Data processing unit

A data processing unit (DPU) is a channel controller, a programmable specialized electronic circuit with hardware acceleration of data processing for data-centric computing.[1][2] The data is transmitted to and from the component as multiplexed packets of information. A DPU generally contains a CPU, NIC and programmable data acceleration engines.[3][4][5] This allows DPUs to have the generality and the programmability of central processing units while being specialized to operate efficiently on networking packets, storage requests or analytics requests.[6][7][8]

The data acceleration engines differentiates itself from a CPU by a larger degree of parallelism (required to process many requests) and from a GPU by a MIMD architecture rather than an SIMD architecture (required as each request needs to make different decisions and follow a different path through the chip).[9][10] DPUs can be either ASIC-based, FPGA-based or SoC-based.[11] DPUs have been increasingly used in data centers and supercomputers since their introduction in the 2010s due to the rise in use of data-centric computing, big data, security, and artificial intelligence/machine learning/deep learning.[12] DPUs are designed to be independent infrastructure endpoints.[13]

DPU vendors & product lines include:

• AMD/Pensando: Capri, Elba & DSC[14][4]
• Broadcom: Stingray[15]
• Fungible: F & S Series[16]
• Intel: Infrastructure Processing Unit (IPU)[17]
• Kalray: Kalray K200-LP[18]
• Marvell Technology: OCTEON and ARMADA[19]
• Nvidia/Mellanox Technologies: Nvidia BlueField, ConnectX, Innova[20]

Software Vendors utilizing DPUs include:

• Bloombase
• Cloudflare
• DDN
• Fortinet
• Palo Alto Networks
• VAST Data
• VMware
• WekaIO