5-cell mode-build sweep

Oct 13, 2026

Throughput

13.0 t/s gen · 382.5 tokens/s prompt processing (prefill)Prompt processing measures input prefill speed; it does not tell you the time to first token.

Quant

Q4_0 (GGUF)

KV cache

F16

Flash Attention

MTP (Multi-Token Prediction)

off

System RAM

128 GB

Summary

First-party historical benchmark from daimonionnn, inspected 2026-10-13. Evidence is pinned to source commit 7de1d467c9f2e8282051d9d2332f5ef584f4fa66. No benchmark was rerun or weights downloaded. Hardware: Minisforum MS-S1 MAX, AMD Ryzen AI MAX+ 395, Radeon 8060S (gfx1151), 128 GB-class unified LPDDR5X-8000, Windows 11 Pro. Machine identity and OS are source-reported. Short-pack CSV captures 128 memory units at 8000 MT/s; its PowerShell helper sums Win32_PhysicalMemory Capacity using 1GB (binary GiB), although the column is named ram_total_gb. This is unified capacity, not dedicated VRAM or resident memory. The reported eight 16 GB firmware entries do not establish removable DIMMs. Physical CPU core count and enabled GPU core/CU count are unknown. Source describes 32 CPU threads; actual benchmark worker threads are 16. Participating GPU count = 1, bound to the measured process: every retained native row has gpu_info exactly AMD Radeon(TM) 8060S Graphics, devices auto, main_gpu 0, n_gpu_layers -1, n_cpu_moe 0, layer split, tensor_split 0.00 and no tensor buffer overrides. Both recorded engine revisions' get_gpu_info iterate every GPU/IGPU and join every description without deduplication. Thus this is one enumerated GPU with full layer offload. The older CSV's backends CUDA,ROCm names registered backend libraries, not two GPUs; GPU enumeration contains only the 8060S. Device identity is not borrowed from another run. Timing: native llama-bench synthetic token IDs, not an HTTP request. Every trial clears KV; prompt and generation warmup precede measured trials. Model load and warmup are excluded. The timer starts before test_prompt/test_gen and stops after synchronization. PP counts decoded input tokens; TG counts serial one-token decode calls. avg_ts is the arithmetic mean of per-trial tokens/elapsed time, and stddev_ts is sample SD; it is not tokens divided by avg_ns. The scripts/CSV retain aggregates, not per-trial samples, so these are publisher-reported native statistics, not independently reconstructed trials. No independent TTFT exists; tokens/rate in the source is a prefill-duration proxy and is excluded from the TTFT field. Settings: batch 2048, microbatch 512, 16 CPU workers, f16 K and V cache, flash attention on, mmap on, direct I/O off, embedding mode off, KV offload on, no CPU MoE, no speculative decoding. Requested context in this benchmark code is n_prompt+n_gen+n_depth; final rounded allocation is not logged and stays unknown. Actual timed tokens and existing context depth are separately recorded below. Model: Gemma 3 27B IT QAT, Q4_0. Exact measured local filename: gemma-3-27B-it-QAT-Q4_0.gguf. Logged tensor parameter count 27009346304; logged model tensor bytes 15560833664. Nominal family size 27B is not a multiplication of architecture labels or measured RAM usage. Artifact provenance: measured local path names lmstudio-community/gemma-3-27B-it-qat-GGUF/gemma-3-27B-it-QAT-Q4_0.gguf. Current HF metadata and the file page were retrieved successfully; current repository revision c29c7d93c128620e932db6a18604a25c4c40a17d is a reference checked today, not a proven tested revision. Historical local weight hash/HF snapshot is absent. HF card lists upstream google/gemma-3-27b-it. Artifact weights were not downloaded. Five retained campaign cells, three repetitions per separate PP512 and TG128 test. PP is 512 timed inputs with 0 generated; TG is 128 serial tokens with 0 prompt and depth 0. These are separate tests, not a 512-in/128-out chat. Sortable headline is the first documented quiet baseline, not the fastest mode. Reported rates below are t/s, mean ± sample SD; raw UTC test_time and nanosecond aggregates are retained in the evidence CSV. - 2026-07-10T18:01:24Z / 2026-07-10T18:01:29Z: quiet, BIOS 1.06/IOMMU on; build 9910/f5525f7e7; PP 382.490359 ± 1.929762; TG 12.962264 ± 0.046976. - 2026-07-10T18:22:06Z / 2026-07-10T18:22:12Z: performance, BIOS 1.06/IOMMU on; build 9910/f5525f7e7; PP 418.012724 ± 4.829387; TG 13.169960 ± 0.004772. - 2026-07-12T06:36:15Z / 2026-07-12T06:36:20Z: balanced, BIOS 1.06/IOMMU on; build 1/049326a; PP 411.467557 ± 1.828297; TG 13.142035 ± 0.064260. - 2026-07-15T08:47:06Z / 2026-07-15T08:47:10Z: performance, BIOS 1.08/IOMMU off, run 1; build 1/049326a; PP 435.193356 ± 6.782130; TG 13.328797 ± 0.118509. - 2026-07-15T08:51:23Z / 2026-07-15T08:51:27Z: performance, BIOS 1.08/IOMMU off, run 2; build 1/049326a; PP 437.558764 ± 4.087500; TG 13.436544 ± 0.078118. Runtime caveat: July 10 quiet/performance use official b9910/f5525f7e7 HIP; July 12 balanced and July 15 BIOS-off cells log build 1/049326a, the source-described lemonade ROCm 7 gfx1151 line. The mode comparison is therefore not a single-runtime controlled experiment. BIOS and IOMMU changed together; neither isolated causality nor a generic speedup is established. Raw date/build data take precedence over prose that calls the pack b9910 throughout. Exact binaries/compiler hashes are unknown. Sources: https://github.com/daimonionnn/amd-rocmfpx-for-win/blob/7de1d467c9f2e8282051d9d2332f5ef584f4fa66/llm-bench/results/rocm-gemma3-27b_128GB.csv https://github.com/daimonionnn/amd-rocmfpx-for-win/blob/7de1d467c9f2e8282051d9d2332f5ef584f4fa66/llm-bench/README.md https://github.com/daimonionnn/amd-rocmfpx-for-win/blob/7de1d467c9f2e8282051d9d2332f5ef584f4fa66/llm-bench/results-strix-halo-rocm.md https://github.com/daimonionnn/amd-rocmfpx-for-win/blob/7de1d467c9f2e8282051d9d2332f5ef584f4fa66/llm-bench/Common.ps1 https://github.com/daimonionnn/amd-rocmfpx-for-win/blob/7de1d467c9f2e8282051d9d2332f5ef584f4fa66/llm-bench/Run-RocmTest.ps1 https://github.com/daimonionnn/amd-rocmfpx-for-win/blob/7de1d467c9f2e8282051d9d2332f5ef584f4fa66/llm-inference/README.md https://github.com/ggml-org/llama.cpp/blob/f5525f7e7/tools/llama-bench/llama-bench.cpp https://github.com/lemonade-sdk/llama.cpp/blob/049326a/tools/llama-bench/llama-bench.cpp https://huggingface.co/lmstudio-community/gemma-3-27B-it-qat-GGUF/blob/c29c7d93c128620e932db6a18604a25c4c40a17d/gemma-3-27B-it-QAT-Q4_0.gguf Structured rate fields are rounded to the form’s 0.1 token/s precision; exact source-reported means, standard deviations and all separate campaign cells are retained above.