Following a fantastic first event in Shanghai it was off to the airport for a short hop to a much chillier Beijing for round two of 2016’s Tech Symposia. On Monday we talked about the keynotes and major product announcements so today we’re taking you to the beautiful Beijing Sheraton ballroom where our technical experts took us through a deeper dive into a huge range of products and processes.
Split into three streams, the Tech Talks covered Next generation processing, Smart embedded & IoT and Intelligent implementation and infrastructure. Focussing on the first stream, first up were ARM Senior Product Managers Dan Wilson and Roger Barker for a closer look at their new products launched on Monday. Having covered the market drivers such as Virtual Spaces and the new Vulkan graphics API in the keynotes, Dan used his session to consider Mali-G51’s Bifrost GPU architecture in greater detail.
The first Bifrost based Mali GPU, Mali-G71 is a high performance product designed for premium devices. In order to facilitate quality graphics on mainstream devices based on Mali-G51, specific architectural optimizations were made to rebalance workloads and prioritise graphics processes. A new shader core design allows partners to choose a flexible implementation of single or dual pixel shader cores, with single pixel cores able to handle one texel per cycle and dual pixel cores handling two texels per cycle. Partners can also choose to implement an asymmetric combination of the two for an MP3 configuration. Not only are there changes to the available shader cores but specific optimizations were also made to the texture and varying units. The texture unit changes have been designed for increased tolerance to high memory system latency while effectively reducing the pipeline length to reduce silicon die area and power consumption. As Mali-G51 based devices start to appear in consumers’ hands in 2018 it will be great to see how our partners have leveraged these developments to provide a greater user experience for mainstream devices.
Following Dan, Roger took to the stage to provide us with more detail about the new video processor, Mali-V61. He highlighted the evolution of video from live stream TV broadcast through on-demand streaming right up to today’s newest use cases where an ever increasing number of users are communicating in live, real-time video. Featuring all-new, super high quality VP9 encode and decode and much improved HEVC encode, the Mali-V61 VPU again provides better than ever configurability and choice to partners. Roger explained that the reasoning behind the importance of this high quality encode was around the emergence of time critical video applications that can’t support the additional processing time required to transcode content from different codecs in the cloud before delivering a VP9 decode, for example. With VP9 encode you can upload your video content in the same format in which it will be decoded, removing the transcoding lag and facilitating the fastest possible video experience. In terms of flexibility, we were given a more rounded view of the options available across mutlicore scaling. With an 8 core configuration partners can choose to support one 4K stream at 120 frames per second (FPS), or alternatively multiple streams at a variety of quality and FPS performance points from 720p right up to 4K. These are of course just a taste of the technical elements so you can visit the Connected Community to read the launch blogs for Mali-V61 and Mali-G51 for all the details.
Another great multimedia session was hosted by our in-house graphics guru Sylwester Bala and boasted the rather poetic title of ‘Virtual Reality: Hundreds of millions of pixels in front of your eyes.’ Sylwester took us through the evolution of mobile gaming from the earliest Gameboy platform right up to the mobile virtual reality applications appearing today. The graphics complexity has increased throughout this timeline but virtual reality has accelerated this trend even further. For starters, we have to render a marginally different view for each eye in order to create a viewpoint that our brain understands, effectively doubling the graphics workload. The lenses within the head mounted display are needed to correct the distance perception in the field of view but have the additional effect of making the final image appear sunken in the middle, known as a pin cushion effect. Barrel distortion therefore has to be applied in post processing to correct this effect, adding another level of processing complexity. Sylwester discussed the factors which often limit the quality of a VR experience such as latency, double CPU and GPU processing (required by the two separate views) and resolution. One of the solutions to reducing graphics workload is the use of foveated rendering. This approach effectively splits the display into two concentric circles to mimic the fovea region of your eye. The central section is rendered in high resolution but the outer section in lower resolution. The reduced quality of the outer section isn’t perceptible to the viewer but greatly reduces the processing power required. It does however, mean that instead of rendering twice, for two eyes, we are rendering four times, for four different sections. Sylwester explained how our Multiview extension can render both views with the required variations, simultaneously. Read more about Multiview in this blog. The really exciting result of the use of Multiview is the impact it has on GPU performance. Graphs demonstrated that GPU utilization could be taken down to from near 100% (for 1024 x 1024 resolution) to 40% for 512 x 512 inset resolution. Further savings can be achieved by reducing the size of the inset section depending on the specific of the device and its optics. It also demonstrated up to 50% bandwidth savings. Further efficiencies could be achieved through the use of ARM Frame Buffer Compression for bandwidth heavy processes like Timewarp, in some cases up to 50%.
There were so many fantastic sessions that I couldn’t possibly hope to cover them all but I hope this has given you at least a glimpse of the vast breadth of knowledge we’re so lucky to be able to share with our partners. From here the Tech Symposia tour moves to Shenzhen on 4th November then on to Taiwan, Korea, Japan and finally, India before finishing up for another year. Don’t hesitate to get in touch if there’s something in particular you’d like to hear more about!