AMD's recent call for community input on ROCm device support has ignited a broader discussion about the company's approach to GPU software development and its competitive position against NVIDIA in the AI and compute market. The discussion reveals deep-seated concerns about AMD's software strategy and highlights the challenges faced by users attempting to leverage AMD GPUs for machine learning and computation tasks.
Software Support Limitations
A major point of contention in the community is AMD's limited and inconsistent software support for their GPU lineup. While NVIDIA provides comprehensive CUDA support across their product range, AMD's ROCm support is notably restricted, with only select high-end cards receiving full support. The situation is particularly problematic for consumer-grade cards, where support can be dropped within a short period after release, leaving users frustrated and questioning their investment decisions.
Currently supported consumer GPUs on ROCm Linux:
- AMD Radeon RX 7900 (XTX, XT variants)
- Select Radeon PRO W7000 series
Key Community Requests:
- Broader support for consumer GPUs
- Longer support lifecycles (minimum 5 years)
- Better documentation and implementation guides
- Consistent support across Linux and Windows platforms
Documentation and Implementation Challenges
Users report significant difficulties in understanding which cards are actually supported, with AMD's official documentation often being contradictory or unclear. The implementation experience varies widely, with some users successfully running applications like Stable Diffusion on technically unsupported cards through community solutions, while others struggle with official channels. This inconsistency in documentation and support has created a barrier to adoption, especially for developers and researchers who require reliable, long-term support for their work.
Hardware vs Software Priority
The community discussion reveals a fundamental criticism of AMD's approach to GPU computing: their hardware-first strategy versus NVIDIA's software-ecosystem focus. While AMD has produced competitive hardware, their software support infrastructure lags significantly behind NVIDIA's CUDA ecosystem. This disparity has led to a situation where, despite having capable hardware, AMD struggles to provide the seamless development experience that has become standard in the industry.
AMD's hardware might be compelling if it had good software support, but it doesn't. CUDA regularly breaks when I try to use Tensorflow on NVIDIA hardware already. Running a poorly-implemented clone of CUDA where even getting Pytorch running is a small miracle is going to be a hard sell.
Signs of Change
Recent developments suggest AMD is beginning to acknowledge these challenges. The company has indicated a shift towards emphasizing software experiences, APIs, and AI, with a roadmap spanning 3 to 5 years. However, the community remains skeptical, citing past promises and the need for immediate, concrete improvements rather than long-term plans.
The situation presents a critical juncture for AMD in the GPU compute market. While the company has shown strength in hardware development and maintains significant partnerships with major tech companies, the lack of comprehensive software support continues to limit their ability to compete effectively with NVIDIA, particularly in the rapidly growing AI and machine learning sectors.
Reference: ROCM Device Support Wishlist #4276