Marc Philipp Menzelstrasse 25 12157 Berlin Germany
Dear AMD Developer Relations and ROCm Engineering Teams, I am reaching out as an early adopter of the AMD RDNA 4 (gfx1200) architecture. My recent efforts to deploy Large Language Model (LLM) inference environments on Ubuntu 25.10 have revealed a systemic failure in the current Linux compute stack that prevents this next-generation hardware from being utilized for AI workloads. I have performed a comprehensive multi-day diagnostic cycle to isolate these failures. I am sharing these findings in the hope of assisting your engineering teams in stabilizing the platform ahead of the major Linux distribution updates expected this spring. Reference Environment: AI Workstation To ensure reproducibility, all diagnostics were performed on the following hardware configuration: GPU: AMD Radeon RX 9060 XT (RDNA 4 / gfx1200) | 16GB VRAM CPU: AMD Ryzen 5 7600 (6C/12T) Motherboard: MSI MAG B650 TOMAHAWK WIFI (BIOS: AGESA 1.2.0.2+) Memory: 32GB DDR5-6000 MT/s (Micron | EXPO Profile 1 active) Storage: Lexar Professional NVMe Gen4 SSD OS: Ubuntu 25.10 (Kernel 6.12+ / Mesa 24.3+) I. Technical Diagnostic Summary My testing was conducted using the llama.cpp (GGUF) framework across various backends. The following critical roadblocks were identified: Vulkan/RADV Compute Instability: While the Mesa RADV driver correctly identifies the gfx1200 ISA, the compute pipeline is currently unstable. During SPIR-V to ISA translation (ACO compiler), the runtime consistently triggers a SIGSEGV (Exit Code 139) or enters an infinite loop at 100% CPU load. This suggests the RADV compute path for RDNA 4 is not yet parity-stable with the graphics path. ROCm Repository Dependency Mismatch: Official ROCm (6.2/6.3) support is currently limited to LTS releases (22.04/24.04). Attempting to use these on Ubuntu 25.10 creates catastrophic dependency breaks with glibc and llvm-roc. There is currently no official “interim” or “distro-agnostic” path for HIP compute on the latest kernels required to fully drive RDNA 4. OpenCL Architecture Mapping: The OpenCL/CLBlast backends fail to correctly map the gfx1200 architecture, defaulting to generic or incompatible mobile kernel parameters. This results in compilation failures due to undefined SIMD-group widths specific to the RDNA 4 instruction set. Container/Host ISA Desync: Attempts to encapsulate the workload in an Ubuntu 24.04 OCI container (to use stable ROCm libraries) failed due to a versioning delta between the Host amdgpu kernel module (required for RDNA 4) and the Container libhsa-runtime. II. Inquiry Regarding the April Roadmap (Ubuntu 26.04 LTS): With the upcoming release of Ubuntu 26.04 LTS, I would like to inquire about AMD’s readiness regarding the following points: Synchronized ROCm Support: Will a native ROCm repository be available for 26.04 LTS on Day‑0? Mesa Upstream Status: Are the compute fixes for gfx1200 scheduled for inclusion in the Mesa versions that will ship with the next LTS? III. AMD Vanguard Admission & Compensation Request The diagnostic work required to map these regressions represents a significant investment of professional time and specialized technical labor. I believe that early adopters who provide this level of deep‑dive feedback are essential to the health of the ROCm ecosystem. I would like to be considered for the AMD Vanguard program to assist in the validation of RDNA 4 compute stability on Linux. Furthermore, given that the hardware’s marketed AI capabilities are currently non-functional in this environment, I would like to ask about potential goodwill gestures or compensation for developers providing these detailed diagnostic reports. This could include: Direct enrollment in the AMD Vanguard or specialized beta-testing programs. Access to pre-release ROCm builds or the AMD Developer Cloud. Credits or discounts toward future AMD compute/accelerator hardware. Conclusion: RDNA 4 is a remarkable achievement in silicon, but the current software fragmentation on Linux is a significant barrier to adoption. I am eager to help bridge this gap and am available to provide full system logs and core dumps to your engineering teams. I look forward to your response regarding the April roadmap and my potential participation in the Vanguard program. Best regards from Berlin, Germany Marc Philipp
