modelrigs

Benchmark database

Every measured run we have, stored individually rather than collapsed into one number per pairing. Where sources disagree — and on Strix Halo they disagree by nearly a factor of two — you see all of them.

Measured runs
15
from published sources
Estimated pairings
1,163
generated by estimator_v1
Sources
12
GitHub, blogs, forums
Coverage
1.3%
of all pairings measured
Measured benchmark density is this database’s biggest weakness. 15 real runs against 1,163 estimated pairings means most intersection pages show a calculated range rather than a measurement. Those pages say so plainly. Benchmark contributions are the highest-value thing anyone could add.
Framework Desktop (Ryzen AI Max+ 395, 128 GB)
ModelQuantEngineContextDecodePrefillTypeDateSource
Qwen3-Coder 30B-A3BQ4_K_Mllama.cpp4K52.00 t/s—Measured (community)10 Feb 2026blog
Qwen3.6 27BQ4_K_MLM Studio4K11.50 t/s—Measured (community)10 Feb 2026blog
gpt-oss-120bMXFP4LM Studio4K30.00 t/s—Measured (community)10 Feb 2026blog
gpt-oss-120bMXFP4llama.cpp4K55.00 t/s—Measured (community)15 Mar 2026blog
gpt-oss-120bMXFP4llama.cpp4K31.00 t/s—Measured (community)2 Apr 2026blog
gpt-oss-120bMXFP4llama.cpp4K34.00 t/s—Measured (community)11 May 2026blog
  • Reported as a 35B-class MoE with 3B active exceeding 50 t/s with a draft model.
  • A dense 27B on Strix Halo: roughly a third the speed of a 120B MoE on the same box. The clearest single illustration of why active parameter count matters more than total size for decode.
  • LM Studio with the ROCm llama.cpp runtime.
  • Substantially higher than other Strix Halo reports for the same model. Retained deliberately: the disagreement between sources is real information and is reflected in the confidence score rather than averaged away.
GMKtec EVO-X2 (Ryzen AI Max+ 395, 128 GB)
ModelQuantEngineContextDecodePrefillTypeDateSource
gpt-oss-120bMXFP4llama.cpp4K32.00 t/s—Measured (community)20 Apr 2026github
  • Same silicon as the Framework Desktop; results track within noise, which is the expected outcome and a useful sanity check on our hardware model.
NVIDIA DGX Spark 128 GB
ModelQuantEngineContextDecodePrefillTypeDateSource
gpt-oss-120bMXFP4llama.cpp
b7db35a7
060.57 t/s1,956 t/sMeasured (community)20 Oct 2025github
gpt-oss-120bMXFP4llama.cpp058.70 t/s1,723 t/sMeasured (independent)25 Oct 2025blog
gpt-oss-120bMXFP4llama.cpp
b7db35a7
4K54.14 t/s1,637 t/sMeasured (community)20 Oct 2025github
gpt-oss-120bMXFP4llama.cpp
b7db35a7
8K51.54 t/s1,512 t/sMeasured (community)20 Oct 2025github
gpt-oss-120bMXFP4llama.cpp
b7db35a7
16K47.45 t/s1,307 t/sMeasured (community)20 Oct 2025github
gpt-oss-120bMXFP4llama.cpp
b7db35a7
32K40.55 t/s1,027 t/sMeasured (community)20 Oct 2025github
  • llama-bench at zero context depth: the best case, and the number most often quoted in marketing. Reported as 60.57 +/- 0.25.
  • Independent reproduction. Slightly below the llama.cpp thread figure, which is normal for a different kernel build.
  • Same run at 4K context depth. Reported as 54.14 +/- 0.08.
  • 8K context depth. Reported as 51.54 +/- 0.14.
  • 16K context depth. Reported as 47.45 +/- 0.08.
  • 32K context depth — decode has fallen 33% from the zero-context figure. This is why the site buckets by context rather than quoting one number.
RTX 5090 workstation (1x 32 GB)
