Skip to content

Latest commit

 

History

1 Commit

Folders and files

NameName
Last commit message
Last commit date
 
 

Repository files navigation

AMD Strix Halo (Ryzen AI Max+ 395) LLM benchmarks — measured

Real measured numbers for local LLM inference on a Strix Halo box: decode pace, time to first answer token, prompt-cache effect, sustained-load drift, quant and backend comparisons. Every value below links to a permanent claim page with its scope, limitations and sealed raw runs — nothing here is hand-typed, the tables are generated from the claim registry.

The box: Beelink GTR9 Pro — AMD Ryzen AI Max+ 395, Radeon 8060S (gfx1151), 128 GB LPDDR5X-8000 unified memory. Two commercially identical units. Runtime: llama.cpp server (Vulkan and ROCm backends), build pinned by image digest. Models: pinned by artifact hash.

Quick answers, from the tables below:

  • How fast is a 35B-class MoE on Strix Halo? Sub-second first answer token and comfortable single-reader decode on Qwen3.6-35B-A3B Q4_K_M (llama.cpp Vulkan) — exact figures in the interactive table.
  • Q8_0 or Q4_K_M? Same box, same build: quality gates saw no difference, Q8_0 decodes slower. Full write-up.
  • Vulkan or ROCm? Vulkan decodes faster in this configuration; details and limits in the comparison.
  • Does it throttle all day? Three continuous hours at concurrency 4: drift between the first and last five minutes is in the endurance table. Write-up.
  • Can a team share one box? The concurrency ladder (1 → 4 → 8) is measured in the capacity report.

The numbers

Interactive assistant — everyday human requests

Metric Model Backend Concurrency Value Evidence level Claim page
answerless (empty) responses in default mode gemma-4-26B-A4B Q4_0 llama.cpp Vulkan 1 8.3 % of requests lab_repeated strix.gemma4.interactive2.c1.answerless-default
time to first answer token in default mode gemma-4-26B-A4B Q4_0 llama.cpp Vulkan 1 10768 ms lab_repeated strix.gemma4.interactive2.c1.ttfa-default
time to the first token of any output gemma-4-26B-A4B Q4_0 llama.cpp Vulkan 1 282 ms lab_repeated strix.gemma4.interactive2.c1.ttft-any-token
answerless (empty) responses at a 1k budget Qwen3.6-35B-A3B Q4_K_M llama.cpp Vulkan 1 58.3 % of requests lab_unit_replicated strix.qwen36.interactive2.c1.answerless-1k
request completion Qwen3.6-35B-A3B Q4_K_M llama.cpp Vulkan 1 100.0 % of requests lab_unit_replicated strix.qwen36.interactive2.c1.completion-nothink
time to first answer token Qwen3.6-35B-A3B Q4_K_M llama.cpp Vulkan 1 210 ms lab_unit_replicated strix.qwen36.interactive2.c1.ttfa-nothink
time to first answer token with reasoning on Qwen3.6-35B-A3B Q4_K_M llama.cpp Vulkan 1 20191 ms lab_unit_replicated strix.qwen36.interactive2.c1.ttfa-thinking
time to the first token of any output Qwen3.6-35B-A3B Q4_K_M llama.cpp Vulkan 1 212 ms lab_unit_replicated strix.qwen36.interactive2.c1.ttft-any-token
request completion Qwen3.6-35B-A3B Q4_K_M llama.cpp Vulkan 4 100.0 % of requests lab_repeated strix.qwen36.interactive2.c4.completion-nothink
time to first answer token Qwen3.6-35B-A3B Q4_K_M llama.cpp Vulkan 4 338 ms lab_repeated strix.qwen36.interactive2.c4.ttfa-nothink
request completion Qwen3.6-35B-A3B Q4_K_M llama.cpp Vulkan 8 100.0 % of requests lab_repeated strix.qwen36.interactive2.c8.completion-nothink
time to first answer token Qwen3.6-35B-A3B Q4_K_M llama.cpp Vulkan 8 865 ms lab_repeated strix.qwen36.interactive2.c8.ttfa-nothink
inter-token latency Qwen3.6-35B-A3B Q4_K_M llama.cpp ROCm 1 18.7 ms/token lab_repeated strix.qwen36q4.rocm.c1.itl-nothink
time to first answer token Qwen3.6-35B-A3B Q4_K_M llama.cpp ROCm 1 215 ms lab_repeated strix.qwen36q4.rocm.c1.ttfa-nothink
inter-token latency Qwen3.6-35B-A3B Q4_K_M llama.cpp Vulkan 1 15.9 ms/token lab_repeated strix.qwen36q4.vulkan.c1.itl-nothink
inter-token latency Qwen3.6-35B-A3B Q8_0 llama.cpp Vulkan 1 18.7 ms/token lab_repeated strix.qwen36q8.vulkan.c1.itl-nothink

Document Q&A session — what the prompt cache buys

Metric Model Backend Concurrency Value Evidence level Claim page
time to first token, first question over a 32k document Qwen3.6-35B-A3B Q4_K_M llama.cpp Vulkan 1 33940 ms lab_repeated strix.qwen36.docsession.c1.ttft-q1-32k
time to first token, second question with prompt cache on Qwen3.6-35B-A3B Q4_K_M llama.cpp Vulkan 1 860 ms lab_repeated strix.qwen36.docsession.c1.ttft-q2-32k-cache
time to first token, second question with prompt cache off Qwen3.6-35B-A3B Q4_K_M llama.cpp Vulkan 1 33728 ms lab_repeated strix.qwen36.docsession.c1.ttft-q2-32k-nocache
time to first token, second question over an 8k document, cache on Qwen3.6-35B-A3B Q4_K_M llama.cpp Vulkan 1 660 ms lab_repeated strix.qwen36.docsession.c1.ttft-q2-8k-cache

