2x R9700 tcclaviger vLLM Qwen3.8-Flash-Next MTP-3 aggregate decode, c1-distinct
100.75 tok/s (aggregate across 1 concurrent requests, short prompt, 2048 output)
Fields
id
benchmarks/llmb-r9700-2x-tcclaviger-qwen38-flash-next-sustained-c1-distinct-tgMetric
tps_aggregate
Value
100.75
Unit
tok/s (aggregate across 1 concurrent requests, short prompt, 2048 output)
Model
Qwen3.8-Flash-Next MXFP4-FP8-GPTQ (tcclaviger)
Methodology
Donato Capitella, local-llm-benchmarks.dev r9700-tcclaviger-sustained.csv, same 2026-10-07 measurement as the tcclaviger writeup at local-llm-benchmarks.dev/tcclaviger.html. Separate workload from the depth curves: short chat prompts of 57-68 tokens, 2048 output tokens per request, temperature 0, top_p 1, top_k -1, thinking disabled, ignore_eos. One batch; all requests completed exactly 2048 tokens. Same engine and weights as run-4d514b0d7034191a: tcclaviger vLLM 0.29.0.dev0+g2bdbbc8080, ROCm 10.0, Qwen3.8-Flash-Next MXFP4/FP8/GPTQ, MTP-3, tensor parallel 2, expert offload 36 GB, PLE on NVMe, two Radeon AI PRO R9700. Batch c1-distinct: concurrency 1, distinct prompts. Aggregate TG spans first output to final stream end. Early/late window rates (all requests still active, first and last 10 s): 95.1 / 106.1 tok/s. MTP draft acceptance 63.87%. AGGREGATE across the batch, not per request.
Measured by
third_party
Reproducible
yes
notes
Sustained batch, not a depth-curve point. The page notes C2/C4 scaled poorly and C8/C16 were not tested.
Notes
Sustained batch, not a depth-curve point. The page notes C2/C4 scaled poorly and C8/C16 were not tested.