Your browser is not supported. Please update it.

15 March 2019

Nvidia snatches Mellanox from Intel, targets continued AI leadership

From the outside the $6.9 billion cash deal announced this week, whereby Nvidia will acquire Mellanox, it looks like Intel, Xilinx and Microsoft were all chasing ownership of the data center interconnect business, so it called up its big brother over at Nvidia, who said, “Don’t worry, we’ll look after you,” and trumped all the other bids with all the cash it had in its locker.

The deal was ushered through mega-fast, with both boards approving a deal by Monday, that was just a rumor on Sunday. The deal will use up all but $500 million of the cash which Nvidia has in its bank, but already it is talking about a unified data center architecture, with processing power coming from Nvidia and interconnects from the ethernet and InfiniBand specialist Mellanox, perhaps sold alongside one another.

The cash is of course not an issue, since the Nvidia share price is already on the rise in the two days since the deal was done showing how much investors like this deal, and with a current market capitalization in excess of $100 million, Nvidia can go to the market any time it needs cash.

The deal certainly unites leaders in processing and interconnect for high performance computing, and a statement was issued that talks about their joint collaboration and innovation in the past and how the deal will be immediately accretive to earnings. It’s not as if Mellanox needed to be acquired at all.

However this is yet another one of those Riot M&A deals based around silicon, which tends to drive very high multiples, and resurrects the moribund M&A market from a dour 4th quarter 2018 (See Riot’s M&A paper this month).

Nvidia’s computing platform and Mellanox’s interconnects power over 250 of the world’s TOP500 supercomputers and have as customers every major cloud service provider and computer maker.

But almost as a Parthian shot, Intel was already talking to US journalists, briefing against the newly formed company, saying that Intel has Nvidia in its sights. Well, it has said that many times before and has yet to make a real dent.

This conversation was really about the coveted leadership that Nvidia has in AI circles. The way it was explained is to say that chips used which train AI models are currently driven more often than not by Nvidia. But the chips which run those pre-trained models are more often than not Intel Xeon’s, because running the models as opposed to building them, takes less compute power.

In the purchase of Nervana with its AI based ASICs, Intel hoped to change all that and begin to offer an integrated platform which shuts the door on Nvidia. Intel then also has its FPGAs which it acquired with Altera. Microsoft and Baidu use already Intel’s FPGAs for AI work.

So Intel wants to create a sandwich around Nvidia, with FPGA’s above and ASICS below, and the purchase of Mellanox (formed initially by refugees from Intel) might have made the trap complete. Which is perhaps why Mellanox did not want to be acquired by Intel (too tight a focus) and why Nvidia was happy to splash out to keep it from Intel’s grasp.

This won’t put off the AI fight between Intel and Nvidia, but it may continue to stack the cards more in Nvidia’s favor.

Nvidia also has a developer community that is completely familiar with the company’s CUDA programming model and related CUDA Deep Neural Network (cuDNN) software library, which makes this something of a software war too.

Intel also has deep learning software libraries optimized for its own CPUs, and now it plans to make them equally target Xeon CPUs, Nervana ASICs and even Nvidia GPUs. That way is only has to win the software war, and then once that is done it can wean developers off Nvidia chips. We’ll see.

Back to the datacenters of the future and Nvidia says these will be architected as giant compute engines with tens of thousands of compute nodes, designed holistically with their interconnects for optimal performance. Enter Mellanox.

When we came across Infiniband back at launch in 2001, we wondered where it sat exactly in the data center architecture, being at that time a 10 Gbps connection, in a Gigabit ethernet world, where the connections inside mainframes (yes they had those back then) were around the 200 Gbps speed. But today’s Infiniband tops out at 200 Gbps and it makes perfect sense as a connection medium which ignores whether or not the message is a transport between CPUs, or to different elements of a large processor or to high speed local or remote storage – it now seems it can handle all of them.

Broadcom will have been unable to acquire Mellanox, primarily because it has a massive overlap with one of its divisions, but it will be eyeing this deal with some suspicion and wishing it had not happened.

Mellanox pioneered the InfiniBand interconnect technology, which along with its high-speed Ethernet products, is now used in over half of the world’s fastest supercomputers and in many leading hyperscale datacenters. In January alone it landed one deal in Finland for a 200 Gigabit HDR InfiniBand system with the Finnish IT Center for Science new supercomputers and in the same month had a deal with the University of Stuttgart for a 5000-node supercomputer named “Hawk” to be built later this year, offering 24 petaFLOPs of compute performance.

Once the combination is complete, Nvidia says it will continue investing in talent from Israel, where Mellanox is based.

This is however Nvidia’s biggest acquisition to date, so let’s hope that it follows its “engineering led” instincts and is not tempted to massage its profitability, which means keeping the design teams separate, but in touch with one another – an alien concept at most of the really big silicon companies.

Mellanox revenues were just $1 billion last year.