Almost every guide to running models at home assumes an NVIDIA card, and anyone holding a Radeon or an Arc gets pointed at CUDA install steps and left there. The 2026 answer for those two vendors is neither a win nor a dead end: it turns on which card, which operating system, and, for Intel, on a change most guides missed.
AMD Support Is Tiered by Card and by OS
AMD’s Windows system requirements mark hardware supported only where “Official software distributions of the current HIP SDK release fully support this hardware.” Three tiers result.
RDNA4 is the top tier. The RX 9070 XT, 9070, 9070 GRE, 9060 XT and 9060 appear in the ROCm compatibility matrix for Linux, and they carry a full Windows 11 25H2 row: runtime, HIP SDK and ROCm Debugger, all supported.
RDNA3 is the mainstream tier: RX 7900 XTX and XT, 7800 XT, 7700 XT, 7650 GRE, 7600 XT and 7600, plus Radeon PRO W7900, W7800 and W7700. Supported on Linux, and on Windows 11 25H2 for the runtime and HIP SDK, but the Debugger column there is unsupported, a tooling limit rather than a blocker.
RDNA2 is where the documentation stops agreeing with itself.
The RX 6000 Discrepancy
On Windows, AMD’s table is unambiguous: ten RX 6000 parts, from the 6950 XT down to the 6600, are marked unsupported in every column, as are four Radeon PRO cards. On Linux the series has no row in AMD’s matrix at all.
Yet Ollama’s GPU documentation lists several of those cards, the RX 6950 XT, 6800 XT and 6800 among them, as usable on Linux through the ROCm build Ollama bundles. Both are primary sources, and they disagree.
The gap is mostly one environment variable. llama.cpp’s build documentation says GPUs outside the official list may work by setting HSA_OVERRIDE_GFX_VERSION, and that this “is not supported on Windows.” Ollama documents the same override plus ROCR_VISIBLE_DEVICES, hedging identically about forcing “a similar LLVM target that is close.” It also requires “an AMD ROCm v7 / HIP7-capable driver stack” on Windows. So the RX 6000 route is Linux, a runner with its own ROCm, and an override: those projects’ claim, not AMD’s.
Flash Attention Is Linux-First and Skips RDNA2
AMD’s flash-attention repository requires ROCm 6.0 and above and, on Linux, names MI200x, MI250x, MI300x, MI355x and RDNA 3/4 for its Composable Kernel backend, CDNA and RDNA for Triton. Windows “might work starting with version 2.3.2,” but “Windows compilation still requires more testing,” which is not a support claim. RDNA2 appears in neither list, an omission rather than a stated exclusion.
Both of Intel’s Own LLM Libraries Are Gone
IPEX-LLM, Intel’s own “LLM acceleration library for Intel GPU … NPU and CPU,” was archived on 28 January 2026 and “is now read-only.” The notice adds that “Patches to this project are no longer accepted by Intel” and that it “has been identified as having known security issues.”
Intel Extension for PyTorch is on the same path: end of life “planned for March 2026,” the 2.8 release getting “critical bug fixes and security patches for two additional quarters,” and Intel’s own advice being “We strongly recommend using PyTorch directly going forward.” That date has passed.
An Arc guide built on either library describes software its vendor has stopped maintaining, one with a security flag attached.
What Still Works on Intel: llama.cpp and SYCL
llama.cpp’s backend table pairs HIP with AMD GPUs and SYCL with Intel GPUs, and marks its OpenVINO entry “[In Progress].” The SYCL documentation says the backend “supports Intel GPU Family: Intel Data Center Max Series, Intel Flex Series, Arc Series, Intel Built-in Arc GPU, Intel iGPU in Core CPU (11th Generation Core CPU and newer…),” with verified systems listed below.
Two numbers matter before setup: “llama-2-7b.Q4_0 requires at least 8.0GB for integrated GPU and 4.0GB for discrete GPU,” and “If the iGPU has less than 80 EUs, the inference speed will likely be too slow for practical use.”
One throughput figure has enough attribution to repeat. On an Intel Arc A770 running llama-2-7b.Q4_0.gguf on llama.cpp’s SYCL backend, a MUL_MAT optimization labelled 2025.2 moved the recorded rate from 42 tok/s to 55 tok/s, a 30 percent gain. No batch size, prompt length or context length is given beside it, so treat it as a before-and-after on one optimization, not a portable benchmark.
Support at a Glance
| Hardware | OS | Stack | Status |
|---|---|---|---|
| AMD RDNA4 (RX 9070, 9060) | Linux | ROCm HIP SDK | Officially supported |
| AMD RDNA4 | Windows 11 25H2 | Runtime, HIP SDK, Debugger | Supported, all three |
| AMD RDNA3 (RX 7000, PRO W7000) | Linux | ROCm | Officially supported |
| AMD RDNA3 | Windows 11 25H2 | Runtime, HIP SDK | Debugger unsupported |
| AMD RDNA2 (RX 6000) | Windows | ROCm | Unsupported, every column |
| AMD RDNA2 (RX 6000) | Linux | AMD matrix | Not listed at all |
| AMD RX 6950 XT, 6800 XT, 6800 | Linux | Ollama, bundled ROCm v7 | Listed by Ollama, not AMD |
| AMD cards outside the matrix | Linux only | llama.cpp HIP | HSA_OVERRIDE_GFX_VERSION, unofficial |
| AMD CDNA, RDNA3/4 | Linux | ROCm flash-attention | Supported, needs ROCm 6.0+ |
| Intel Arc A-series, B580, Max, Flex, iGPUs | Ubuntu 22.04, Fedora Silverblue 39, Arch, Windows 11 | llama.cpp SYCL | Verified by the project |
| Intel GPUs | Windows, Linux | IPEX-LLM | Archived 28 Jan 2026, security flag |
| Intel GPUs (XPU) | Windows, Linux | Intel Extension for PyTorch | End of life planned March 2026 |
| AMD and Intel | Linux, Windows, macOS | llama.cpp Vulkan | Generic GPU backend |
What This Article Does Not Cover
No AMD throughput number appears above: none could be traced to vendor or project documentation for this piece. LM Studio and vLLM are absent for the same reason, their install docs being unreachable. No model licence, parameter count or context length is stated here either.
What This Means
Two questions decide the rest: which card, and which OS. On AMD, RDNA4 and RDNA3 sit inside official support on Linux and Windows 11, minus the debugger for RDNA3 on Windows. RX 6000 sits outside it on both, and the route Ollama and llama.cpp document is Linux-only. Flash attention narrows further, to ROCm 6.0 or newer on CDNA and RDNA3/4.
On Intel, the vendor’s own tooling is no longer the answer. IPEX-LLM is archived and flagged, Intel Extension for PyTorch has passed its planned end of life, and Intel points at upstream PyTorch instead. What still carries current documentation is llama.cpp’s SYCL backend.
The Bottom Line
AMD support is real and tiered: full on RDNA4, near-full on RDNA3, officially absent on RDNA2 even though Ollama’s docs list several RX 6000 cards as working on Linux through an override those same docs say Windows does not support. Intel’s is the sharper story. Both libraries Intel built for this are wound down, an older Arc guide points at dead software, and llama.cpp’s SYCL backend is what is left.