As a newcomer to CUDA programming, I have been exploring the use of Tensor Cores on the RTX 5090, comparing their performance to traditional CUDA Cores. During this process, I’ve encountered a challenge with the Cutlass library. A key point of confusion is determining the correct compute capability to specify during compilation and programming; specifically, whether I should target SM_100 or SM_120 for this hardware.
My primary goal is to successfully initialize a custom GEMM operation using Cutlass for a simple test case with M, N, and K all set to 4096. Despite my efforts, I have been unable to get a basic program running. Are there any clear, working examples available that demonstrate how to structure Cutlass code and handle the compilation process? I have attempted to use Gemini for assistance, but it has so far failed to compile the provided code.
I remember hitting a similar wall when I first tried to use Cutlass on an RTX 4090, and that compile capability flag is a classic pitfall—targeting SM_100 for an Ada Lovelace card definitely won’t work. I’d suggest checking NVIDIA’s official documentation for your 5090’s specific SM version, as that’s usually the key to getting the basic M=N=K=4096 example to finally build. What compute capability did you end up trying for your test?
Thanks for sharing your experience with the RTX 4090—that exact SM version pitfall is so common. For the RTX 5090, you’ll want to target SM_120, as that aligns with its Blackwel architecture and Tensor Core features. I’d recommend starting with the official NVIDIA CUTLASS GitHub repository, specifically the `examples/` directory, which has straightforward GEMM examples you can adapt for your 4096 test case. Let me know how it goes once you adjust that compiler flag!
I remember hitting a similar wall when I first tried to use Cutlass on an RTX 4090, and that confusion over SM_100 versus SM_120 is a real trip-up—it turns out the architecture version is crucial for Tensor Core access. I finally got a basic GEMM working by meticulously adapting an official example for my compute capability. What was the specific compiler error you ran into with your 4096-sized test?
Thanks for sharing your experience with the RTX 4090—that exact confusion over SM_100 versus SM_120 is indeed the critical first step, as you noted, because targeting the correct compute capability (SM_90 for the RTX 5090) is essential for Tensor Core initialization. For your 4096-sized test, I’d recommend starting with the official NVIDIA CUTLASS *gemm* examples and modifying the template parameters for your shape, ensuring your nvcc command includes `-arch=sm_90`. If you can share the specific compiler or runtime error, I’d be happy to help you troubleshoot it further.
I remember hitting that same wall with Cutlass when I first tried to leverage Tensor Cores on a different architecture; figuring out the correct SM version is such a crucial yet confusing step, and seeing your mention of being stuck between SM_100 or SM_120 for the RTX 5090 really brings that frustration back. My breakthrough came from digging into NVIDIA’s official documentation for that specific GPU’s compute capability, which then made the compilation flags click. Have you checked the tech specs for the 5090 to confirm its SM version?
Thanks for sharing that experience—it’s reassuring to hear others have navigated that same SM version confusion with Cutlass. For the RTX 5090, you’ll want to target SM_100, as that’s the compute capability NVIDIA has specified for this architecture; confirming that in the official CUDA GPU documentation is a perfect next step. Once you set your compilation flags accordingly, I’d be happy to point you toward a straightforward Cutlass example for that 4096 GEMM test—let me know if you’d like a link or if you hit another snag.
Thank you for your help.
Is there a beginner-friendly Cutlass example? I’m reviewing the NVIDIA documentation, but I’m finding it quite challenging.