There hardly seems to be a month that doesn’t go by without a different IT supplier being rumored to want to acquire network device provider Mellanox Technologies, which has an intense focus on the high end, scale out, low latency segment of the datacenter switching market.
Two years ago, in thinking about the members of the OpenPower consortium, we did a thought experiment about how if IBM really wanted to shake things up, it would buy GPU maker Nvidia, Xilinx, and Mellanox all in one fell swoop and create a real OpenPower company that could take on a dominant Intel and a resurging AMD head to head. Today, before the rumors of Intel reportedly wanting to pay $5.5 billion to $6 billion to acquire Mellanox surfaced, we suggested at the end of our coverage of AMD’s financial results for Q4 2018 that maybe it should buy Mellanox before someone else does, and last summer, we even suggested recently that Mellanox should buy supercomputer maker Cray, get it back into the networking game, keep the Cray name, and really go after HPC, hyperscale, and cloud markets with zeal and a much bigger revenue stream to allow it to take down exascale deals. (This was before Cray announced its “Slingshot” variant of 200 Gb/sec Ethernet tuned for HPC, which it unveiled last fall.)
We like to play Legos with the computer industry as much as anyone else does, just to see the effect of interesting combinations. A few months ago, it was FPGA maker Xilinx that was supposedly going to acquire Mellanox; we thought this would have been better as a full-on system play with the OpenPower collective being literally brought together as one. And then we heard it might be Microsoft who was chasing Mellanox, and this sort of made sense since Microsoft is one of the largest customers for both InfiniBand and Ethernet gear that Mellanox makes, but why buy the company when you can just offload all the engineering and risk to Mellanox? And finally today, as Mellanox was reporting its own financial results for the fourth quarter, there was a rumor in Globes, the Israeli business newspaper, about Intel being interested in Mellanox and paying a pretty substantial 30 percent premium over the market capitalization that Mellanox had before the Intel rumors hit.
Mellanox did not say one tiny peep about this Intel acquisition rumor on its call with Wall Street analysts today.
We don’t think many of these acquisition possibilities make a lot of sense, and the Intel one in particular is, to be honest, upsetting. Intel has not been a good home for the Fulcrum Microsystems Ethernet switch ASICs that blazed the trail to 10 Gb/sec networking a decade ago and then were never heard from again. You could argue that Cray would not have had to do its own Slingshot interconnect if the original plans to merge the InfiniBand interconnect from QLogic and the “Aries” XC interconnect from Cray – both of which Intel bought earlier in the decade – had panned out as hoped. Intel is still moving ahead with the Omni-Path 200 series, as we have reported, but this has not been an easy path forward and Cray decided to get back into interconnects rather than get fully behind Omni-Path 200, for whatever reason.
We are happy to see Cray get back into the interconnect game, since interconnects are central to high performance distributed computing – and we have similarly been happy to watch Mellanox grow its InfiniBand business over the past decade and expand, sometimes with a fit or a start, into the Ethernet space and compete for deals at the hyperscalers and cloud builders.
We think that there needs to be a balance of power in high performance networking and high performance compute and storage, and we would rather see Mellanox leveraged to counterbalance Intel’s hegemony in the datacenter than to extend it further and perhaps stifle competition instead of enhancing it. We surely do not want more consolidation in the HPC interconnect market. Hence our thought experiments about Mellanox doing a deal with IBM, Cray, or AMD, and hence we never once pondered a scenario where Intel would buy Mellanox. Intel has enough might and money to take the pieces it has and create great things. And it should by all means do that.
Having said all of that, Mellanox, which broke through the $1 billion revenue barrier in 2018, would be a good fit for any company that wanted to take on stalwarts Broadcom, Cisco Systems, or Juniper Networks in datacenter switching – with Cisco and Juniper really focusing on large enterprises – or upstarts like Barefoot Networks, Innovium, and Nephos, who are coming at Ethernet switching from the hyperscaler angle. Or even an upstart switch maker like Arista Networks, which leveraged merchant silicon and created its own Linux-based network operating system to ratchet up the competition in datacenter networking in the past decade.
