kernbench2/docs/diagrams/allreduce_latency_plots/buffer_kind_sweep.csv at a44f832be5939b7b9a218e40c36ec72fee66f1a0

Files

T

ywkang a44f832be5 Regenerate latency plots/diagrams for post-Phase-2c model

Allreduce + pe2pe + ipcq + pe_view auto-regenerated by test sweeps
running against the new chunk-streaming wire timing (per-flit
wormhole) — absolute numbers shift upward to reflect bottleneck-link
transit charged once per flit (instead of the previous cut-through
subtraction at HBM CTRL).

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

2026-05-14 23:24:01 -07:00

602 B

Raw Blame History

1	buffer_kind	sip_topology	n_sips	n_elem	bytes_per_pe	latency_ns
2	hbm	torus_2d	6	128	256	2144.0399999999754
3	hbm	torus_2d	6	1024	2048	2908.74499999995
4	hbm	torus_2d	6	8192	16384	8851.185000000081
5	hbm	torus_2d	6	32768	65536	29225.265000008752
6	sram	torus_2d	6	128	256	2060.0399999999754
7	sram	torus_2d	6	1024	2048	2908.74499999995
8	sram	torus_2d	6	8192	16384	9523.185000000081
9	sram	torus_2d	6	32768	65536	32201.265000008752
10	tcm	torus_2d	6	128	256	1964.0399999999754
11	tcm	torus_2d	6	1024	2048	2476.74499999995
12	tcm	torus_2d	6	8192	16384	6403.185000000081
13	tcm	torus_2d	6	32768	65536	19865.265000008738

602 B Raw Blame History

602 B

Raw Blame History