The Issue:
Despite successful compilation, ‘gmx mdrun’ fails at findallgputasks.cpp (line 89):
“Fatal error: Cannot run short-ranged nonbonded interactions on a GPU because no GPU is detected.”
Questions:
Is gfx1201 fully supported in the 2026.3-dev branch?
Are there specific sanity checks preventing the kernel dispatch on this architecture?
Attached is the log output of the failing run. Any advice is appreciated.
The rest are not needed, because a correct configuration picks up those settings automatically. Just out of curiosity, where did you get the CMake command line from?
I just build it myself just a few days ago and am running without issues on my 9070XT.
I wanted to follow up and let you know that your advice worked perfectly.
After cleaning my build directory and using the exact flags you provided, GROMACS 2026.3-dev compiled flawlessly. I am now running a production MD simulation on my AMD Radeon RX 9070 XT (gfx1201), and the performance is fantastic. I am getting 100% GPU utilization during the NPT equilibrium and production phases, and the temperature remains very stable.
Thank you again for the support and for all the work you’ve put into the AMD/HIP backend. It’s a huge relief to have a fully functional environment for my research.
To answer your curiosity: I am a student currently self-studying computational chemistry. Because the RX 9070 XT is so new, I was searching through old community forum posts for AMD compilation guides and relied on an AI assistant to help me piece together the CMake command.
It turns out the AI completely hallucinated the GMX_HIP_ARCHITECTURES flag! It seems to have spliced the GROMACS GMX_ prefix with the standard CMAKE_HIP_ARCHITECTURES variable. Since CMake silently ignored this non-existent flag, the build fell back to a default architecture. I actually tried compiling without those two cluster flags initially, but it still failed, which I now realize was because I used the wrong GMX_HIP_ARCHITECTURES flag instead of your corrected GMX_HIP_TARGET_ARCH!
Thanks again for your help in getting this running!
To answer your question about the performance on the RX 9070 XT , I am currently getting 30.373 Matom*steps/s (which translates to about 57.083 ns/day for my specific system). The GPU seems to be handling the workload very smoothly. According to my monitoring tools during the run, the card operates stably at around 62.0°C, drawing roughly 208W with a core clock of 3277 MHz.
By the way, my next research step involves running Absolute Binding Free Energy calculations (specifically, alchemical transformations). Since this heavily relies on free energy kernels, I was wondering: are the free energy implementations fully supported and optimized for the RDNA4 / HIP backend in the current 2026.3-dev branch? Are there any known limitations or specific mdrun flags I should be aware of when running alchemical steps on this specific GPU architecture?
Let me know if you are curious about any other metrics or need further testing data on this card!
the free energy kernels are currently still a MR for the main branch, and should be included in the 2027 release. If you are willing to try the AMD feature branch I’m maintaining you can use them for the current 2026 branch. The branch is 4947-hip-feature-enablement and always tracks the latest GROMACS release. The only additions are AMD specific optimizations that haven’t made it into the official release yet.