Software-defined datacenters provide a lot of flexibility and enable operators to quickly configure and scale their hardware infrastructure without having to rebuild it from scratch. But software-defined infrastructure is not the most efficient when it comes to performance and power. As a result, multiple special-purpose accelerators for modern datacenter workloads emerged in the recent years.
Among these accelerators are Mellanox’s BlueField-2 data processing units (DPUs) that have evolved from being just SmartNICs to fully-fledged processors that can offload networking, storage, and security workloads while adding manageability.
In addition, for GPU-accelerated AI and ML workloads that need fast and efficient networking connectivity, Nvidia offers its BlueField-2X card with Ampere GPU that promises an orders of magnitude performance increase when compared to traditional AI setups.
Nvidia strengthens datacenter efforts
The Nvidia Mellanox BlueField-2 DPU uses a custom-designed system-on-chip that integrates eight Arm Cortex-A72 cores and special-purpose VLIW acceleration engines to perform compute tasks that is combined with Mellanox’s ConnectX-6 DX 200Gbps Ethernet or Infiniband NIC for connectivity. The general-purpose Arm cores can run Linux operating system and offload a variety of workloads from host CPUs. The VLIW engines provide some additional acceleration.
Nvidia says that its Mellanox NVMe SNAP (Software-defined Network Accelerated Processing) technology can provide 2.5M+ IOPS read/write access for remote storage, which is line-rate performance at 100Gb/s with 4-KB block sizes. By contrast, an entry-level NVMe SSD can offer something like 300K IOPS with 4-KB blocks. In addition, BlueField-2 DPUs effortlessly add IPSec encryption and decryption capabilities at 100 Gb/s speed.
Meanwhile, the BlueField-2X card adds an Ampere GPU that offers 60 TOPS performance for AI. The key advantage of the BlueField-2X is that it uses hardware acceleration for remote direct memory access (RDMA/RoCE) communications that enables zero-copy data transfer between GPUs to maximize efficiency of compute clusters.
Nvidia’s plans include BlueField-3 DPU as well as BlueField-3X DPU card with a GPU for AI applications that are due in 2022. Furthermore, by 2023 the company intends to launch its BlueField-4X DPU that will finally combine DPU and GPU in the same chip.