Continuing this morning’s run of GTC-related bulletins, NVIDIA is providing yet one more replace on the state of their Data Processing Unit (DPU) undertaking. An initiative inherited from Mellanox as a part of that acquisition, NVIDIA and Mellanox have been speaking up their BlueField-2 DPUs for the higher a part of the final yr. And now the corporate is lastly nearing a launch date, with BlueField-2 DPUs sampling now, and set to ship in 2021.
Originally hatched by Mellanox earlier than the NVIDIA acquisition, the DPU was Mellanox’s thought for the next-generation of SmartNICs, combining their networking gear with a modestly highly effective Arm SoC to dump varied duties from the host system, comparable to software-defined networking and storage, in addition to devoted acceleration engines. Mellanox had been engaged on the undertaking for a while, and whereas the unique BlueField merchandise noticed a comparatively low-key launch final yr, the corporate has been laborious at work on Bluefield-2, which NVIDIA has since elevated to a a lot larger place.
This second technology of DPU-accelerated {hardware} will go below the BlueField-2 title, and the 2 firms have been speaking about it for a lot of the previous yr. Based on a customized SoC, the BlueField 2 SoC makes use of 8 Arm Cortex-A72 cores together with a pair of VLIW acceleration engines. All of that is then paired with a ConnectX-6 DX NIC for precise community connectivity. At a excessive degree, the DPU is meant to be the following step within the gradual motion in the direction of domain-specific accelerators throughout the datacenter, providing a extra specialised processor that may offload networking, storage, and safety workloads from the host CPU.
Coming off of their success within the datacenter market from broadening purposes for GPUs, it’s straightforward to see NVIDIA’s curiosity within the DPU undertaking: that is one other piece of silicon they’ll promote to server builders and datacenter operators, and additional undermines the significance of the one factor NVIDIA doesn’t have, a server-class CPU. So though not a undertaking said by NVIDIA, it’s a undertaking they’re full embracing and increasing upon.
As the majority of at present’s DPU announcement is a recap for NVIDIA, the precise product plans for BlueField-2 haven’t modified. NVIDIA can be releasing two DPU-equipped playing cards, the BlueField-2, and the BlueField-2X. The former is a extra conventional SmartNIC with the DPU and a couple of 100Gb/second Ethernet/InfiniBand ports. This permits it for use for networking in addition to storage duties like NVMe-over-Fabrics.
Meanwhile the bigger Bluefield-2X incorporates a DPU in addition to one in all NVIDIA’s Ampere-GPUs for additional acceleration through in-network computing, as NVIDIA likes to name it. NVIDIA hasn’t disclosed the GPU used on the BlueField-2X, but when these renders are correct, then the variety of reminiscence chips signifies it’s GA102, the identical chip going into NVIDIA’s high-end video playing cards. Which would make BlueField-2X a really potent card close to compute efficiency.
And NVIDIA’s plans don’t cease with the BlueField-2 merchandise. The firm has deliberate out a collection of playing cards primarily based on BlueField-2, which can be launched because the BlueField-Three and BlueField-Four household in successive years. BlueField-Three can be a souped-up model of BlueField-2, with separate DPU and DPU + GPU playing cards. Meanwhile BlueField-Four would be the first half the place NVIDIA’s affect makes it into the core silicon, with the corporate planning a single high-performance DPU that will be capable of considerably outperform the simpler discrete DPU + GPU designs. All instructed, NVIDIA is anticipating BlueField-Four to supply 400 TOPS of AI efficiency.
All of this, in flip, will include NVIDIA’s conventional embrace of each {hardware} and software program. The firm is seeking to mirror its CUDA technique with…
