I would like to know if the newer RTX 50xx GPUs would have hardware for faster synchronization across thread blocks on the same SM cluster (the feature released with hopper GPUs). I assume it would be yes as Blackwell supersedes Hopper but we haven’t had any RTX GPUs with that capability yet (to the…

For anyone seeing this later (after Jan 25th). The SM version for the blackwell cards is >10.0 and it is going to support thread block clusters.

Thanks, this is a good way to think about it!

[image] Curefab: or that the sections about forward compatibility are changed to 10.0 only and RTX 5000 is 10.5. Seems like they’ve allocated 12.8 for RTX50XX: The Ultimate GeForce GPU Comparison 50 Series Specs

If I am not wrong cuda version and SM architecture can be unrelated

That would be the largest jump since Maxwell (3.7 → 5.0) for this technical version number. And even, if true, would Nvidia have called it Blackwell as well? Perhaps they mixed up with the needed Cuda Toolkit version? Other 3rd party pages specify 10.1. Here is a bit discussion about the consumer B…

It’s the CUDA toolkit (CTK) version. Remember 11.8 was the first to provide some support for cc8.9/cc9.0. I’m guessing that 12.8 will be the first CTK version to provide some support for the “new” GPUs. [image] samkawtikwar: If I am not wrong cuda version and SM architecture can be unrelated …

It is 10.0 for the datacenter blackwell cards (e.g. B100, B200) and 12.0 for the consumer blackwell cards (e.g. RTX 5090). There seems to be a compute capability 10.1 card planned, could be the Jetson embedded cards or the announced Digits AI Workstation. Or some datacenter refresh. The consumer bl…

Where does it say the RTX50xx cards are compute capability 12.0? I only see 12.8 on the spec sheet which as discussed seems like the CTK version.

For example, the rtx Blackwell whitepaper mentions that each SM in the GB202 gpu has 2 fp64 cores. In the programming guide the description of cc 12.0 mentions 2 fp64 cores, as well.

For example here you can see, which compute capability version was supported with with toolkit: https://docs.nvidia.com/cuda/parallel-thread-execution/index.html#release-notes 10.0 and 10.1 are new for toolkit 12.7 and 12.0 for toolkit 12.8. Then as @striker159 said, look at the programming guide …

Thread block clustering in Blackwell GPUs

Accelerated Computing CUDA CUDA Programming and Performance

rs277 January 22, 2025, 12:45am 9

Nvidia might want to correct the graphic I linked to above then, as all the previous generations listed show the CC value, not the Toolkit release supporting them.

Topic		Replies	Views
How does the Thread Block Cluster of the Nvidia H100 work concurrently? CUDA Programming and Performance gpu	26	7840	October 12, 2022
Does only H100 support the kernel launch configuration attribute for clustering? CUDA Programming and Performance	10	260	March 18, 2025
No thread block clusters on GB10? DGX Spark / GB10	6	160	January 22, 2026
CUDA Toolkit 12.8 what GPU is 'sm_120'? CUDA NVCC Compiler	8	10813	February 8, 2025
Confusion about setting kernel block and grid size for maximum occupancy CUDA Programming and Performance cuda	11	1280	March 30, 2024
Maximum number of threads How to find maximum number of threads your Card can support CUDA Programming and Performance	16	10614	July 7, 2009
Cluster size limitation CUDA Programming and Performance	4	1070	February 1, 2024
Confusing H100 SXM thread block cluster CUDA Programming and Performance cuda , kernel	12	334	October 29, 2025
CUDA 9 Features Revealed: Volta, Cooperative Groups and More Technical Blog	45	1154	November 2, 2017
Block Clustering in Ada architecture CUDA Programming and Performance	2	135	September 4, 2024

Thread block clustering in Blackwell GPUs

Related topics