“Future datacenters of all kinds will be built like high performance computers,” said Nvidia CEO Jensen Huang during a phone briefing on Monday after Nvidia revealed scooping up the high performance networking company Mellanox for $6.9 billion.
If that’s your vision – as it is Nvidia’s – that’s a darn good reason to buy one of the world’s leading HPC interconnect companies. Of course, there are other reasons too, playing keep-Mellanox-away-from-Intel for one. Another that might not have been top of everyone’s mind: getting a foothold in Israel — although if you watch the space closely, you know what a hotbed of technological innovation Israel is.
But the primary reason is Nvidia’s conception of the changing strategic role, and changing architecture, of the datacenter of the 2020s.
“Datacenters are the most important computers in the world today, and in the future – as the workloads continue to change triggered by artificial intelligence, machine learning, data analytics and data sciences – future datacenters of all kinds will be built like high performance computers,” said Huang.
“We believe that in future datacenters, the compute will not start and end at the server, but the compute will extend into the network. And the network itself, the fabric, will become part of the computing fabric. Long-term, I think we have the opportunity to create datacenter-scale computing architectures; short-term, Mellanox’s footprint in datacenters is quite large. […] We will be in position to address this large market opportunity much better,” he said.
Rumors are that the Mellanox bidding process, which kicked off last year under pressures from activist investor firm Starboard (which purchased a 10.7 percent stake in the company in November 2017 – see our coverage at the time), was highly competitive and that Intel was a top contender. Huang would not or could not confirm reports that Intel was in the mix. But he said the bidding was competitive owing to the interconnect – specifically the intelligent interconnect – becoming more important than ever, with more of the computing workload conducted on the interconnect fabric. This is the so-called offloading approach that is fundamental to Mellanox’s technology strategy.
Huang also praised Mellanox’s style of networking and their “extraordinary” software stack that has been “galvanized and integrated into a large body of work over a long period of time.”
In explicating the hot bidding process, he also pointed out the attractiveness of Israel, where Mellanox is headquartered. “It’s a technology center – the culture, the spirit of the people, the richness of technical excellence makes it a great place,” he said. “It’s one of the world’s great AI development centers and tech development centers.”
Taken together, with its high powered GPUs, its DGX machines, its development on NVLink and NVSwitch, and now an end-to-end interconnect portfolio, Nvidia is more than a de facto systems company, a point which Huang essentially conceded, although he said he prefers the term “datacenter scale computing company.”
“We were a GPU company and then we became a GPU systems company. We became a computing company which started from the chip up, now we are extending ourselves into a datacenter computing company,” he said.
Nvidia sees itself as a different style of company from other vertically integrated vendors. “We create the entire architecture, but we componentize it in such a way that we can partner with the entire IT ecosystem. We componentize our technology by thinking about scalability from one GPU to thousands of GPUs and in the future millions of GPUs. We think about it from a scalability perspective and from a compatibility perspective, so as a result we can offer our components to the entire IT industry, so that everybody can build computers and configurations that pleases them, and that solves their problems,” Huang said.
“Our business model won’t change even though we continue to expand into a systems sensibility; we’re not a systems company, if you will. We are really a systems architecture company,” Huang added, underscoring the distinction. “We open ourselves to partner with our customers and partners however they like. If they would like to bundle our DGX servers like NetApp and Pure and DDN, we’re delighted by that. If they would like to purchase our HGX motherboard, basically the systems board inside our DGX, to put into their cloud, like what Google has done and Microsoft has done and others have done, we’re delighted by that. If you would like to buy our chips and build your own systems, we’re delighted by that or if you would like to buy it in add-in card form. The thing we work hard is to make sure they are all compatible so the CUDA acceleration libraries work perfectly on top of all of them.”
There’s a reason for the caveating and careful language, the hedging with regard to Nvidia committing to the systems business – and it’s a point that Addison Snell, CEO of Intersect360 Research shared with us. “Nvidia does need to be careful that its aspirations do not threaten their important partners on the server side. While Nvidia can sell components and reference architectures directly to hyperscale companies and ODMs, they shouldn’t take the traditional OEM channel for granted.”
Really the only missing element from a complete systems stack at Nvidia is the general purpose CPU. Nvidia had intentions to get into the CPU game reaching back to 2011 with Project Denver, but discarded those plans. With Arm momentum high and Arm looking for partners to make its Neoverse chips, many are thinking it’s a natural next step for Nvidia.
But Huang, when asked about this, seemed less than excited by CPUs – as a market opportunity, that is. He repeated his stance that the company believes in serial processing (acknowledging “you cannot parallelize everything”), but affirmed that while Nvidia is happy working closely with all the major CPU companies – IBM, AMD and yes Intel (“people think we are antagonistic, but it’s just not true”) as well as Cavium/Marvell, Ampere and Broadcom on the Arm side – it sees more benefit to focusing its R&D on engineering capable of delivering “x factors of improvement.”
“I think if we were to pour a bunch of R&D into building CPUs today, the x factor you get, after five years, is about 15 percent — and [the CPU makers] battle over 15 percent!” said Huang. “And since they’re doing such a good job battling over 15 percent we could invest our R&D on the areas that get big x factors – the places where people aren’t investing. And I think that accelerated computing was severely underinvested, and so our return there has been fantastic and it’s going to continue to be fantastic and getting better.”
Getting back to the Mellanox deal, it’s largely held as a strategic buy by market-watchers.
“Nvidia at $10 billion capex versus Mellanox at $1 billion suggests that Nvidia sells many more customers, has a more extensive sales/distribution channel and a global service infrastructure,” said one long-time industry observer we spoke with. “Via economies of scale, by simply dropping the Mellanox product line into the Nvidia portfolio should drive the Mellanox line of business significantly. This could be the principle reason they paid nearly $7 billion. They looked at Mellanox’s leading technology, factored in their customer base and channels, and estimated the sales opportunity over a period of time, typically three-to-five years. Nvidia must feel they can drive the Mellanox business unit to $7 billion in three-to-five years, as well as enhance their core products.”
Back on the technology side, Snell noted that “the combination of Nvidia and Mellanox is intriguing for HPC and AI. The companies are respective leaders in their high-performance technology areas, and each has ridden important growth trends. Nvidia’s purchase of Mellanox not only locks in potential synergies in advancing high-bandwidth connections for accelerated systems, but also protects the technologies from Intel, which is a natural competitor to both companies.”