Strix Halo + dGPU real-world test: low-context benchmarks oversell the speedup
Hrethric · reddit · 2026-08-25
The author attached an R9700 dGPU to a Strix Halo machine via Oculink. Running Unsloth quants of MiniMax M-2.7, Q4KXL on the dGPU was slower than IQ4XS on the APU alone; after hours of tuning, IQ4XS matched APU token generation with a 2-3x prompt-processing boost and a modest context bump.
Many published dGPU speedups were run at low context, he notes. There is a real tradeoff: offload more layers to the dGPU at the cost of context size/quality, or keep a higher-quality context with fewer dGPU layers. Full Docker command and models.ini settings (tensor-split, cache-type, batch sizes) are included.
More from Infra
- Next-gen GraalVM benchmarks: 4x faster startup, 2x less memory, no build step needed — RealGeneKim · 2026-10-15
- How Microsoft optimized Windows for Nvidia's Arm-based RTX Spark AI laptops — tomwarren · 2026-10-15
- How Microsoft optimized Windows for Nvidia's Arm-based RTX Spark laptops — The Verge AI · 2026-10-15
- Global smartphone shipments plunge 22% in worst quarterly decline on record as AI demand fuels memory costs — Polymarket · 2026-10-15
- SK Hynix drops 11% again, now below 3x FY2027 earnings as AI memory selloff deepens — JOBhakdi · 2026-10-15
- ASML Sells Zero Lithography Machines into Europe as US Hosts 75%+ of AI Compute — NinaDSchick · 2026-10-15