kernbench2/docs/diagrams/allreduce_latency_plots/summary.csv at 4ded67a44371b1e5f046658d9d87ff9d455aa6cd

Files

T

ywkang b6315c3c90 paper(1H): reflect D8 single-op cost model in §2/§3.4 + figure/diagram regen

Two strands bundled as the 1H-codesign-paper refresh unit:

(A) This session — single-op cost-model reflection (depends on 2d8271c):
  - §2 Table 2 (tab:hw): split "FIXED per command" into "FIXED per
    single-op command" (8 cycles) and "FIXED per composite command"
    (40 cycles); §2 dispatch-overhead prose updated to the two-class split.
  - §3.4 (sec:gemm-vs-async): rename paragraph headers + prose to
    async-full / async-tiled; "atomic" -> "single-op" throughout; reframe
    mechanism #3 from the old DMA-only fast-path to the single-op
    fast-path. Headline narrative now: even with EVERY single-op cmd
    (96 DMA + 48 dot + 47 add) charged the light 8-cycle FIXED, composite
    still wins ~2.8x at K=3072 purely on command-count structure (1 vs
    192 commands) -- down from the pre-D8 ~6.3x, and explicitly NOT a
    modelling artifact. Numbers refreshed from the regenerated sweep:
    async-full 3.83->3.91, async-tiled 1.14->~2.53, under-tile corner
    1.06->1.21, depth-2 vs depth-inf spread <1%. New figure wired in.
  - build/main.pdf rebuilt (tectonic); pdftotext-verified (no broken
    refs; Table 2 split, single-op terms, 2.8x/2.53/192-host-commands
    all present).

(B) Prior-session paper work riding along uncommitted: §4 all-reduce
    deep-edit, §5 GQA, §6 discussion trims; milestone_1h_ccl.py plot
    label "FSIM" -> "H2 2025 SW queue baseline"; regenerated diagrams
    under docs/diagrams/** and gemm output PNGs under
    1H_milestone_output/gemm/. (Composite-window gemm plots are
    unaffected by D8 — D8 only changes single-op dispatch FIXED, which
    the composite window excludes.)

All TODO items for the D8 single-op extension are now complete and
pushed across 3 commits (2d8271c cost-model+ADR+tests, 821bbf2 bench
harness, this paper refresh). Full regression green (826 passed, 1
skipped). No remaining work.

NOTE for review (carried from 2d8271c): ADR-0065's "2x CPU-offload win"
headline for GQA decode opt2 may want a refresh to the post-D8 ~1.87x.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

2026-06-17 15:12:31 -07:00

2.7 KiB

Raw Blame History

1	algorithm	sip_topology	n_sips	n_elem	bytes_per_pe	bytes_per_sip	latency_ns
2	lrab_hierarchical_allreduce	mesh_2d_no_wrap	6	8	16	256	2918.5525000000157
3	lrab_hierarchical_allreduce	mesh_2d_no_wrap	6	32	64	1024	2999.740000000016
4	lrab_hierarchical_allreduce	mesh_2d_no_wrap	6	64	128	2048	3107.990000000019
5	lrab_hierarchical_allreduce	mesh_2d_no_wrap	6	128	256	4096	3324.4900000000207
6	lrab_hierarchical_allreduce	mesh_2d_no_wrap	6	512	1024	16384	3589.1133333333582
7	lrab_hierarchical_allreduce	mesh_2d_no_wrap	6	1024	2048	32768	3960.0333333333692
8	lrab_hierarchical_allreduce	mesh_2d_no_wrap	6	2048	4096	65536	4701.873333333393
9	lrab_hierarchical_allreduce	mesh_2d_no_wrap	6	4096	8192	131072	6185.020000000124
10	lrab_hierarchical_allreduce	mesh_2d_no_wrap	6	8192	16384	262144	9152.379999999861
11	lrab_hierarchical_allreduce	mesh_2d_no_wrap	6	16384	32768	524288	15087.099999999224
12	lrab_hierarchical_allreduce	mesh_2d_no_wrap	6	32768	65536	1048576	26956.540000000765
13	lrab_hierarchical_allreduce	mesh_2d_no_wrap	6	49152	98304	1572864	38825.97999999701
14	lrab_hierarchical_allreduce	ring_1d	6	8	16	256	2628.2558333333477
15	lrab_hierarchical_allreduce	ring_1d	6	32	64	1024	2699.943333333348
16	lrab_hierarchical_allreduce	ring_1d	6	64	128	2048	2795.526666666683
17	lrab_hierarchical_allreduce	ring_1d	6	128	256	4096	2986.693333333351
18	lrab_hierarchical_allreduce	ring_1d	6	512	1024	16384	3311.635000000021
19	lrab_hierarchical_allreduce	ring_1d	6	1024	2048	32768	3656.4016666666957
20	lrab_hierarchical_allreduce	ring_1d	6	2048	4096	65536	4345.401666666714
21	lrab_hierarchical_allreduce	ring_1d	6	4096	8192	131072	5721.801666666768
22	lrab_hierarchical_allreduce	ring_1d	6	8192	16384	262144	8479.934999999887
23	lrab_hierarchical_allreduce	ring_1d	6	16384	32768	524288	13996.201666665835
24	lrab_hierarchical_allreduce	ring_1d	6	32768	65536	1048576	25028.73500000064
25	lrab_hierarchical_allreduce	ring_1d	6	49152	98304	1572864	36061.26833333154
26	lrab_hierarchical_allreduce	torus_2d	6	8	16	256	1925.6025000000104
27	lrab_hierarchical_allreduce	torus_2d	6	32	64	1024	1978.290000000011
28	lrab_hierarchical_allreduce	torus_2d	6	64	128	2048	2048.540000000012
29	lrab_hierarchical_allreduce	torus_2d	6	128	256	4096	2189.040000000012
30	lrab_hierarchical_allreduce	torus_2d	6	512	1024	16384	2421.8183333333463
31	lrab_hierarchical_allreduce	torus_2d	6	1024	2048	32768	2702.2783333333473
32	lrab_hierarchical_allreduce	torus_2d	6	2048	4096	65536	3263.1983333333583
33	lrab_hierarchical_allreduce	torus_2d	6	4096	8192	131072	4384.5050000000665
34	lrab_hierarchical_allreduce	torus_2d	6	8192	16384	262144	6628.18500000005
35	lrab_hierarchical_allreduce	torus_2d	6	16384	32768	524288	11115.5449999995
36	lrab_hierarchical_allreduce	torus_2d	6	32768	65536	1048576	20090.265000000378
37	lrab_hierarchical_allreduce	torus_2d	6	49152	98304	1572864	29064.985000000597

2.7 KiB Raw Blame History

2.7 KiB

Raw Blame History