A100 removes the raytracing cores and adds a lot of tensor cores. How does it affect the Accerlartion of the Raytracing application compare to T100? Will the Optix code only running on Cuda core A100?
Correct, OptiX will run on the CUDA cores for any supported GPU that does not come with RT cores, including the A100.
Are you asking about the T1000 vs the A100? They are such completely different GPUs that I don’t know how to compare them. Neither one is optimal for general ray tracing performance. The A100 has a lot of memory, and the T1000 has very little. The A100 is a large high end GPU and the T1000 is a low power small GPU. The expected single precision floating point performance is around 20x difference.
The main reason to consider the A100 for ray tracing is if you need double precision floating point, and your application is dominated by and bottlenecked on floating point performance. If you truly require double precision and your shading is the bottleneck, then using the A100 instead of (for example) the A6000, you might achieve faster performance. If your ray tracing is not bottlenecked on shading, but is limited by traversal & ray intersection speeds, then the A6000 will be considerably faster than the A100 for tracing rays. Another reason to use A100 is if you have large compute workloads that depend on Tensor Cores or double precision, and your ray tracing workload is a relatively small part of your overall needs. Or if you have access to A100s but not RTX GPUs for whatever reason.
For GPU without RT core like V100 or A100, what is the main benefit of using Optix instead of writing ray tracing code in CUDA? Does Optix provide any hardware-level advantage?
The benefits of using OptiX without RT Cores include:
- The BVH building & ray tracing (traversal) code is written for you
- Triangle and curve intersectors are included
- OptiX is optimized for performance on the CUDA cores
- The OptiX ABI (calling convention) is optimized compared to CUDA functions
- OptiX offers support for compilation, and includes a compile cache
- OptiX offers features like motion blur, multi-level traversal, validation mode, exceptions, etc.
- The OptiX programming model is established and well designed for scalability, and matches the key concepts in DXR and Vulkan
- You can switch to GPUs that have RT cores without changing any code, and it will go much faster
- OptiX Denoiser is available