ModelQuantEngineContextDecodePrefillTypeDateSource
Qwen3 30B-A3BQ4_K_Mllama.cpp0352.06 t/s—Measured (community)30 Nov 2025github
gpt-oss-20bMXFP4llama.cpp0419.14 t/s—Measured (community)30 Nov 2025github
  • llama-bench tg128, Qwen3 30B-A3B Q4_K_M (17.28 GiB), full GPU offload, flash attention enabled and GGML_CUDA_GRAPH_OPT=1. Reported as 352.06 +/- 1.76 t/s. The source does not pin a llama.cpp commit, which caps confidence.
  • llama-bench tg128, native gpt-oss 20B MXFP4 (11.27 GiB), full GPU offload, flash attention enabled and GGML_CUDA_GRAPH_OPT=1. Reported as 419.14 +/- 0.89 t/s. The source does not pin a llama.cpp commit, which caps confidence.
Sources
TitleTypeAuthorAccessedNotes
AMD Ryzen AI Max+ 395 (Strix Halo) for Local LLMs in 2026blogRunAIHome6 Sept 2026Community review with throughput figures for MoE models on Strix Halo.
GPT-OSS 120B & 20B on AMD Ryzen AI Max+ 395blogakehir6 Sept 2026Kubernetes-hosted llama.cpp on Strix Halo. Reports a materially higher figure than other Strix Halo sources, which is preserved rather than smoothed away.
Local LLM Speed Test: GPT-OSS, Qwen3.6 and Hermes on 128GB Unified MemoryblogMindStudio6 Sept 2026LM Studio with the ROCm llama.cpp runtime on a 128 GB Ryzen AI Max+ 395.
Strix Halo Tokens Per Second 2026: Real Speed, Dense vs MoEblogDataHardware6 Sept 2026Aggregated Strix Halo measurements across dense and MoE models.
llama.cpp on NVIDIA DGX Spark — BenchmarksblogJetsonHacks6 Sept 2026Independent reproduction of the llama.cpp DGX Spark benchmarks.
Strix Halo (Ryzen AI Max+ 395) LLM Benchmark ResultsforumLevel1Techs forum6 Sept 2026Community benchmark thread with a wide spread of results depending on backend (Vulkan vs ROCm).
AMD Strix Halo — Backend Benchmarks (Grid View)githubkyuz06 Sept 2026Systematic backend comparison grid for Strix Halo. One of the few Strix Halo sources with full method documentation.
Optimizing Token Generation in llama.cpp's CUDA Backendgithubllama.cpp contributors7 Sept 2026Published llama-bench decode results for Qwen3 30B-A3B Q4_K_M and gpt-oss 20B MXFP4 on RTX 5090. The source gives the command, backend flags, model sizes and repeated-result dispersion, but not a pinned llama.cpp commit, so confidence remains community-level.
Performance of llama.cpp on Apple Silicon M-seriesgithubllama.cpp contributors6 Sept 2026The long-running Apple Silicon llama-bench collection. Uses Llama 7B as a fixed reference model across chips, which makes it excellent for calibrating a hardware bandwidth coefficient and poor for predicting any specific modern model.
Performance of llama.cpp on NVIDIA CUDAgithubllama.cpp contributors7 Sept 2026Maintained llama-bench CUDA scoreboard. The imported RTX 5090 and RTX PRO 6000 Llama 7B Q4_0 rows retain their individual llama.cpp commits and are calibration-only because the reference model is not in the catalogue.
Performance of llama.cpp on NVIDIA DGX Sparkgithubllama.cpp contributors6 Sept 2026Maintained benchmark thread with llama-bench output at several context depths. Build 7db35a7. Numbers include the reported standard deviation, which is why the confidence weight is high for a community source.
Performance of llama.cpp with Vulkangithubllama.cpp contributors7 Sept 2026Maintained Vulkan llama-bench scoreboard with a standardized Llama 7B Q4_0 workload. The R9700 row is used only for hardware calibration.