in the datacenter running cloudy workloads. And, it looks like Cavium, thanks to its acquisition of the “Vulcan” Arm server processors, now given the ThunderX2 label and originally developed by Broadcom and sold off after it was acquired by Avago Technologies.
Intel may be regretting letting its XScale ARM processors be sold off to Marvell back in the summer of 2006, which was the foundation of that company’s Armada family of embedded processors, just like Broadcom/Avago might regret selling off the Vulcan ARM chips. (We will be drilling down into these separately, with details coming out at the recent SC17 supercomputing conference.) While there were a bunch of prototype systems based on 32-bit Armada processors back in the early days of Arm servers – Dell had one that it was peddling, notably – Marvell did not catch the server bug as hard as Calxeda, Applied Micro, Samsung, AMD, Cavium, Broadcom (the original one), and finally Qualcomm, the last one to the party. Calxeda is gone, Applied Micro and its X-Gene family of chips has been passed around and has landed in the hands of private equity firm The Carlyle Group (to what end we are not sure), and Samsung and AMD backed out of Arm server chips.
With the relative success of the ThunderX processor, which was really a development platform and which was not really widely deployed, that gave Cavium a foundation on which to extend to its own ThunderX2 designs, which are aimed at highly threaded cloud applications and the impetus to snap up the Vulcan intellectual property and expertise as soon as Avago lost patience with the whole thing. (We think it will regret that decision in the long run.) The Vulcan chips have had the ThunderX2 brand slapped on them and were the start of the SC17 conference, with systems available from a slew of vendors and venerable Cray putting ThunderX2 nodes officially beside Intel Xeon and Xeon Phi processors in the XC50 product line and linking them with its “Aries” XC interconnect. Cavium has yet to put out its original ThunderX2, which has a different balance of compute, memory, I/O and which also packs all kinds of accelerators for commercial workloads, to further flesh out the Thunder line of chips. But we have been assured by the server top brass at Cavium that this is indeed the plan and that Cavium has no intention of being a one-trick server pony, whether or not it is acquired by Marvell or anyone else.
According to Murphy, the timing of the Broadcom-Qualcomm and Marvell-Cavium deals are coincidental, and while this is true, some kind of matchup between various semiconductor players trying to take on Intel’s very profitable position as hegemon of datacenter compute was inevitable. Nvidia is taking a good-sized bite out of compute with its Tesla GPU accelerators, AMD is hanging in there with its Epyc X86 chips, which were all over SC17, too, and IBM is finally starting to get its “Nimbus” Power9 chips out the door and into HPC systems with more to follow across technical and commercial computing. There has never been a better time for these companies to try to break into the datacenter, with Intel having difficulties with advancing its manufacturing process to the 10 nanometer node at the same time that Intel is charging a hefty premium for its Skylake Xeons. Xeons are still the easy choice, and safest bet, and the “Purley” platform that uses Skylake Xeons does have lots of benefits that can justify some of that premium, and for some customers who get good discounts, all of it perhaps.
Whether or not all of these events coming together is coincidence or cross-reactionary is not the point, even though it is intellectually interesting to ponder it. What can be said is we have not seen such a strong desire for alternatives since the dot-com bust more than a decade and a half ago. And both Marvell and Cavium know this and see the opportunity ahead of them. It is going to take a lot of money, time, energy, and heart to take on Intel, which is the toughest and most ambitious competitor there is. But between AMD, Cavium, Qualcomm, and IBM, there is a chance to chip away at the King of Compute.
Cavium would have had a much tougher slog going this alone, and Marvell is now the beneficiary of the significant amount of internal and acquired investment in ThunderX and Vulcan Arm server processors, XPliant Ethernet switch ASICs, and QLogic host bus adapters. Marvell has a large networking and storage chip business in its own right, but it is dependent on chips it sells for disk drives and notebooks, and with Cavium, which mostly sells compute and thus far mostly its Octeon III MIPS-based network processors, it is less dependent on these markets. And importantly, because there is not a lot of overlap between the two company’s customer bases and addressable markets, this is actually a good fit that doubles the total addressable market and will quite possibly see the embiggened Marvell balance out at a third networking, a third storage, and a third compute over the near term if the ThunderX2 lines take off.
