#amd

Is AMD a real option for local LLMs in 2026, or still the 'it works but...' alternative? The honest answer, from running RDNA3 and RDNA4 cards in the lab: usable, genuinely good value at the 24GB tier, with one persistent caveat you need to know before you buy.

We ran the same two 8B models on both cards with byte-identical weights and identical settings. The RX 7900 XT is ~3.9x faster — consistently across both architectures. Here's the full 2x2, what it actually tells you, and where we got the framing wrong the first time.

First cross-machine baseline of the lab. The same model — Qwen3.6-27B at Q4_K_M, identical Ollama digest a50eda8ed977 (byte-identical weights) — run on all four lab machines under controlled conditions (same 37-token prompt, 512 generated tokens, temperature 0, seed 42, one warmup rep). Result: the three CUDA/ROCm machines cluster within ~10% for generation (Ray 26.74, Evo-X2 25.21, Victor 24.28 tok/s), so for a 27B Q4_K_M model dual-12GB NVIDIA is NOT meaningfully faster than single-20GB AMD. The dual-GPU advantage only matters once the model exceeds ~18GB. Jitori's 8.41 tok/s is Vulkan-fallback performance, NOT the Arc B60's real capability (SYCL is broken on Battlemage — see bm-012). This directly answers Research Question RQ-001.

Vulkan on Intel Battlemage lands within 12% of dual-Radeon AI PRO R9700 for code generation at Q4_K_M, at roughly half the system cost. Arc is the price-performance leader for sub-$2K builds; multi-GPU AMD wins raw throughput once you accept the complexity.

MTP speculative decoding on Qwen3.6-27B (single Radeon AI PRO R9700, llama.cpp Vulkan) holds 55-65 tok/s generation across 8K-80K context - degrading roughly 1.3x over that range - while prompt processing falls from ~760 to ~580 tok/s as context fills. The practical finding: MTP keeps working at long context (acceptance 66-75%), but KV-cache pressure steadily erodes both prompt and generation speed. Founder's prior short-prompt runs hit ~75-86 tok/s; the numbers here use real document prompts and are correspondingly a bit lower.

LFM2.5-8B-A1B (Q4_K_M, 8.47B total / ~1.7B active MoE params, 4.79 GiB) measured on TWO GPUs for a clean MoE-vs-GPU comparison: EvoX2's AMD Radeon RX 7900 XT (gfx1100, ROCm) at 269.01 tok/s gen / 7464.85 tok/s prompt using 6.81GB VRAM, and Jitori's Intel Arc Pro B60 (Battlemage, Vulkan) at 68.79 tok/s gen / 1688.53 tok/s prompt. The RX 7900 XT is ~3.9x faster on this MoE than the Arc B60 - the SAME ratio as on the dense Llama-3.1-8B (see bm-008), confirming the GPU gap is consistent across architectures. Companion to bm-011 (the APU iGPU result on the same machine). Closes the small-model and real-app workload gaps.

Llama-3.1-8B-Instruct (Q4_K_M, 8.03B dense params, 4.58 GiB) measured on TWO GPUs for a clean dense-vs-GPU comparison: Jitori's Intel Arc Pro B60 (Battlemage, Vulkan) at 27.01 tok/s gen / 706.61 tok/s prompt, and EvoX2's AMD Radeon RX 7900 XT (gfx1100, ROCm) at 105.22 tok/s gen / 3210.77 tok/s prompt. The RX 7900 XT is ~3.9x faster on this dense 8B than the Arc B60 - the SAME ratio as on the MoE LFM2.5-8B (see bm-007), confirming the GPU gap is consistent across architectures. Closes the general-work workload and dense-general model-class gaps; the 9th of 10.

LFM2.5-8B-A1B (Q4_K_M, 8.47B params, 4.79 GiB) on the integrated Radeon 8060S GPU of EvoX2's AMD Ryzen AI MAX+ 395 (gfx1151, ~124GB unified memory pool visible to HIP). Generates at 150.16 tok/s and processes prompts at 3661.53 tok/s via ROCm 7.2.2. This is the only unified-memory APU in the fleet - the same machine's discrete RX 7900 XT (bm-007) is ~1.8x faster on this small model, but the APU's point is not raw speed: unified memory removes the VRAM ceiling that bounds every discrete card here. Closes the unified-memory / APU platform gap and answers the 'should I buy Strix Halo / Ryzen AI MAX for local AI?' question with first-party data.

On a single Radeon AI PRO R9700 (gfx1201, RDNA4), llama.cpp Vulkan with the MTP speculative-decoding head is the only path that activates Qwen3.6's Multi-Token Prediction — delivering 66 tok/s, roughly 2.5x the 27 tok/s class of every other backend. If you don't load the -mtp.gguf draft head, Vulkan, HIP, and Ollama all land within a narrow 26-32 tok/s band and the architecture's main speed advantage is left on the table.

For the Ornith-1.0-35B MoE model (qwen35moe family) on a single Radeon AI PRO R9700, Ollama is the clear winner at 78.3 tok/s - roughly 2.9x faster than llama.cpp Vulkan (27.3 tok/s), which also spilled past the 32GB VRAM limit into system RAM. The HIP backend failed outright: it could not parse the Ornith GGUF (version mismatch). On MoE models without an MTP head, Ollama's expert routing currently beats hand-tuned llama.cpp on this build.