What are Nvidia and AMD doing getting involved with SmartNICs and VMware?
A gentle primer on these chip giants' accelerators – and your essential guide to the conference so far
VMware Explore SmartNICs have long been the domain of hyperscale and cloud datacenters, but they remain relatively uncommon in enterprise environments due in large part to the lack of software compatibility. But with official support for AMD and Nvidia's data processing units (DPUs) now baked into VMware's popular hypervisor, that may not be the case for much longer.
At the VMware Explore conference this week, Nvidia and AMD announced support for running the vSphere 8 virtualization stack on their respective BlueField and Pensando DPUs. The announcement represents the culmination of more than two years of work under VMware's Project Monterey, which solicited Nvidia, AMD-Pensando, Intel, and others to offload vSphere's virtualization functions onto dedicated accelerators.
These accelerators, often called SmartNICs, DPUs, or infrastructure processing units (IPUs), typically combine high-speed networking and a combination of fixed-function ASICs or configurable FPGAs with general compute processors. In this regard, the DPU essentially functions as a coprocessor for I/O intensive workloads. The ultimate goal is for these accelerators to pick up work from a host server, handling those tasks – such as network packet filtering – at high speed and leaving the host's processor cores to run applications.
These devices have been deployed in public cloud and hyperscale datacenters for years. They isolate tenant workloads from networking, security, storage, and other infrastructure operations associated with running the datacenter, freeing up host CPU cycles in the process. VMware's Project Monterey sought to bring these capabilities to a more mainstream enterprise audience.
While VMware worked to bring its software stack to SmartNICs, Nvidia and AMD were on a buying spree, snapping up networking vendors that could give them an edge in what they saw as an emerging market.
In 2019, Nvidia announced the acquisition of Mellanox, which was in the midst of bringing its 200Gbit/sec BlueField-2 SmartNIC to market. Similarly, over the past two years, AMD has paid well over $50 billion to expand its data processing capabilities with the acquisition of FPGA vendor Xilinx and later SmartNIC startup Pensando. The chip house detailed its plans for melding FPGAs, ASICs, and Arm cores for its first in-house SmartNIC system-on-chip at the Hot Chips event last week.
Clearing the way for SmartNIC adoption
The software necessary to make use of these accelerators, however, has remained somewhat elusive, with many relying on highly proprietary and vendor-specific software stacks.
By extending support for vSphere to these devices, VMware wants to change that by enabling its customer base to achieve cloud-like efficiencies in their datacenters using their preferred virtualization stack. Meanwhile, for Nvidia and AMD, it's an opportunity to capitalize on years of research and development.
"We've seen an evolution where companies have moved from really bespoke appliances, with networking, security, storage, and management, to a software-defined datacenter," Kevin Deierling, Nvidia SVP of networking, said during a press briefing last week. "So with VMware, now with vSphere 8.0, we have actually moved that process – the infrastructure management, and storage, and software defined networking – all of that is now running on the BlueField."
It's a similar story for Pensando's Distributed Services Card, which AMD claims can improve efficiency by freeing up host CPU cycles and enhance security by isolating infrastructure services from tenant workloads.
"VMware vSphere Distributed Services Engine, enabled by AMD's Pensando DPUs, is a key step in bringing the industry closer to composable hardware systems," Forrest Norrod, SVP of AMD's Datacenter Solutions Group, boasted in a statement.
- AMD smartNICs to meld ASICs, FPGAs, Arm cores
- SmartNICs to make up 38% of network market by 2026
- Intel's smartNICs probably aren't for you (yet) says Intel
- Alibaba Cloud challenges AWS with its own custom smartNIC
Both vendors tout lower operating costs as one of the clearest benefits to deploying DPU-equipped systems. Nvidia's Deierling claims a total cost-of-ownership saving of $8,200 per server over its useful lifespan. "For an enterprise with 1,000 node servers, this improvement yields $1.8 million in savings over three years," he claimed.
This week's announcement of vSphere support also addresses one of the biggest challenges for customers looking to deploy SmartNICs in their datacenters: enterprise support. While SmartNICs have been available on the open market, they haven't necessarily been validated for use in OEM systems.
At VMWare Explore this week, AMD said it would begin offering its 400Gbit/sec-capable Pensando DPUs through leading server OEMs such as Dell, HPE, and Lenovo in the next few weeks. Likewise, Nvidia will begin shipping its 200Gbit/sec BlueField-2 DPUs in Dell PowerEdge systems beginning in November. Users can also test out the BlueField-2 as part of Nvidia's cloud LaunchPad service.
More to come
While Nvidia and AMD may be among the first to support VMware's vSphere software stack, they won't be the last, if current trends continue.
Intel has developed a number of FPGA and ASIC-based SmartNICs — or IPUs as it prefers to call them — and has also been heavily involved in Project Monterey from the start. The chip giant's upcoming Mount Evans and Oak Springs Canyon accelerators are slated to begin shipping "later this year," we're told.
Marvell also has its sights set on claiming a chunk of the DPU market with its Octeon 10 DPUs, announced last year. The cards feature an integrated terabit switch for high-speed network offload.
As these accelerators become more accessible, so is the software. Palo Alto Networks, for example, demoed its virtualized firewall functions, including intrusion detection and protection running on a BlueField-2 DPU last summer.
And, in June, the Linux Foundation, backed by a slew of DPU, OEM, and software vendors like Nvidia, Intel, Marvell, F5, Keysight, Dell Tech, and Red Hat, launched the Open Programmable Infrastructure project to accelerate development of smartNIC-enabled applications.
The question that the OPI project aims to solve is "how do we generalize, and how do we free up this technology so it will be consumable by everyone," Yan Fisher, Red Hat global evangelist for emerging technologies, told The Register at the time. There are now certainly enough players to find out. ®
Your cut-out-and-keep guide to our 2022 VMware Explore coverage
- News: VMware CEO bats away fears over Broadcom takeover
- News: Dell and VMware reveal updated VxRail HCI systems with support for network offloading to DPUs, and more
- News: VMware launches vSphere 8
- News: NetApp and VMware join forces in Amazon’s cloud and reveal work on NVMe-oF and NFS functionality
- Analysis: Amazon-style DPU offload, meet the VMware customer base
- Analysis: A closer look at vSAN and vSphere 8