Since Mellanox announced last year that it was for sale, there have been rumors of many potential buyers— Microsoft, Intel, Xilinx, AMD , and even IBM. Interestingly, the final winner, NVIDIA, was rarely mentioned, even though the companies have a long-standing partnership and common interests in high-performance computing. Let’s take a quick look at the three big reasons why this deal makes more sense.
1. HPC matters to both companies
The High-Performance Computing market is critical to both companies, both in terms of revenue and culture. NVIDIA and Mellanox have collaborated on HPC projects for many years, such as the Summit supercomputer at Oak Ridge National Labs, the fastest computer in the world (a joint effort with IBM). HPC needs fast networking, visualization, and AI, and these two companies are by far the leaders in those technologies; they both share a performance-centric culture. HPC has regained respect as a growth market, especially for NVIDIA. With their powers combined, they will be a strong contender for future HPC business, in the public and private sector.
2. Networking matters to NVIDIA
GPUs can deliver as much as 95% of the computation of supercomputers. Since compute nodes gain over 10X performance when deployed with NVIDIA GPUs, they need to be fed massive amounts of data to be kept busy. This is why IBM chose Mellanox as the interconnect technology for Summit, and why so many supercomputers are built with Mellanox and NVIDIA. As AI and HPC continue to converge, NVIDIA stands to gain more than its share of new installations. Last year, NVIDIA announced its own switching technology, called NVSwitch. Essentially, NVSwitch is a fabric of 16 interconnected GPUs that appear as one massive GPU, with shared data space, to solve problems that scale well. With Mellanox, NVIDIA can ramp that approach to even larger and more capable GPU clusters, using Mellanox’s RDMA to feed the beasts.
3. Both NVIDIA and Mellanox provide platforms
NVIDIA and Mellanox are more than just semiconductor companies; they deliver comprehensive platforms that add value beyond the advanced chips they sell. As platform companies, they have both become quite “sticky.” When Intel introduced the embedded OmniPath interconnect, some pundits predicted this would lead to the death of Mellanox. Instead, Mellanox demonstrated how difficult it is to displace. It provides libraries of software that can exploit its clients’ advanced HPC needs, tuning its interconnects to deliver high bandwidth and low latency for specific use cases. Likewise, NVIDIA delivers tremendous value through its software, such as CUDA, CuDNN, TensorRT, and most recently its RAPIDS library for Machine Learning acceleration. Again, these software tools will reduce the temptation to consider replacing NVIDIA with other products, such as Google GOOGL -0.98% TPU. As a combined company, they can leverage their platforms to each other’s advantage.
In short, I am not surprised these two companies decided they could benefit from each other. The deal also makes financial sense, since it will be immediately accretive to NVIDIA’s earnings when it closes. I’d also like to note that the timing is good for improving their combined ability to respond to the next era of Exascale supercomputing. This should be fun to watch!