Corsair AI Workstation 300 · MiniV Vulkan UMA

DFlash Pilot Benchmark

Run dflash-matrix-20260708_153700 · 2026-07-08 · public sanitized metrics

Decision

Do not promote DFlash on MiniV yet.

DFlash activated and accepted draft tokens, but the isolated Vulkan DFlash build was slower than baseline across all tested compatible models. Baseline and DFlash used the same build; only speculative mode changed.

results.json summary.csv comparisons.csv

Context8192

Output cap256 tokens

Drafter quantQ8_0

BackendRADV Vulkan

Buildllama.cpp e8f1015

Qwen3.6 / Qwopus3.6 27B Q6_K

Baseline decode9.48 tok/s

DFlash decode8.34 tok/s

DFlash delta-12.1%

Draft acceptance30.8%

Mean accepted length3.45

Qwen3.6 35B-A3B Q6_K_XL

Baseline decode51.07 tok/s

DFlash decode33.72 tok/s

DFlash delta-34.0%

Draft acceptance37.4%

Mean accepted length3.92

Gemma4 26B-A4B-it Q6_K_XL

Baseline decode44.83 tok/s

DFlash decode31.10 tok/s

DFlash delta-30.6%

Draft acceptance34.3%

Mean accepted length3.64

Per-case metrics

Prefill and decode are separated. Decode uses llama-server predicted timing; wall decode is shown separately. Status SPEED_VALID_EMPTY_CONTENT means 256 completion tokens were generated, but content was empty due thinking-template behavior in this isolated PR build; this report is runtime-only, not quality scoring.

Model	Mode	Load sec	Prefill tok/s	Decode tok/s	Wall tok/s	Draft accept %	Status
Qwen3.6 / Qwopus3.6 27B Q6_K	baseline	64.465	45.255	9.481	9.021	—	SPEED_VALID_EMPTY_CONTENT
Qwen3.6 / Qwopus3.6 27B Q6_K	dflash	70.16	41.807	8.337	7.949	30.8	SPEED_VALID_EMPTY_CONTENT
Qwen3.6 35B-A3B Q6_K_XL	baseline	91.402	77.259	51.072	43.802	—	SPEED_VALID_EMPTY_CONTENT
Qwen3.6 35B-A3B Q6_K_XL	dflash	94.191	83.797	33.717	30.603	37.4	SPEED_VALID_EMPTY_CONTENT
Gemma4 26B-A4B-it Q6_K_XL	baseline	64.135	176.638	44.831	41.136	—	SPEED_VALID_EMPTY_CONTENT
Gemma4 26B-A4B-it Q6_K_XL	dflash	65.16	159.646	31.1	29.207	34.3	SPEED_VALID_EMPTY_CONTENT

Methodology guardrails

YAGNI pilot: one deterministic prompt, three compatible model families, Q8 drafters only.
Production TurboQuant router was not replaced; this was a separate DFlash build lane.
Public artifact excludes internal paths, IPs, process commands, and local configuration.
No combined TPS metric is used: load, prefill, decode, wall decode, and draft acceptance are separate.