ROCm
An open-source GPU computing platform developed by AMD, serving as an open-source alternative to NVIDIA's CUDA ecosystem. It supports mainstream AI frameworks and targets scenarios including large model training, inference, and AI agents.
Core Facts
ROCm is AMD's open-source software platform for high-performance computing and AI workloads, including the HIP programming interface, which allows CUDA code to be ported to AMD GPUs at relatively low cost.
ROCm (Radeon Open Compute) is AMD's core open-source GPU computing platform competing against the NVIDIA CUDA ecosystem.
HIP is AMD's solution competing with CUDA, featuring syntax highly similar to CUDA, and designed to achieve cross-platform portability of GPU code.
Timeline (last 90 days)
Open-source alternatives (such as ROCm and OneAPI) and domestic GPU software ecosystems are often seen as key paths to breaking vendor lock-in.
Because CUDA is deeply tied to Nvidia hardware and has accumulated more than a decade of ecosystem inertia, developers migrating to other platforms need to rewrite large amounts of code, which creates extremely high switching costs.
vLLM's PagedAttention kernels are tightly bound to NVIDIA CUDA and AMD ROCm, with only experimental support on Apple Silicon.
For buyers, a near-doubling gap in tokens per dollar is enough to change TCO calculations; for developers, the combination of ROCm and vLLM is beginning to reach production readiness.
This performance improvement is mainly attributed to three kinds of work: scheduling and memory management optimizations at the vLLM level, kernel implementations customized for the AMD architecture, and fine-tuning of hardware-level details.
On the PS5, Llama.cpp most likely relies mainly on CPU inference, because the PS5's GPU driver, firmware, and API are all Sony proprietary, no general-purpose compute interface is exposed, and ROCm is almost unusable in a game console environment.
TokenSpeed requires an NVIDIA GPU host (the nightly wheel requires CUDA 13) or an AMD host with ROCm 7.2, and Python 3.11 to 3.13.
On an AMD W7900 GPU using llama.cpp with the ROCm backend, the Q8 quantized version of LFM2.5-24B-A2B sustains an inference speed of about 95 tokens/second.
AMD's ROCm started late, has limited support for consumer-grade GPUs, and its optimization libraries for large language model inference are far less mature than CUDA.
GPUs were originally designed for parallel graphics rendering, and thanks to their large number of general-purpose compute cores and mature programming models (such as CUDA, ROCm), they have become the mainstay of current AI training and inference, balancing flexibility and performance.
40 more timeline events
All Facts (20)
ROCm is AMD's open-source software platform for high-performance computing and AI workloads, including the HIP programming interface, which allows CUDA code to be ported to AMD GPUs at relatively low cost.
90%VerifiedROCm (Radeon Open Compute) is AMD's core open-source GPU computing platform competing against the NVIDIA CUDA ecosystem.
75%VerifiedHIP is AMD's solution competing with CUDA, featuring syntax highly similar to CUDA, and designed to achieve cross-platform portability of GPU code.
75%VerifiedAMD's ROCm and Intel's oneAPI still have significant gaps compared to CUDA in terms of ecosystem completeness, community support, and performance tuning tools.
75%Verifiedllama.cpp supports GPU acceleration via CUDA for NVIDIA GPUs, ROCm for AMD GPUs, and Metal for Apple Silicon, and also supports pure CPU inference.
70%VerifiedROCm is an open-source GPU computing platform launched by AMD in 2016; the current ROCm 6.x version already provides official support for PyTorch and TensorFlow.
65%VerifiedAMD's ROCm platform was launched in 2016, providing HIP as a compatibility layer for CUDA, and PyTorch officially provides precompiled packages for ROCm.
65%VerifiedHIP, the core component of ROCm, provides an API highly similar to CUDA, enabling CUDA code to be converted via the hipify tool to run on AMD GPUs, and the ROCm version of PyTorch reuses the torch.cuda namespace.
65%UnverifiedAMD显卡可通过ROCm或DirectML方案运行Stable Diffusion，但兼容性和性能不如NVIDIA
85%UnverifiedThe new version adds an AMD ROCm Docker image containing Unsloth Studio and JupyterLab, and the ROCm image can access the GPU via WSL2's DXG bridge.
70%UnverifiedWhen the ROCm version cannot run, the system falls back to a Vulkan sd.cpp build.
70%UnverifiedvLLM supports the AMD ROCm platform and is continuously improving that support.
70%UnverifiedAMD installations automatically select the optimal build version on Windows and Linux, and enable BF16 on more GPUs.
70%UnverifiedSimultaneous support for Metal, CUDA, and ROCm is uncommon among open-source inference projects
70%UnverifiedSGLang 的开源属性使其能够比 NVIDIA 主导的 TensorRT-LLM 更快速地集成 AMD ROCm 生态的新特性
70%Unverifiedds4 is written in pure C with native support for Metal, CUDA, and ROCm backends
60%Unverifiedllama.cpp was introduced in early 2023 to run LLaMA models using only a CPU, and subsequently added support for hardware acceleration backends such as Metal, CUDA, and ROCm.
60%UnverifiedBecause CUDA is deeply tied to Nvidia hardware and has accumulated more than a decade of ecosystem inertia, developers migrating to other platforms need to rewrite large amounts of code, which creates extremely high switching costs.
50%UnverifiedOpen-source alternatives (such as ROCm and OneAPI) and domestic GPU software ecosystems are often seen as key paths to breaking vendor lock-in.
50%UnverifiedvLLM's PagedAttention kernels are tightly bound to NVIDIA CUDA and AMD ROCm, with only experimental support on Apple Silicon.
50%