Pull requests / #2951
#2951 ship a gfx1031 rocBLAS tuning table (RX 6700 XT)
open · @matthewjt · 0 comments · View on GitHub
BenchmarksSetup & installAMD / HIPModels & quantsDocumentation
Description
Summary Adds tools/hip/gfx1031-rocblas-5.4.0.ad84daa81c.txt, measured with tools/hip/tune_rocblas on an RX 6700 XT under the pip rocm-sdk 7.13 stack (rocBLAS 5.4.0.ad84daa81c). gfx1031 (Navi 21: RX 6700 XT / 6600 XT / 6650 XT) previously had no table, so setup left those engines on rocBLAS's default kernel picks. With the table, setup auto-points gfx1031 engines at the directory; the engine log reports rocBLAS tuning enabled (112 rows, 80 with a solution, gfx1031, rocBLAS 5.4.0.ad84daa81c). Results (RX 6700 XT, Swift IQ3_S, 262K ctx, membind, 3× bench medians, STRATA_HIP_PROMPT_F16=1) - 4K prompt: 579.7 → 625 tok/s (+7.8%); 4K TTFT 6.98 → 6.46 s - 16K prompt: 724 → 734 tok/s (+1.4%); 32K: 720.7 → 725.3 - Decode unchanged (the table touches only the prompt-path FP16 GEMMs) - Per-shape wins up to 2.5× (GDN input projection below the T = 1152 flip), all solution-checked against the default kernel's output (padding_writes=0) Contents - tools/hip/gfx1031-rocblas-5.4.0.ad84daa81c.txt — 112 rows, 80 winning solutions - docs/AMD_HIP.md: one line in the shipped-tables list
Related on this site
Quick links to related pages.