Skip to content

Instantly share code, notes, and snippets.

@bjacob
Created September 11, 2026 20:55
Show Gist options
  • Select an option

  • Save bjacob/69ee844687c54c3603c7efc40fb783ce to your computer and use it in GitHub Desktop.

Select an option

Save bjacob/69ee844687c54c3603c7efc40fb783ce to your computer and use it in GitHub Desktop.

ConSan gfx1201 benchmark status

For each mode, Startup is the total latency through the first synchronized run and its selected evidence checkpoints, including instrumentation, loading, binding, and warm-up; Run is the absolute second-run latency followed by its ratio to the matching uninstrumented second run.

Workload Uninstrumented Startup Uninstrumented Run SuperCollider Startup SuperCollider Run RecordReplay Startup RecordReplay Run Sampled Startup Sampled Run InlineShadow Startup InlineShadow Run
PyTorch synthetic dense prefill (32-token prompt) 0.937 s 0.00209 s (1×) 132 s 0.00332 s (1.59×) 182 s 0.805 s (386×) 163 s 0.0092 s (4.41×) 159 s 0.431 s (207×)
PyTorch synthetic dense decode (one continuous-batch tick) 1.04 s 0.00193 s (1×) 134 s 0.00495 s (2.56×) 175 s 0.682 s (353×) 159 s 0.00748 s (3.87×) 159 s 0.448 s (232×)
PyTorch synthetic four-expert top-1 MoE prefill (16-token prompt) 1.23 s 0.00745 s (1×) 340 s 0.0124 s (1.66×) 442 s 3.18 s (427×) 412 s 0.0215 s (2.89×) 412 s 1.27 s (171×)
Gluon verified shared-memory round trip (1024 elements) 0.414 s 0.000113 s (1×) 0.404 s 0.000113 s (0.999×) 0.769 s 0.0163 s (144×) 0.435 s 0.000168 s (1.49×) 0.582 s 0.0118 s (104×)
hipBLASLt/Tensile verified FP16 GEMM (512×512×512) 0.0000165 s 0.0000134 s (1×) 0.521 s 0.0000387 s (2.9×) 1.25 s 0.0363 s (2,720×) 0.785 s 0.000127 s (9.52×) 0.961 s 0.06 s (4,490×)
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment