Different workloads require different architectures, and there are two classic ways to seek improved performance: scaling your architecture up and scaling it out.
A scaled up architecture seeks increased capability through additional or more performant resources. A scaled out architecture, by contrast, optimizes for parallelized capacity. In these architectures, many concurrent requests need to be turned around quickly, and the design optimizes for speed of many small jobs rather than size of one large job.
NVIDIA has historically been optimized for environments that prioritized scaling up, but as market factors combine to drive scale outwards the company is embracing this shift.
Mellanox and NVIDIA’s Push to Data Centers
NVIDIA’s raison d’être is to create high performance compute options, and the business model to accomplish that has evolved over time. They began by offering GPUs, or chips that offer more accelerated compute than standard CPU chips. With CUDA they expanded to a platform designed to help developers build on top of GPUs. With the recently announced plans to acquire the networking company Mellanox for $6.9B, NVIDIA is now pushing into the world of broader data center architecture.
On its face it may be surprising to see a company historically focused on accelerated compute make a networking play, but NVIDIA is making a strategic bet that the shift in application development trends towards containerized workloads will impact the way data centers are designed and constructed.
As my colleague Stephen O’Grady noted in his piece covering VMware’s Heptio acquisition in 2018:
From a macro perspective, then, whether a business is working with the largest of the large cloud providers, tiny nimble startups or more traditionally focused enterprise infrastructure suppliers, it is likely to be encountering Kubernetes – whether it likes it or not.
It is this market context that helps us understand the Mellanox acquisition.
According to NVIDIA CEO Jensen Huang, microservices have created an “exponential” amount of east-west traffic (or traffic within the data center) as containers communicate via API to other services within the data center; as application architectures become more distributed, the amount of inner-service communication required to run apps increases. (As an aside, this spike in network traffic is one of the primary justifications for service mesh projects like Istio.)
By comparison, north-south network traffic (or ingress/egress traffic into and out of the data center) has a slower growth trajectory. The trajectories have the potential to change the future of data centers design.
NVIDIA envisions a future where distributed applications will run across the data center, which in turn puts an increased emphasis on the importance of networking performance. The ability to more tightly integrate compute and networking can lead to large overall performance gains, and it is in this context that the acquisition of Mellanox helps NIVIDA augment their competitive positioning.
Evidence of the strategic importance of data centers is evident beyond their investment in Mellanox. NVIDIA has grown both their absolute dollar and percentage share of revenue related to their data center market segment. While the bulk of their revenue still comes sales related to the gaming segment, the business NVIDIA has brought in from data center sales has grown to be a significant factor in the company’s revenue mix.
NVIDIA’s intent to acquire Mellanox thus aligns with both development trends in the wider market along with their journey towards becoming a datacenter-centric company.
Related Posts:
- Stephen O’Grady, The Kubernetes World: VMware Acquires Heptio
- James Governor, Kubernetes won – so now what?
- Fintan Ryan, NVIDIA GTC 2018: Kubernetes, TensorFlow and as many GPUs as you can consume
Disclosures: Nvidia is a client and paid for my T&E to GTC.
No Comments