In a conference call going over the deal, Murphy said that the combined companies were only pulling about 10 percent of their $3.4 billion in annual revenues from the cloud and datacenter markets, but that this part of the business was growing fast, particularly for Cavium. Murphy did not elaborate on how much money and what kind of growth was involved here. But he did say that prior to doing this deal, the core Marvell networking and storage businesses were growing at 6 percent and it was modeling for 5 percent growth, and with Cavium it should be able to grow at the high single digits. That is as good as Intel is doing in the datacenter these days. Perhaps the new Marvell can do better, and perhaps Qualcomm will shop around for its own networking and storage chips to flesh out its datacenter portfolio.
There is always safety in building a platform so long as you can get the customer to buy the whole thing, which is what Intel has really done so brilliantly with the Xeon chips and what system makers from days gone by did. Intel has showed that component suppliers can control a platform while letting it grow in a more flexible way than actually being a system vendor itself, and this seems to be the idea with the combination of Marvell and Cavium, at least as it relates to the datacenter.
Marvell’s Prestera line of Ethernet ASICs are used by service providers and small businesses to get 10 Gb/sec and slower networks, and Cavium’s XPliant programmable switch ASICs (which came to it by virtue of an acquisition) scale from 10 Gb/sec all the way up to 100 Gb/sec for leaf and spine architectures for massive datacenters. Cavium has been mixing its processors, switch chips, and various accelerators for SDN and security into a platform called CloudScale Rack, and this shows the aspiration that Cavium has to be a full platform provider for the datacenter, much as Intel does already with its Xeon processors, server chipsets, 3D XPoint and flash memories, Ethernet and Omni-Path network interface cards, and Omni-Path switches. (Why Intel has sat on its Fulcrum Microsystems Ethernet switch ASICs, acquired six long years ago, is a mystery. It picked QLogic InfiniBand over Ethernet and Aries, both if which it also owns.)
The one thing that the Marvell-Cavium combination doesn’t have is main and flash memory, and with a market capitalization of $55 billion, they are not about to buy Micron Technology, which had $20.2 billion in sales in its fiscal 2017. Maybe Micron, which needs a compute and networking platform, should buy Marvell/Cavium? This seems like a good fit, excepting that very close relationship with Intel on flash and 3D XPoint memory. But if you want to build a platform, this might be what it will take.
The other thing it takes is shared and focused investment, and this is something that Ali called out in talking about the deal, using the 10 nanometer and 7 nanometer process nodes for chips as an example. With the two being separate, both Marvell and Cavium would have to spend about the same amount of dough perfecting their designs and learning these processes from their fab partners and driving volumes to improve yields.
There are other synergies, too. Cavium will be benefitting from its MIPS products for many years to come because design wins at telcos and service providers result in products that are sold for eight or nine years before they are upgraded. So it will be able to milk that Octeon III chip line for many years to come as it ramps up its Arm server chip efforts. Moreover, the cores, caches, memory controllers, and accelerators used in the original ThunderX2 and Vulcan ThunderX2 chips can be redeployed in other products, like Armada embedded chips, the kind of trickle down from on high we believe strongly in here at The Next Platform. This is how the new Marvell will get more than a 20 percent slice of its current $16 billion total addressable market.
Now won’t it be funny if Broadcom prevails in its deal to acquire Qualcomm, and then spins out the Centriq line for Marvell to catch it? This doesn’t seem likely, but so many things that happen are not.
Of the $6 billion for the deal, half of that is coming from 170 million new shares of Marvell stock that will be given to Cavium shareholders. Marvell will be lining up $1.75 billion in new debt financing and doing the rest from the $1 billion in cash it has on hand and probably tapping some of its $500 million revolving credit facility. The deal is expected to close some time in middle of next year, provided the global antitrust regulators give it the nod and the US doesn’t make a fuss about Marvell being based in the Bermuda tax haven.