clm-0091
measured-heremed ●●○
citable URL: https://halobench.com/records/clm-0091/ — this address never moves; the anchor /records/#clm-0091 keeps resolving
Nemotron-3-Super-120B-A12B now has a single record-backed Vulkan d204800 throughput cell under the existing cfg-0108 fingerprint: 130.8127 t/s prefill (run-0492) and 15.8832 t/s decode (run-0493), rc=0, contention=false, with the same stock llama.cpp 3653e6d6d Vulkan / UD-Q4_K_M / f16 KV / --load-mode none boundary as the prior d131072 Vulkan rows. This is a narrow backend-depth extension, not a production/capability verdict: compared only to the existing matched records, Vulkan d204800 is 20.05% slower than ROCm d204800 on prefill and 5.81% faster on decode, while versus Vulkan d131072 it is 11.98% slower on prefill and 4.45% slower on decode.
verified 2026-08-20 · volatility medium
Note — the record's own working
METHOD — HO-006 admitted evidence from /home/aihydra/bench-results/ho006-nemotron3-super-vulkan-d204800-crossover-r1, variant vulkan-f16-d204800-rep1. Command: llama-bench on /home/aihydra/src/llama.cpp-vk3653e6d/build/bin/llama-bench, model NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_M-00001-of-00003.gguf, -ngl 999 -fa 1 -b 2048 -ub 512 --load-mode none -dev Vulkan0 -ctk f16 -ctv f16 -t 16 -p 1024 -n 256 -d 204800 -o json. Run-meta anchor: 2026-08-19T23:34:06Z to 2026-08-19T23:58:15Z, 1449 s, rc=0, contention=false, host aihydra. Postrun validation reported no fault markers, no residual llama processes, and released box claim.
ENERGY STATUS — wall-meter energy is now joined as eng-0198 from the Home Assistant counter-difference over the preserved run-meta window: 69.1300 Wh gross, 171.75 W mean. This is throughput-window energy only, not a correct-answer denominator.
Cited by — computed at build time, never stored
model pages nemotron3-super
candidate gate history nemotron3-super