A good write-up on AMD RDNA3 graphics.

[iPwnz]

Especially for those who aren't well aware of.

https://www.angstronomics.com/p/amds-rdna-3-graphics

 

[NanditaRoy]

Heard The AMD Radeon RX 7900 XT is supposed to get over 75 tera Flops of computing power.

 

[Psycho_McCrazy]

12288 ALUs x ~3GHz x 2 Ops/cycle = sounds right, about 75 TFlops

 

[kvn95ss]

Wait, really? Correct me if I'm wrong but are ALUs responsible for AI operations as well? Will it compete with Tensor core of nvidia?

 

[Psycho_McCrazy]

Wait, really? Correct me if I'm wrong but are ALUs responsible for AI operations as well? Will it compete with Tensor core of nvidia?

The above is FP32 performance.

AI inferencing is at INT8 operations, which depending on the architecture may be multiples of this TFlops rating.

For AMD, check the latest CDNA professional compute orientated architecture for real powerhouses of compute.

 

[tech.monk]

Will it compete with Tensor core of nvidia

AMD (read ROCm) is doing it and so does Intel (for ARC GPUs) but AI game is blocked by Nvidia+Google.... Most of the organizations prefer the Tensorflow (read Google) and TF favours CUDA (read Nvidia).

PyTorch is bringing in support for Radeon GPUs but that is limited and most of the stable frameworks are FB or Chinese and hence adoption is lower in Enterprises. If you're interested, you can check DLPrimitives (https://github.com/artyom-beilis/dlprimitives). Its fairly good approach by the Git Repo Owner on going for OSS Cross Platform lib.

 

[AINDRASIS_GUPTA]

Unless AMD brings major improvements in their lacklustre driver support and AI/ML functionality, Nvidia will still remain the king.

Gone are the days when pure compute power wins. It's the age of AI/ML and even though AMD FSR 2.0 is good in the sense that it's open-source and even supported by Nvidia's GPUs, it can't match the prowess that DLSS 2.0 delivers. Not to mention Nvidia brings a whole slew of other features with their driver technology that AMD can't match.