Adds a consumer Ampere card to the throughput table, which currently has no Ampere GPU other than the A100.
Measured with llama-bench from the official prism-b10743-adfffbe Windows CUDA 12.4 release, -ngl 99 -fa on, separate runs for -p 512 -n 0 and -p 0 -n 128, 5 repetitions each, batch size 1, depth 0, no vision tower. Energy is board power from nvidia-smi (power.draw), sampled every 100 ms and averaged over the samples where GPU utilisation was at least 80%, so that model loading does not dilute it; J/tok is that average divided by TG128 throughput, which matches how the existing rows relate power to token generation.
Setup: RTX 3080 10 GB, driver 610.74, Windows 10 (WDDM). The card also drives the desktop display, which keeps about 1.3 GB of VRAM and a few watts busy; both packs still fit entirely in VRAM.
The TG128 figures agree with an independent reproduction on the same card in PrismML-Eng/llama.cpp#283 (61.38 for PQ2_0 and 52.18 for PTQ1_0, on Linux with CUDA 12.8), and with an earlier run of mine on this machine (62.88 and 52.71).
Two observations that fit the notes already in the card: PQ2_0 is faster than PTQ1_0 in decode on this Ampere card, as the "Native low-bit kernels" note anticipates; and the energy per token is higher than on the Ada and Blackwell cards, because the board stays close to its 320 W limit even during bandwidth-bound decode.
