Dense geometry is expensive in storage, even before arriving on the GPU. With version 1.2 of its DGF SDK, AMD adds a layer of supercompression intended to reduce the size of assets, while retaining the possibility of exactly reconstructing the original DGF blocks.
DGF SuperCompression adds a more compact storage format
AMD has released version 1.2 of its DGF SDK with a new feature called DGF SuperCompressionor DGFS. The objective is to reduce the disk footprint of geometric data used by the Dense Geometry Format, its block compression format for dense meshes.
The Dense Geometry Format is thought to be an efficient format for future GPU architectures capable of directly consuming this type of data. At the same time, the SDK remains open source and can run on current GPUs via DirectX 12 and Vulkan, without requiring native DGF hardware support.
DGFS is not a format consumed directly by hardware. It serves as a more compact representation for DGF data, with the ability to exactly reconstruct the original DGF blocks, but also decode to classic vertex buffers and index buffers. This approach therefore allows us to maintain a fallback route for GPUs that do not support DGF hardware.
Announced gains and decoding times
In the examples published by AMD, DGFS takes up about 30% less than raw DGF. The Dragon model goes from 29.25 MB to 20.15 MB, while Statuette goes down from 40.99 MB to 29.31 MB.
With GDeflate applied on top, AMD says DGFS remains about 20% more compact than DGF. The figures communicated go up to 22.22% difference in favor of the new format on certain models.
Tests publiés sur Ryzen 9 7950X et Radeon RX 9070 XT
AMD also detailed decoding performance on a platform equipped with a Ryzen 9 7950X, 64 GB of DDR5-6000 and a Radeon RX 9070 XT. A Statuette model of 10 million triangles is decoded into meshlets in 0.15 seconds, while decoding DGF blocks takes 0.22 seconds.
These measurements were carried out on a single CPU core. AMD specifies that a GPU decoder remains possible, but this does not mean that RDNA 4 already has native DGF hardware support. The results above all show that software decoding is fast enough to be considered in asset streaming scenarios.
A Distinct Approach to NVIDIA RTX Mega Geometry
AMD DGF may recall certain issues addressed by NVIDIA RTX Mega Geometry around dense geometry and ray tracing, but the two approaches are not based on the same principle. DGF remains primarily a geometric compression format, while RTX Mega Geometry mainly targets the management and construction of clustered acceleration structures for ray tracing.
The interesting point is elsewhere: AMD retains a fallback route to conventional mesh data. For content engines and pipelines, this backward compatibility matters as much as the compression ratio, especially while no mainstream GPU yet exhibits native DGF hardware support.
Source : AMD GPUOpen, via VideoCardz
