TokenMark own benchmark run: NVIDIA A10, 2026-10-14

Rented from Lambda Cloud, 1 GPU. Measured by TokenMark, not a vendor or community figure.
nvidia-smi (name, driver, power limit, max SM clock, PCIe gen, PCIe width, memory): NVIDIA A10, 570.148.08, 150.00 W, 1695 MHz, 4, 16, 23028 MiB
llama.cpp b11609, built from source with CUDA 12.8 for sm_86 (the rows' "CUDA" setting is the driver's CUDA version).
Benchmark: poor-pauls-benchmark (ppb) 2668a04ce007d8b7aecb2af2e896435cc2506cef, llama-server runner, no extra flags.
Settings: ctx 8192, batch 512 and 1024, 1 concurrent user, 10 ShareGPT prompts x 256 generated tokens.
tok/s = generated tokens / wall time per request, includes time to first token. Each row is the median
of its baseline runs (rows whose runs disagree by more than 25% are dropped), same rule as the ppb rows
from other users on tokenmark.app.

Pre-flight override: ppb's VRAM pre-flight flagged some files "OOM likely" on this card (its estimate was
over 90% of the card's memory), so the main pass ran 0 combos for them. They were run once more, same
settings, answering ppb's own "proceed anyway" prompt. Rows from that pass are marked below.

Not run on this card (file larger than its 24 GB less 1.5 GB headroom): Qwen3.6-35B-A3B UD-Q5_K_M and UD-Q6_K, gemma-4-26B-A4B-it UD-Q8_K_XL, Qwen3.6-27B Q8_0. Those rows exist for the H100 and A100 runs.

Rows:
  Qwen3.5-27B (unsloth/Qwen3.5-27B-GGUF) Q4_K_M: 23.48 tok/s, TTFT 326.6 ms, 146.4 W, 2 run(s)
  Qwen3.6-27B (Jackrong/Qwen3.6-27B-GGUF) Q4_K_M: 24.05 tok/s, TTFT 321.2 ms, 145.8 W, 2 run(s)
  Qwen3.5-35B-A3B (unsloth/Qwen3.5-35B-A3B-GGUF) Q4_K_M: 97.86 tok/s, TTFT 142.3 ms, 138.1 W, 2 run(s) (pre-flight override)
  Qwen3.6-35B-A3B (unsloth/Qwen3.6-35B-A3B-GGUF) UD-Q4_K_M: 95.37 tok/s, TTFT 163.9 ms, 132.5 W, 2 run(s) (pre-flight override)
  Qwen3.6-35B-A3B (unsloth/Qwen3.6-35B-A3B-GGUF) UD-Q4_K_XL: 91.3 tok/s, TTFT 159.3 ms, 133.4 W, 2 run(s) (pre-flight override)

results.jsonl: every ppb result row from this run, one per line, in the flat column layout of the
public ppb dataset (made with ppb's own flattener at the SHA above).