Nvidia Corp. early today announced a new accelerator card for servers and unveiled Selene, a supercomputer it recently added to its internal research cluster that ranks as the seventh most powerful system in the world.

The announcements were made during the virtual ISC Digital high-performance computing event. 

First, the server accelerator: It’s essentially a new version of Nvidia’s flagship A100 graphics processing unit (pictured) in a more compact package. Nvidia has placed a single A100 chip on a PCIe card that companies can plug into a server via the PCIe port to provide a speed boost for artificial intelligence applications, scientific simulations and other workloads that typically use GPUs.

Nvidia already has one accelerator based on the A100. That product, which debuted earlier this year, comes with either four or right onboard GPUs. The new single-GPU PCIe card announced today can be easily plugged into a firm’s existing servers thanks to the PCIe support, which will let Nvidia better target a key market segment: organizations looking to make use of existing infrastructure in GPU projects.

The  PCIe card (below) is mostly the same as far as core GPU specifications are concerned with two main differences. The product’s thermal design power, a measure of how much heat a chip gives off, is 38% lower, which allows it to run cool enough to be connected to a server with no need for a dedicated fan. To facilitate this increased efficiency, the PCIe card trades off some speed and provides 90% the relative performance of its bigger sibling as a result.

Servers that feature Nvidia’s A100 GPUs under the hood are poised to hit the market imminently. Nvidia said a dozen server makers, including Dell Technologies Inc. and Hewlett Packard Enterprise Inc., are planning to launch more than 50 A100-powered systems between them, 30 expected this summer.

Enterprise data centers aren’t the only environment in which Nvidia silicon can be found. The chipmaker’s GPUs are also offered by all three major public cloud providers and, not least, they’re widely used in supercomputers. At ISC Digital, Nvidia provided an update about its efforts on that front, revealing that 20 of the 25 most powerful supercomputers on the global TOP500 ranking use its graphics cards under the hood.

The list includes Selene, a supercomputer the chipmaker has built in-house for its internal research cluster. The system is the seventh most powerful in the world with performance equivalent to 27.5 quadrillion operations per second and, just as notably, is also one of the most power-efficient.

“Selene is the only top 100 system to crack the 20 gigaflops/watt barrier,” Nvidia senior product marketing manager Dion Harris detailed in a blog post.

Besides expanding Nvidia’s research cluster, Selene could also serve as a useful sales showcase for Nvidia’s DGX appliance series. The chipmaker says it has assembled the supercomputer in just four weeks by cobbling together DGX A100 appliances, which provide eight A100 GPUs per system. That speed could prove appealing for potential customers given how building a large supercomputer can take more than a year in some cases.

Mellanox, Nvidia’s networking subsidiary, was not absent from its ISC Digital presentation either. Nvidia acquired the network equipment maker last year for $6.9 billion partially because it has a major presence in the high-performance computing market, where its products power 305 of the TOP500 supercomputers. The chipmaker revealed at ISC Digital that its GPUs and Mellanox’s gear now power a combined two-thirds of the TOP500. 

Nvidia also used the occasion to unveil the Mellanox UFM Cyber-AI, a new network monitoring appliance. The system scans the traffic in a data center or supercomputer to identify potential component failures, performance issues and suspicious activity that may suggest a breach. 

Photos: Nvidia