cc@cudawarped

@johnnynunez Which version of the CUDA toolkit are you using to compile for Blackwell. The latest version I have access to (12.6 Update 3) does not support compute capability 10. I assume Nvidia will release a version 13.0 to coincide with the release of the first Blackwell cards.

Isn't Blackwell compute capability 10.0, where does 12.0 come from?

@johnnynunez Which version of the CUDA toolkit are you using to compile for Blackwell. The latest version I have access to (12.6 Update 3) does not support compute capability 10. I assume Nvidia will release a version 13.0 to coincide with the release of the first Blackwell cards.

Isn't Blackwell compute capability 10.0, where does 12.0 come from?

Hello,
B100 is coming support with 12.7 cuda
Rtx50 is coming with 12.8

@johnnynunez Which version of the CUDA toolkit are you using to compile for Blackwell. The latest version I have access to (12.6 Update 3) does not support compute capability 10. I assume Nvidia will release a version 13.0 to coincide with the release of the first Blackwell cards.

Isn't Blackwell compute capability 10.0, where does 12.0 come from?

Hello,
B100 is coming support with 12.7 cuda
Rtx50 is coming with 12.8

also thor is 10.1 capability.
I have rtx 5090 card.

@johnnynunez So you haven't tested that this works on Blackwell 🤯? I would wait until a version of the CUDA toolkit is release which supports compute capability 10.0 to be 100% sure your change don't break anything. I would also remove compute capability 12.0.

Note: If you add 10.1 for Thor you also want to update this filter

@johnnynunez So you haven't tested that this works on Blackwell 🤯? I would wait until a version of the CUDA toolkit is release which supports compute capability 10.0 to be 100% sure your change don't break anything. I would also remove compute capability 12.0.

Note: If you add 10.1 for Thor you also want to update this filter

opencv/cmake/OpenCVDetectCUDAUtils.cmake

Line 274 inea023b7

ocv_filter_available_architecture(${nvcc_executable} __cuda_arch_bin

But why you remove 12.0?
My 5090 output is 12.0. I think that digits will be 11.0

@johnnynunez So you haven't tested that this works on Blackwell 🤯? I would wait until a version of the CUDA toolkit is release which supports compute capability 10.0 to be 100% sure your change don't break anything. I would also remove compute capability 12.0.
Note: If you add 10.1 for Thor you also want to update this filter

opencv/cmake/OpenCVDetectCUDAUtils.cmake

Line 274 inea023b7

ocv_filter_available_architecture(${nvcc_executable} __cuda_arch_bin

But why you remove 12.0? My 5090 output is 12.0. I think that digits will be 11.0

Well, today NDA is removed

@johnnynunez So you haven't tested that this works on Blackwell 🤯? I would wait until a version of the CUDA toolkit is release which supports compute capability 10.0 to be 100% sure your change don't break anything. I would also remove compute capability 12.0.
Note: If you add 10.1 for Thor you also want to update this filter

opencv/cmake/OpenCVDetectCUDAUtils.cmake

Line 274 inea023b7

ocv_filter_available_architecture(${nvcc_executable} __cuda_arch_bin

But why you remove 12.0? My 5090 output is 12.0. I think that digits will be 11.0

I added support on pytorch, xla etc

But why you remove 12.0?
My 5090 output is 12.0. I think that digits will be 11.0

Can you provide me with a source to indicate that your 5090 will be compute capability 12.0?

I added support on pytorch, xla etc

Pytorch merged compute capability 10.0 and 12.0 without any build testing (I guess its a python first library)? Not sure I can see a reason for not waiting especially in OpenCV when you can manually select the compute cabability using combinations ofCUDA_ARCH_BIN andCUDA_ARCH_PTX.

But why you remove 12.0?
My 5090 output is 12.0. I think that digits will be 11.0

Can you provide me with a source to indicate that your 5090 will be compute capability 12.0?

I added support on pytorch, xla etc

Pytorch merged compute capability 10.0 and 12.0 without any build testing (I guess its a python first library)? Not sure I can see a reason for not waiting especially in OpenCV when you can manually select the compute cabability using combinations ofCUDA_ARCH_BIN andCUDA_ARCH_PTX.

more references:
pytorch/pytorch#145270 I added to pytorch
Dao-AILab/flash-attention#1436

@johnnynunez 🤯 Nvidia is going up 3 compute capabilities with a single generation, that's going to be really confusing considering they previously did a compute version per generation.

What is the output from

nvidia-smi --query-gpu=compute_cap --format=csv

If it says

compute_cap
12.0

Do you think its possible that the driver is outputing the wrong info?