Mellanox has been growing pretty steadily in the past two years after a hiccup in early 2016, and it has managed to be reasonably profitable in 2018 – something that was important since activist investor Starboard Value was pressing the company to cut costs to get profitable and to find a buyer.
In the fourth quarter, Mellanox grew revenues by 22.1 percent, to $290.1 million, and net income came in at $42.8 million, a big reversal from the $2.7 million loss it posted in the year-ago period. That net income is a very respectable 14.8 percent of revenue, which is not too bad for a hardware maker.
The hyperscalers and cloud builders, as we have reported elsewhere in Intel’s financial results for the fourth quarter, took a little bit of a breather after pouring enormous sums into cluster rollouts in the first three quarters of 2018, and Mellanox similarly was anticipating a slowdown in Ethernet switch and adapter and ASIC sales as these companies tapped the brakes. And, as Mellanox expected, Ethernet sales did hit an air pocket:
Sales across all Ethernet products fell by 15 percent sequentially, to $148.5 million. “Consistent with our expectations, Ethernet revenue declined on a sequential basis as several of our large hyperscale customers digest inventory they have built up during the prior quarters,” explained Eyal Waldman, chief executive officer at Mellanox. “While our visibility is limited, we expect revenue from hyperscale customers to grow progressively throughout the year and currently forecast 2019 revenue from hyperscalers will exceed 2018 revenue.”
Mellanox has been holding at around 70 percent share of the Ethernet adapter and adapter ASIC market for cards that provide 25 Gb/sec and faster ports to servers, which is about as good as it can get given the competition that Mellanox is seeing from Intel and Marvell for adapter cards and chips (for those who make their own cards or have ODMs do it). Interestingly, Waldman says that sales of Ethernet adapters that run at 100 Gb/sec speeds rose by a factor of 2.5X in 2018, and Ethernet switch revenues exited the fourth quarter at a $100 million annualized run rate and rose by 70 percent for all of 2018. Waldman added that Mellanox expected for Ethernet switch sales to accelerate even more in 2019, driven by adoption of its 200 Gb/sec and 400 Gb/sec Spectrum-2 Ethernet switches, which were unveiled back in July 2017 and which are starting to ship to customers now. Let’s say it will more than double, given the growth rate and despite the fact that the hyperscalers and cloud builders will be taking a breather probably into the second quarter of 2019.
During the fourth quarter, InfiniBand products in the aggregate – chips, adapters, cables, and switches – did not outsell Ethernet products, but sales were up nicely as 100 Gb/sec EDR Quantum InfiniBand hummed along and as 200 Gb/sec HDR Quantum-2 InfiniBand both picked up. Here is how we think the various speeds of InfiniBand sold each quarter since the Great Recession:
InfiniBand sales overall rose by 32.5 percent, to $135.2 million, and we think HDR InfiniBand ramped fast and that EDR InfiniBand, despite all of the benefits that the newer and we think better Quantum-2 devices offer, grew nicely just the same and turned in its best quarter. (These are our estimates above, based on trend data that Mellanox used to supply for sales of each generation of InfiniBand but no longer does.) The 200 Gb/sec HDR InfiniBand is tied very closely to exascale systems that will be coming in the next couple of years, although faster 400 Gb/sec InfiniBand may be available by the time exascale machines come to market.
“We expect InfiniBand revenue to grow in the upper single to lower double digit range in 2019, supported by the race to exascale as well as the increasing demands to analyze growing amounts of data, enabling advances in research and industry product development,” Waldman told Wall Street. “InfiniBand’s performance advantage and network computing capabilities make it the interconnect of choice for large scale, high performance computing, artificial intelligence, cloud, storage, and other datacenter deployments.”
Waldman also said that Mellanox did not see a lot of deals these days for Intel’s Omni-Path interconnect, a superset of InfiniBand that Intel has been pushing hard into the HPC and AI arenas in recent years against the Switch-IB2 and Quantum InfiniBand from Mellanox. You can see why Intel might want to buy the competition.