OPEN LINK ↗
// 69d agoINFRASTRUCTURE
RX 9070 XT hits LM Studio driver wall
A user reports full system hangs when running a 18.4B MoE model on the new AMD RX 9070 XT, highlighting ongoing stability challenges for RDNA 4 in local LLM workloads. The crash occurs when increasing the context window, suggesting VRAM over-allocation triggers a driver timeout.
// ANALYSIS
AMD's RDNA 4 hardware is a leap forward, but Windows driver stability for local AI still trails NVIDIA's robust memory management.
- –The 16GB VRAM on the 9070 XT is enough for the model but not for large context windows with unquantized KV caches.
- –Unlike NVIDIA's "System Memory Fallback," AMD's ROCm/Windows stack often results in a hard freeze when VRAM is exhausted.
- –MoE models (8x3B) have higher memory overhead than dense models, complicating VRAM offloading calculations for new users.
- –KV cache quantization (Q4/Q8) is now a mandatory optimization for AMD users wanting larger context windows on consumer hardware.
// TAGS
lm-studioamdgpullmrocmrdna-4llama-3.2
DISCOVERED
69d ago
2026-03-21
PUBLISHED
69d ago
2026-03-21
RELEVANCE
7/ 10
AUTHOR
redfukker