@asmorkalov Either way I would suggest it would be better to wait until more info is available before merging this PR.

nvidia-smi --query-gpu=compute_cap --format=csv

It's okay, flash attention v4 is coming for blackwell also. They have 100 and 120.
I totally agree haha but well, nvidia now have a lot of products...
ARM upcoming laptops, gpus desktops, digits, jetson, and data centers

@johnnynunez 🤯 Nvidia is going up 3 compute capabilities with a single generation, that's going to be really confusing considering they previously did a compute version per generation.

What is the output from

nvidia-smi --query-gpu=compute_cap --format=csv

If it says

compute_cap
12.0

Do you think its possible that the driver is outputing the wrong info?

@asmorkalov Either way I would suggest it would be better to wait until more info is available before merging this PR.

I share you in the following hours because I'm not at home. But we have press release driver 571.86 whl

@cudawarped Thanks a lot for the analysis. I'll review hardware specs and return back soon.

But why you remove 12.0?
My 5090 output is 12.0. I think that digits will be 11.0

Can you provide me with a source to indicate that your 5090 will be compute capability 12.0?

I added support on pytorch, xla etc

Pytorch merged compute capability 10.0 and 12.0 without any build testing (I guess its a python first library)? Not sure I can see a reason for not waiting especially in OpenCV when you can manually select the compute cabability using combinations ofCUDA_ARCH_BIN andCUDA_ARCH_PTX.

But why you remove 12.0?
My 5090 output is 12.0. I think that digits will be 11.0

Can you provide me with a source to indicate that your 5090 will be compute capability 12.0?

I added support on pytorch, xla etc

Pytorch merged compute capability 10.0 and 12.0 without any build testing (I guess its a python first library)? Not sure I can see a reason for not waiting especially in OpenCV when you can manually select the compute cabability using combinations ofCUDA_ARCH_BIN andCUDA_ARCH_PTX.

new drivers are showing cuda 12.8 and same 12.0 codename

more references:
NVIDIA/cccl#3493

10.0 b100 b200
10.0a arm laptops, digits?
10.1 thor
10.1a arm laptops, digits?
12.0 rtx50

@johnnynunez I still suggest we wait I can't see any downside with OpenCV because the compute capability can be manually selected.

I still wouldn't be suprised if the 12 is the version of the CUDA toolkit due to the returned compute capability not being valid because the CUDA toolkit used by pytorch and used to compile nvidia-smi pre-date compute capabilities >= 9.0.

@johnnynunez I still suggest we wait I can't see any downside with OpenCV because the compute capability can be manually selected.

I still wouldn't be suprised if the 12 is the version of the CUDA toolkit due to the returned compute capability not being valid because the CUDA toolkit used by pytorch and used to compile nvidia-smi pre-date compute capabilities >= 9.0.

yeah! Totally agree

cuda 12.8 is out

@johnnynunez Compute capability 12 looks to be official. Consumer cards look to have less resident threads per SM.

https://docs.nvidia.com/cuda/cuda-c-programming-guide/index.html?highlight=compute%2520capability#features-and-technical-specifications

@asmorkalov Builds on both Windows 11 and Ubuntu 22.04 (WSL) with CUDA Toolkit 12.8 using default architectures selection (CUDA_ARCH_BIN andCUDA_ARCH_PTX not specified)

--   NVIDIA CUDA:                   YES (ver 12.8, CUFFT CUBLAS NVCUVID NVCUVENC)--     NVIDIA GPU arch:             50 52 60 61 70 75 80 86 89 90 100 120--     NVIDIA PTX archs:            120----   cuDNN:                         YES (ver 9.7.0)

and-DCUDA_GENERATION=Blackwell

--   NVIDIA CUDA:                   YES (ver 12.8, CUFFT CUBLAS NVCUVID NVCUVENC)--     NVIDIA GPU arch:             100 120--     NVIDIA PTX archs:----   cuDNN:                         YES (ver 9.7.0)

@asmorkalov Builds on both Windows 11 and Ubuntu 22.04 (WSL) with CUDA Toolkit 12.8 using default architectures selection (CUDA_ARCH_BIN andCUDA_ARCH_PTX not specified)

--   NVIDIA CUDA:                   YES (ver 12.8, CUFFT CUBLAS NVCUVID NVCUVENC)--     NVIDIA GPU arch:             50 52 60 61 70 75 80 86 89 90 100 120--     NVIDIA PTX archs:            120----   cuDNN:                         YES (ver 9.7.0)

and-DCUDA_GENERATION=Blackwell

--   NVIDIA CUDA:                   YES (ver 12.8, CUFFT CUBLAS NVCUVID NVCUVENC)--     NVIDIA GPU arch:             100 120--     NVIDIA PTX archs:----   cuDNN:                         YES (ver 9.7.0)

Yeah I compiled it pytorch, xformers, etc and opencv with my rtx5090 and it works. I just couldn't comment on anything because of NDA. But it was lifted yesterday

@asmorkalov feel free to merge! thanks