Feature: System design
Equally, a GPU’s computational power can drastically
accelerate the algorithms needed in training AI and ML models, as well as reduce the time spent on running them. For example, inspection autofocus zoom cameras used in industrial inspection constantly adjust their focus to capture items on a production line. Products can vary in shape, colour and size – like apples, for example – so the GPU is programmed to extract information from these images, and the CPU running the AI program can then recognise any problems, such as insect damage, fungi or bruising, for example, and reject apples that don’t meet the specification. Using a frame grabber in a vision system reduces latency
between image capture and transfer, allowing real-time image processing, in some cases without requiring CPU overhead at all. A frame grabber captures individual frames from a video stream. In many applications frame grabbers are used in
combination with GPUs, where with various APIs that enable filter, convolution and matrix-vector operations, the GPU performs processing directly on data from the frame grabber, without involving the CPU or system buffers. This makes data acquisition very fast with very low latency as the GPU memory is made directly accessible to the frame grabber.
GPU API comparison In March this year, Nvidia introduced its next-generation GPU – the Hopper-based Nvidia H100 – made up of 80 billion transistors. It features Nvidia’s highly-scaleable NVLink interconnect for advancing large AI language models, deep recommender systems, genomics and complex digital twins. Along with its new Grace CPU (a Grace Hopper architecture named after the famous computer scientist), Nvidia is upping the processing power of thousands of systems.
Traffic within a CPU moves in a one-way, linear process, so when high volumes of
data such as 3D images are involved, bottlenecks occur
Also this year, AMD is launching its RDNA 3
GPU, a Radeon RX 7000 device. Made on a 5nm process node and with optimised graphics pipeline architecture, it promises up to 50% more performance per Watt than previous versions. The company is also working on its RDNA 4 Navi 4x
generation, for release in 2024. Both companies, Nvidia and AMD, offer their own GPU APIs.
DirectGMA DirectGMA (Direct Graphics Memory Access) is AMD’s proprietary method and API for low- latency peer-to-peer data transfers between PCI Express devices. The API gives access to part of the GPU memory to other devices on the bus, such as Active Silicon’s acquisition cards, for example. This allows the frame grabber to image data directly into GPU memory, with no CPU involvement at all and bypassing system memory completely, resulting in data transfer with minimal latency and saved memory bandwidth.
Active Silicon GPU without and with GPUDirect for video
www.electronicsworld.co.uk November 2022 37
Page 1 |
Page 2 |
Page 3 |
Page 4 |
Page 5 |
Page 6 |
Page 7 |
Page 8 |
Page 9 |
Page 10 |
Page 11 |
Page 12 |
Page 13 |
Page 14 |
Page 15 |
Page 16 |
Page 17 |
Page 18 |
Page 19 |
Page 20 |
Page 21 |
Page 22 |
Page 23 |
Page 24 |
Page 25 |
Page 26 |
Page 27 |
Page 28 |
Page 29 |
Page 30 |
Page 31 |
Page 32 |
Page 33 |
Page 34 |
Page 35 |
Page 36 |
Page 37 |
Page 38 |
Page 39 |
Page 40 |
Page 41 |
Page 42 |
Page 43 |
Page 44
