In my last thread on Betelgeuse (bel-cli / bel-server), the main pushback in the comments was: "Okay cool, 22 tok/s on IQ2_XS, but 2-bit is too low—what happens on an actual 4-bit quant like IQ4_XS where the experts are almost 2x larger?" "Why were you testing on a single 16GB stick of DDR4 instead of normal dual-channel 2x 8GB?" Fair questions. So I upgraded the oracle + victim-cache DMA path, packed IQ4_XS (~100GB+ class, full 4-bit experts), and tested it on the exact same RTX 3060 12GB under both RAM configs (still 16GB total system RAM in both tests): Config A: 1x 16GB DDR4 (single-channel, ~19 GB/s host bandwidth) Config B: 2x 8GB DDR4 (normal dual-channel 16GB gaming PC setup, ~38 GB/s host bandwidth) Measured Results on IQ4_XS (RTX 3060 12GB + 16GB Total RAM) Model / Quant RAM Config (16GB Total) Stock llama.cpp Strata (--low-ram) Betelgeuse (bel-server) Output Parity vs. Stock IQ4_XS (4-bit) 1x 16GB (Single-Channel) ~0.5 tok/s (thrash) ~4–5 tok/s 14.0 tok/s 100% Byte-Exact IQ4_XS (4-bit) 2x 8GB (Dual-Channel) ~0.6 tok/s (thrash) ~5–6 tok/s 24.0 tok/s 100% Byte-Exact IQ2_XS (GSQ-RCO) 1x 16GB (Single-Channel) 1.47 tok/s 8.11 tok/s 22.03 tok/s (48 tok/s bursts) 100% Byte-Exact Q8_K_XL (8-bit) 1x 16GB (Single-Channel) OOM /   submitted by