Skip to contentSLQS001-GPUSiliq datasheet
GPU table
Median sustained score for each GPU model on each of the 5 instruments. A median shows after three verified runs on that instrument.
GPU models with a median verified run
How to read a median
A score is the sustained rate the tab held for the whole run, capped at the ceiling for its vendor family. Before launch, trial and genesis runs count toward a median. After launch, only paid runs do.
The same numbers come from /api/gpus. Every row behind them is in /api/dataset.
| Instrument | Unit | Workload |
|---|---|---|
| SLQ-01 compute | TFteraflops per second | chained fused multiply-add over a million vec4 lanes |
| SLQ-02 memory | GB/Sgigabytes per second | streaming upload, GPU copy and strided readback through 256 MB |
| SLQ-03 render | MP/Smegapaths per second | path tracer, 4 bounces, seeded sphere field over a ground plane |
| SLQ-04 matrix | TFteraflops per second | tiled SGEMM with workgroup memory, 1024 x 1024 |
| SLQ-05 inference | TFteraflops per second | transformer forward pass, 12 layers, 128 tokens, GPT-2 small shape |