AMD Releases Orochi 1.0 For Dynamic Runtime Switching Between Radeon HIP & NVIDIA CUDA

[phoronix]

Phoronix: AMD Releases Orochi 1.0 For Dynamic Runtime Switching Between Radeon HIP & NVIDIA CUDA

Earlier this year AMD's GPUOpen team announced the Orochi project for dynamic HIP/CUDA run-time handling. Orochi makes it easier for application developers to ship AMD HIP and NVIDIA CUDA support within a single code-base / binary that is then selected at run-time based on the GPU in use...

https://www.phoronix.com/news/AMD-GPUOpen-Orochi-1.0

[tildearrow]

The only problem is that it's not CUDA.

All these efforts to free from CUDA are slowly causing fragmentation, which may slow down adoption and therefore makes dethroning CUDA even harder...

[Guest]

Can I as a random end-user take CUDA pre-compiled software, and run it on AMD through HIP as-is without having to re-compile it or ask the dev to?

[Ladis]

Can I as a random end-user take CUDA pre-compiled software, and run it on AMD through HIP as-is without having to re-compile it or ask the dev to?

Are you a Linux fan or why do you visit this Linux-oriented web? Of course you don't want to touch any compiled blob, compile it yourself (then you don't have problem with the AMD and recently also Intel's solution).

[Ladis]

The only problem is that it's not CUDA.

All these efforts to free from CUDA are slowly causing fragmentation, which may slow down adoption...

CUDA won, nobody is trying to kill it anymore. So now they go the path "you can compile your CUDA code with our compiler and run on our GPU".

[Paradigm Shifter]

CUDA won, nobody is trying to kill it anymore. So now they go the path "you can compile your CUDA code with our compiler and run on our GPU".

At this point, close enough.

CUDA won before the battle even began. AMD consistently tripping over its own feet with aborted attempts with OpenCL, heterogeneous computing and so forth just cemented the win.

[boboviz]

AMD consistently tripping over its own feet with aborted attempts with OpenCL, heterogeneous computing and so forth just cemented the win.

Now AMD seems to be on the right way, with ROCM.
But it's poor and late.

[qarium]

CUDA won, nobody is trying to kill it anymore. So now they go the path "you can compile your CUDA code with our compiler and run on our GPU".

no one and nobody ? google: WebGPU...

[Ladis]

no one and nobody ? google: WebGPU...

Can WebGPU run general programs? Because neither Vulkan and DirectX 12 can. Only CUDA and Metal support general memory pointers. Vulkan with some uncommon extension.

[qarium]

Can WebGPU run general programs? Because neither Vulkan and DirectX 12 can. Only CUDA and Metal support general memory pointers. Vulkan with some uncommon extension.

maybe i do not unterstand what you mean... but in my knowlege it can...

WebAssembly+WebGPU can run general programs inside the browser and outside the browser.

also vulkan bytecode can run in WebGPU and also Metal shader language can run in WebGPU...

the only one in the game who is not compatible with WebGPU is DirectX12..