ARM, like many other large companies, are putting the focus next year on virtual reality. They see the future in mobile VR content, and want to be able to provide a definitive and visceral experience with their new Cortex A73 processor design and the Mali G71 GPU that goes along with it.
Cortex A73 and Mali G71 are posed to take mobile VR by storm
The new processor architecture is part of the “Sophia” family of CPUs, so named for the technology park in France in which it was designed. It’s a significant step forward in terms of performance, allowing for a much more efficient CPU overall. It’s still an OoO design, but it has some changes from the A72 that help to optimize how instructions are handled. The new design has a 2-wide decoder, down from a 3-wide setup that was introduced in the Cortex A72. The instruction fetch pipeline is 4-stages long with the decoder itself able to take care of most operations in a single-cycle, as opposed to up to three-cycles of the A72 generation.
Overall the entire new architecture is capable of dispatching and queuing far more operations at any given time compared to the previous generation. This means more work can be done per cycle, leading this to be far more efficient, in theory. They’ve optimized the pipeline to get things done more quickly. Interestingly, they’ve implemented a sort of “slot” mechanism, where each of those eight available slots (in the Cortex A73) are capable of completing all of their assigned tasks with every resource available. They can also talk to and retrieve information from other slots to help help stymie any redundancy.
The Branch Prediction of the Cortex A73 is state-of-the-art too. It has a much larger cache structure and a massive 64-bit micro Branch Target Address Cache as well. It’s huge and capable of far better prediction of the operations it’ll perform. This is likely but a part of why the whole design of the 2-wide architecture is more efficient than a wider design. Memory improvements, too, are included in the new design. A larger cache system overall has been added to help speed up internal operations as well as a full OoO dual-issue load and store area. This means it’s able to sustain ultra-high memory speeds for much longer.
Mali G71 GPU is the future, a very fast future
Moving on we see that they’ve released a new GPU design as well, the Mali G71. This boasts a 50% increase in performance while being nearly 20% more power efficient at the same time. This new GPU uses their newest Bifrost architecture. The Mali G71 has a very low latency, being able to start operations in as low as 4ms. The top-end will have 32 unified shader cores, which have been completely redesigned with a new core fabric and support for quad-parallel execution of instructions. The result is an architectural change into a more TLP-centric design that’s quite similar to the one that AMD did when the introduced Graphics Core Next in 20111. Each lane can have all kinds of different information to be computed, at the same time.
They’re also including support for OpenCL 2.0 to allow it to be used as a compute framework. Crucially, the new GPU does not fully support the Heterogeneous System Architecture standard, though the shared virtual memory and fine grained buffers move them in the direction of being a much more compliant heterogeneous compute platform. It’s efficiency and ability to complete a variety of different workloads quickly could see being useful with a variety of different mobile compute workloads. Perhaps even as a backbone for having a more complete and efficient deep learning framework on your mobile device itself. Of course it supports the latest iterations of OpenGL and Vulkan as well.
The end result of all the optimizations and changes within the architecture mean we’ll continue to see more efficient mobile GPUs that could very well exceed the performance of certain lower-tier laptop dedicated GPUs in certain workloads. The future of mobile CPUs and GPUs are both quite exciting. It’ll be fun to see just what kinds of performance increases we’ll see in the real-world and how that’ll translate into the overall user experience.