Powered by a 64-core Grace CPU and ConnectX-9 SuperNIC, the BlueField-4 DPU is ‘designed to power the operating system of AI factories’ and will deliver six times more compute compared to BlueField-3 when it debuts next year, according to Nvidia.
Nvidia said Tuesday that it plans to integrate its Grace CPU and ConnectX-9 SuperNIC into the next-generation BlueField-4 DPU to bring 800 Gbps of network throughput to future AI data centers for high-performance inferencing.
Designed to offload and accelerate networking, storage and security workloads from the server’s host CPU, BlueField-4 is set to debut in Nvidia’s Vera Rubin rack-scale platforms next year. The DPU, which was slated to be announced during Nvidia CEO Jensen Huang’s keynote at its GTC DC event, is expected to become available for other server platforms.
[Related: Nvidia Wants To Bring New BlueField DPUs To ‘Every Server’]
In a briefing with journalists and analysts on Monday, Dion Harris, a senior director at Nvidia, said BlueField-4 is “designed to power the operating system of AI factories” and will deliver six times more compute compared to BlueField-3, which became generally available in 2023 and offers 400 Gbps of network throughput.
Nvidia said BlueField DPUs have been “widely adopted” by AI infrastructure and cybersecurity companies thanks to their support of the product line’s underlying DOCA software framework. These vendors include Cisco Systems, DDN, Dell Technologies, Hewlett Packard Enterprise, IBM, Lenovo, Supermicro, Vast Data and Weka on the server and storage side.
The AI infrastructure giant also touted BlueField support from security vendors Armis, Check Point, Forescout, Palo Alto Networks and Trend Micro as well as cloud and AI providers such as Akamai, Crusoe, Lambda, Oracle Cloud Infrastructure, Together.ai and xAI. In addition, the DPU line has backing from cloud-native and infrastructure software platform providers Canonical, Mirantis, Nutanix, Rafay, Red Hat, Spectro Cloud and SUSE.
Among the BlueField supporters in the channel are global solution provider powerhouses Accenture, Deloitte and World Wide Technology, which are preparing to offer BlueField-4 to enterprises, governments and manufacturers, according to Nvidia.
BlueField-4 Combines Grace CPU And ConnextX-9 SuperNIC
BlueField-4’s major gain in compute performance is made possible by its 64-core Grace CPU, which relies on the server-grade Arm Neoverse V2 microarchitecture. The BlueField-3, on the other hand, features up to 16 CPU cores based on Arm’s Cortex-A78 microarchitecture that is typically marketed for smartphones and other mobile devices.
On the networking side, the BlueField-4 takes advantage of Nvidia’s upcoming ConnectX-9 Spectrum-X SuperNIC, which Harris said pushes “the boundaries of AI networking scale and functionality” by providing 1.6 Tbps per GPU of network throughput.
According to Harris, ConnectX-9 features the “most advanced” capabilities for remote direct memory access (RDMA) and 48 switch lanes with PCIe Gen 6 connectivity.
“This means we’re unlocking peak efficiency for AI workloads, from training trillion-parameter models to running massive inference jobs,” he said.
ConnectX-9 also comes with hardware crypto acceleration and advanced security features to provide security “at line speed,” Harris said.
The SuperNIC is also designed to be “highly extensible” thanks to “enhanced programmable I/O and comprehensive SDKs” that “give developers and partners the ability to tailor the infrastructure to unique AI use cases and future workloads,” he added.
Other BlueField-4 features include multi-tenant networking, rapid data access, AI runtime security, cloud elasticity and the BlueField Advanced Secure Trusted Resource architecture for enabling bare-metal compute instances with zero-trust tenant isolation.