Nvidia Corp. is making some big claims about the performance of its specialized BlueField-2 data processing units, saying they offer more than four times faster input/output operations per second than any other currently available chip of its kind.

That’s according to the results of Nvidia’s own benchmark tests. Published today, the findings show that a single Nvidia BlueField-2 DPU achieved an impressive 41.5 million IOPS using standard networking and open-source software.

Nvidia’s BlueField-2 DPUs debuted last year. They are highly specialized chips that come with various optimizations that enable them to handle certain types of infrastructure administration tasks more efficiently, including network traffic inspection. By offloading this work to the DPU, a server’s central processing units can focus solely on the compute tasks they have been given, thereby improving overall performance. Nvidia has previously claimed that a single BlueField-2 DPU can handle data center infrastructure administration tasks that would otherwise have to be performed by as many as 125 CPU cores.

To back up those claims, Nvidia staged a series of tests that involved connecting two fast Hewlett Packard Enterprise Co. Proliant DL380 Gen 10 Plus servers, one of which was used as an application server and one as the storage system. Each server was powered by two Intel “Ice Lake” Xeon Platinum 8380 central processing units clocked at 2.3 gigahertz per second, giving 160 hyperthreaded cores per server, along with 512 gigabytes of dynamic random-access memory, 120 megabytes of L3 cache and a PCIe Gen4 bus.

Each of those servers was then configured with two BlueField-2 P-series DPU cards, each with two 100 gigabit per second Ethernet network ports, resulting in four network ports and 400Gb/s wire bandwidth between initiator and target, connected back-to-back using Nvidia’s LinkX 100GbE Direct-Attach Copper passive cables. Both servers had Red Hat Enterprise Linux version 8.3 running on them.

The tests demonstrated that the BlueField-2 DPUs delivered “record-breaking performance” using standard networking protocols and open-source software, reaching more than 5 million 4KB IOPS, and from 7 million to more than 20 million 512B IOPS for NVMe over Fabrics, which are commonly used to access storage media.

In testing, the company added, BlueField supercharged performance as both an initiator and target, using different types of storage software libraries and different workloads to simulate real-world storage configurations. In short, Nvidia said, its tests show there are big advantages to using its DPUs, enabling higher performance and better efficiency for almost any kind of enterprise workload across both application servers and storage appliances.

Image: Nvidia