Sustained load — 3 continuous hours at concurrency 4

Metric Model Backend Concurrency Value Evidence level Claim page
request completion over three hours Qwen3.6-35B-A3B Q4_K_M llama.cpp Vulkan 4 100.0 % of requests lab_single_run strix.qwen36.endurance.c4.completion-180m
decode-pace drift over three hours Qwen3.6-35B-A3B Q4_K_M llama.cpp Vulkan 4 1.6 % lab_single_run strix.qwen36.endurance.c4.itl-drift-180m
inter-token latency Qwen3.6-35B-A3B Q4_K_M llama.cpp Vulkan 4 30.0 ms/token lab_single_run strix.qwen36.endurance.c4.itl-median

Long context — needle retrieval and unanswerable control

Metric Model Backend Concurrency Value Evidence level Claim page
unanswerable-control honesty gemma-4-26B-A4B Q4_0 llama.cpp Vulkan 1 75.0 % of requests lab_repeated strix.gemma4.longctx.c1.control-success
needle retrieval success gemma-4-26B-A4B Q4_0 llama.cpp Vulkan 1 95.8 % of requests lab_repeated strix.gemma4.longctx.c1.needle-success
unanswerable-control honesty Qwen3.6-35B-A3B Q4_K_M llama.cpp Vulkan 1 100.0 % of requests lab_unit_replicated strix.qwen36.longctx.c1.control-success
needle retrieval success Qwen3.6-35B-A3B Q4_K_M llama.cpp Vulkan 1 100.0 % of requests lab_repeated strix.qwen36.longctx.c1.needle-success
time to first token at a 2k-token document Qwen3.6-35B-A3B Q4_K_M llama.cpp Vulkan 1 1907 ms lab_repeated strix.qwen36.longctx.c1.ttft-2k-en
time to first token at a 32k-token document Qwen3.6-35B-A3B Q4_K_M llama.cpp Vulkan 1 33965 ms lab_repeated strix.qwen36.longctx.c1.ttft-32k-en

Strict-JSON automation

Metric Model Backend Concurrency Value Evidence level Claim page
strict-JSON task success gemma-4-26B-A4B Q4_0 llama.cpp Vulkan 1 93.8 % of requests lab_repeated strix.gemma4.structured.c1.task-success
end-to-end time to a complete answer Qwen3.6-35B-A3B Q4_K_M llama.cpp Vulkan 1 844 ms lab_repeated strix.qwen36.structured.c1.e2e-nothink
end-to-end time with reasoning on Qwen3.6-35B-A3B Q4_K_M llama.cpp Vulkan 1 14677 ms lab_repeated strix.qwen36.structured.c1.e2e-think4k
strict-JSON task success Qwen3.6-35B-A3B Q4_K_M llama.cpp Vulkan 1 100.0 % of requests lab_repeated strix.qwen36.structured.c1.task-success-nothink
strict-JSON task success with reasoning on Qwen3.6-35B-A3B Q4_K_M llama.cpp Vulkan 1 100.0 % of requests lab_repeated strix.qwen36.structured.c1.task-success-think4k
strict-JSON task success Qwen3.6-35B-A3B Q4_K_M llama.cpp ROCm 1 100.0 % of requests lab_repeated strix.qwen36q4.rocm.c1.task-success
strict-JSON task success Qwen3.6-35B-A3B Q8_0 llama.cpp Vulkan 1 100.0 % of requests lab_repeated strix.qwen36q8.vulkan.c1.task-success

Interactive assistant v1 — superseded corpus, kept for comparison

Metric Model Backend Concurrency Value Evidence level Claim page
answerless (empty) responses at a 1k budget Qwen3.6-35B-A3B Q4_K_M llama.cpp Vulkan 1 75.0 % of requests lab_repeated strix.qwen36.interactive.c1.answerless-1k
time to first answer token Qwen3.6-35B-A3B Q4_K_M llama.cpp Vulkan 1 220 ms lab_repeated strix.qwen36.interactive.c1.ttfa-nothink
time to first answer token with reasoning on Qwen3.6-35B-A3B Q4_K_M llama.cpp Vulkan 1 23453 ms lab_repeated strix.qwen36.interactive.c1.ttfa-thinking
time to the first token of any output Qwen3.6-35B-A3B Q4_K_M llama.cpp Vulkan 1 221 ms lab_repeated strix.qwen36.interactive.c1.ttft-any-token

How these numbers are made

  • Frozen workloads with published corpora; failed requests stay in the denominator.
  • Three repeated runs per cell (evidence level lab_repeated), or the level the table states; lab_unit_replicated means the pattern reproduced on both identical units.
  • Quality gates grade the answer text (format, language, repetition, groundedness), not the HTTP status.
  • Harness, corpora and gate tests are open: agmind-bench. Sealed run bundles: agmind-lab.

Reproduce or dispute

  • Reproduce a claim on your own Strix Halo hardware — confirmed external reproductions raise the claim's evidence level, failed ones get published too.
  • Got a number that disagrees? Submit it — the config delta behind a disagreement is usually more informative than either figure.

Deep dives


Data: CC BY 4.0 — take the tables, keep the claim links. Maintained by AGmind Systems Lab · Telegram

About

Measured LLM benchmarks for AMD Strix Halo / Ryzen AI Max+ 395 (Radeon 8060S, 128 GB unified): llama.cpp Vulkan & ROCm — decode pace, TTFA, prompt cache, quants, sustained load. Every number links to raw runs.

Topics

Resources

Stars

1 star

Watchers

0 watching

Forks

Releases

Packages

Contributors