Nvidia claims it achieved a record input/output operations per second (IOPS) with its dedicated BlueField GPUs operating as SmartNICs. Thereby, the organization aims to demonstrate that its DPUs deliver cutting edge storage IOPS for data centers.
DPUs, also called SmartNICs, are essentially network cards equipped with a powerful processor. SmartNICs are often used in data centers to run specific workloads for firewalls and encryption, freeing up other processors to run more essential tasks for AI and analytics. AWS, Azure, Google and other hyperscalers often use SmartNICs in their data center environments.
Test configuration and execution.
Nvidia conducted a test using two 100 GB Ethernet BlueField GPUs in two separate HPE ProLiant DL380 Gen 10 Plus servers. Furthermore, the configuration consisted of Intel Ice Lake Xeon Platinum 8380 CPUs, 512GB DRAM, 120MB L3 cache (60MB per socket) and a PCIe Gen4 bus.
To start, the SmartNICs delivered a wired bandwidth of 4000 Gbps between the initiator and the target. Meanwhile, the Storage Performance Development Kit (SPDK), FIO test and Linux kernel were booted up.
The configuration yielded 41.5 million IOPS — up to four times more than any other DPU currently available, according to Nvidia. The record IOPS shows that Nvidia’s BlueField DPUs can deliver significant performance for application servers and storage appliances.
Furthermore, Nvidia says the record underscores BlueField’s capability of utilizing 4000 Gbps networks. This is particularly practical for AI applications that need data as fast as possible.
In addition, Nvidia states its Smart NICS are increasingly useful for standard workloads. VMware’s Project Monterey, a version of VMware’s hypervisor running on SmartNICs, promises to accelerate the latter with its release in 2022.