deepseek-v4.1 on a 3090 + 3080 ti with 80gb of ram: dropping to 5 experts gains 15%, everything else did nothing

what i wanted to know

the project maya engineering report promised almost twice the throughput by forcing the pcie path for deepseek. i wanted to see what that gives on my box, and above all which settings really move decode when the model's experts don't fit in ram: 132 gb of experts for 80 gb of ram, the game is played between VRAM, ram and the ssd. i kept one thing as the result: tokens per second in generation.

the setup

rtx 3090 + rtx 3080 Ti (each on pcie 4.0 x8), ryzen 7 5700X, 80 gb of DDR4-3200, a pcie 4 nvme. the model: deepseek-v4.1-flash-1017 in GSQ-RCO IQ2_XXS, the compact format that lets this monster fit on gaming hardware, plus the Engram modules in IQ4_NL. two engines on the same package: strata 0.1.51 (ready-made binary) and maya-ds4 from project maya v1.1.28 (compiled by hand for sm_86, the ampere cards). measured with the engine's own CLI, 256 tokens from a fixed prompt, greedy, context 8192, one model load per run, variants interleaved, two runs each unless noted.

the numbers

one single knob moved anything: going from 6 routed experts to 5. decode climbs from 13.75 to 15.76 tok/s, +14.6%. and that isn't noise from my machine: strata announces +16% for this mode on a rtx 5070, a box with 36 gb of VRAM. skipping the light experts on the ssd (DS4_DISK_SKIP_W) changed nothing: at 0.12 as at 0.14, you stay glued to the default value. doubling the cpu threads, from 8 to 16: nothing either.

dspark, the speculative decoding that makes the model guess several tokens instead of generating them one by one, loses big: 9.69 tok/s against 13.75 on average, about a 30% loss. the engine expects this itself when most of the experts live outside the gpu.

and maya? its two default runs give 8.25 and 9.69 tok/s; forcing the pcie path gives 8.16 and 9.99. twice as fast on paper, 35% slower on my machine.

why it happens this way

the engines' logs, on the same package, explain the difference. strata keeps 3068 experts in VRAM with a 64.6 gb store in ram for the 6035 hottest experts: the model sizes that store from the free ram. maya has no ram store at all: 1653 experts in VRAM, the rest read from disk, about 103 mb per token. the recent strata build is 40 to 65% faster than maya on this box.

what doesn't work here

the forced pcie path from the maya report needs the experts copied into ram: 108 gb for this model, my 80 gb wasn't enough. it is also reserved for a single gpu. so the promised gain doesn't come down to a machine where ram can't fit the experts; and maya v1.1.28 dropped streaming for one format after its measurements, with all the experts in ram it announces itself +34.6% slower. on the build side: compiling maya-ds4 failed on a missing cuda_profiler_api.h in the CUDA 12.8 toolkit, the nvidia-cuda-profiler-api headers installed by pip fixed that.

what i keep

5 experts instead of 6: that is THE setting to turn on for a box where the experts spill over into ram, +14.6% here. for the rest, my list of received ideas doesn't hold: dspark loses 30%, more cpu threads are useless when the work is tied to disk data, skipping ssd experts changes nothing (probably because the cpu lane reads in place from files already in the page cache), and maya has no interest on a machine that can't fit the experts in ram. the quality of the 5-expert mode, i didn't measure it: worth checking before making it a default setting.