Microsoft has introduced its brand new preview of the Agility SDK which adds support for GPU Work Graphs, Wave Matrix, and AV1 across various GPU vendors.
Microsoft's Agility SDK Gets Work Graphs, Wave Matrix & AV1 Support Across AMD, NVIDIA, Intel & Qualcomm GPUs
Today, Microsoft has announced its latest Agility SDK 1.711.3-preview which brings with it three brand new features which include Work Graphs, Wave Matrix, and AV1 support. Work Graphs are one of the major features that will allow the GPU to work autonomously & address the limitations associated with general compute workloads. Epic Games will also be utilizing Work Graphs for its Unreal Engine 5 since features such as Nanite and Lumen are already hitting the limits of current compute shader paradigm and Work Graphs can not only optimize them but also unlock various capabilities in the future.
Epic Games has been searching and advocating for a better solution to the GPU generated work problem for a while now. UE5 rendering features such as Nanite and Lumen are hitting the limits of the current compute shader paradigm of chains of separate dispatches issued by the CPU.
Work graphs directly address that problem in a way that not only allows us to do things we previously could not but also enables us to do them in ways that should be far easier to write. We have already started exploring how we can optimize our current features with work graphs and are excited about what possibilities they could unlock in the future.
— Brian Karis, Epic Games
Work Graphs
Work graphs introduce a powerful new form of GPU autonomy.
This first release of work graphs allows compute shaders to request other compute shaders to run asynchronously, for tasks such as culling, binning, or chaining of compute work. These work requests can include a data payload if desired, managed by the system.
Hardware has the flexibility to schedule work efficiently without the developer needing to understand the specifics of every device. The programming model is easier to use and more flexible in many ways than the existing serial ExecuteIndirect model for GPU work generation.
In cases where work graphs aren’t a full replacement for ExecuteIndirect yet, like launching the rasterizer, they can still set up ExecuteIndirect buffers as needed. New capabilities can be added to work graphs over time by Microsoft. AMD's GPUOpen blog post on Work Graphs expalins that this tech could:
- … enable more direct methods of solving complex problems.
- … reduce memory constraints and improve cache utilization.
- … simplify inter-pass dependencies and barrier-induced complexity.
- … improve GPU thread saturation.
Wave Matrix
GPUs and compute devices have begun adding dedicated silicon to their hardware to support matrix multiplication at higher bandwidths for usage in machine learning and imaging applications. To allow access to this dedicated silicon, HLSL is adding Wave Matrix instructions to the language, also known as Wave Matrix Multiply Accumulate (WaveMMA). This addition defines several new abstract Wave Matrix data types, which allows the underlying hardware to store, rearrange, and duplicate data across all threads in a wave. Here are the specs.
Microsoft's Wave Matrix driver support:
- AMD: AMD will be releasing a preview AMD Software: Adrenalin Edition driver with WaveMMA support in the coming weeks, watch here for an updated link to that driver.
- Intel: For Intel drivers, please contact your developer representative
- NVIDIA: For NVIDIA drivers, please reach out to your developer engagement representative
- Qualcomm: Future support is planned
AV1 Encoding
AV1 is a new codec that is gaining adoption across the industry by providing higher compression performance while maintaining the same quality as other available codecs.
In recent years dedicated hardware for encoding video using AV1 has become available. By extending the existing D3D12 Video Encode API that was previously announced, Microsoft is able to provide a unified interface to access the AV1 video encode hardware. Here are the specs.
AV1 Encoding driver support:
- AMD: An AMD Software: Adrenalin Edition driver with support for AV1 Encode on AMD Radeon RX 7000 Series graphics GPUs will be available estimated Q4 of 2023
- Intel: For Intel drivers, please contact your developer representative
- NVIDIA: RTX 40 series GPUs, driver version 545.31+
- Qualcomm: Future support is planned
Refference- https://wccftech.com
0 Comments