At the GPU Technology Conference, the CEO of NVIDIA presented a new family of processors under the brand BlueField-2 which were dubbed DPU, to which the developers had access through an SDK called DOCA, which is the equivalent for the DPU to what is CUDA for GPUs.
The NVIDIA DPU
In fact, the NVIDIA DPU is nothing more than a new name for Mellanox SmartNIC, remember that NVIDIA bought this company for almost $ 7 billion.
BlueField-2 DPUs are nothing more than programmable network and storage interfaces that relieve the CPU workload when handling these functions.
This means that applications perform better and don’t have to worry about repetitive tasks based on distributing and storing data over a network.
What is NVIDIA’s interest in entering the data center market?
In its simplest definition, a DPU is nothing more than what we colloquially call SmartNIC, that is, an intelligent network controller responsible for managing the movement of data and its access in a system.
In cloud computing, virtualization means that it is the processor that has to completely manage the infrastructure by software, there is no supporting hardware for this. Which means that the processor, most of the time an X86, has to take care of everything, not only running the operating system and applications, but also network traffic, storage management, security and more.
In other words, everything is software and the applications compete with the software in charge of the movement of data to obtain the resources of the processor, hence the need for a SmartNIC which frees the processor from these tasks and allows it to concentrate. its resources not on the processor. data movement but when running applications.
How does this affect the future of graphics cards in our PCs?
Current GPUs have an SoC-like infrastructure, but with the implementation of chip-based configurations, these will evolve into a NoC-like configuration, which is based on a network infrastructure to communicate the various elements where a SmartNIC is. in charge of communication management. between the different elements.
It is at this stage that a DPU will be necessary to be in charge of effectively managing the movements of data under the new paradigm. In other words, the instructions for moving data in the GPU would not work as shader programs in the compute units but would be performed by calling the DPU, and the DPU will be used to manage both the RAM and the accesses to the computer. SSD. .
NVIDIA CEO Jensen Huang showed on his DPU roadmap where the BlueField-4 stands out, which will release in 2023 and will include a GPU inside, or would it be the other way around? In any case, given the timing of the launch of the different GeForce architectures, BlueField-4 may be NVIDIA Hopper.
The reality is that 2023 is a very important year for NVIDIA, as that’s the date they’ve marked to leave out the traditional rasterization-oriented GPU. Is DPU and GPU integration part of this roadmap? This is very likely, because for ray tracing you are working with extremely complex databases.