Per-SM Scheduling Model — Sample Tables

Addresses apply to ptxas v13.0.88 (CUDA 13.0). VA base 0x400000 (non-PIE).

Representative slices of the three per-SM scheduling table families. The full per-SM TSVs (11 dependency-rule sets, 3 latency families, 7 scoreboard configs) live in the repo at decoded/ptxas-sched-full/. See Latency Model and Scoreboards for the field semantics.

Dependency rule (40 B) — sm_90 vs sm_90a (first 12 classes + the WGMMA split)

rule_type 4 = disabled/unit-absent (always paired with latency=255). The six classes 41, 561, 562, 563, 566, 567 are disabled on sm_90 but active on sm_90a — the WGMMA / async-MMA tensor classes. This is the entire sm_90 vs sm_90a divergence.

sm	idx	unit	rule	lat	tput_inv	bar_lat	bar_tput	rd_lat	wr_lat	stall	slots
sm_90	0	2	1	17	0	56	8	-1	-1	0	1
sm_90	1	3	1	17	0	56	8	-1	-1	33	2
sm_90	2	4	1	42	15	56	4	-1	-1	1	1
sm_90	3	5	1	42	15	56	4	-1	-1	1	1
sm_90	4	6	1	42	15	56	4	-1	-1	1	1
sm_90	5	7	1	42	15	56	4	-1	-1	2	3
sm_90	6	8	1	42	15	56	4	-1	-1	2	3
sm_90	7	9	1	42	15	56	4	-1	-1	1	1
sm_90	8	10	1	42	15	56	4	-1	-1	4	1
sm_90	9	11	1	17	0	56	8	-1	6	1	1
sm_90	10	12	1	17	0	56	8	-1	-1	1	1
sm_90	11	15	0	22	2	56	2	-1	-1	39	4
sm_90	33	41	4	255	35	56	-1	-1	-1	39	4
sm_90	171	561	4	255	35	56	-1	-1	-1	39	4
sm_90	172	562	4	255	35	56	-1	-1	-1	39	4
sm_90	173	563	4	255	35	56	-1	-1	-1	39	4
sm_90	174	566	4	255	35	56	-1	-1	-1	39	4
sm_90	175	567	4	255	35	56	-1	-1	-1	39	4
sm_90a	0	2	1	17	0	56	8	-1	-1	0	1
sm_90a	1	3	1	17	0	56	8	-1	-1	33	2
sm_90a	2	4	1	42	15	56	4	-1	-1	1	1
sm_90a	3	5	1	42	15	56	4	-1	-1	1	1
sm_90a	4	6	1	42	15	56	4	-1	-1	1	1
sm_90a	5	7	1	42	15	56	4	-1	-1	2	3
sm_90a	6	8	1	42	15	56	4	-1	-1	2	3
sm_90a	7	9	1	42	15	56	4	-1	-1	1	1
sm_90a	8	10	1	42	15	56	4	-1	-1	4	1
sm_90a	9	11	1	17	0	56	8	-1	6	1	1
sm_90a	10	12	1	17	0	56	8	-1	-1	1	1
sm_90a	11	15	0	22	2	56	2	-1	-1	39	4
sm_90a	33	41	1	42	15	56	4	-1	-1	33	2
sm_90a	171	561	0	15	19	17	8	-1	-1	39	4
sm_90a	172	562	0	15	19	17	8	-1	-1	39	4
sm_90a	173	563	0	14	19	18	16	-1	-1	39	4
sm_90a	174	566	0	15	19	17	8	-1	-1	39	4
sm_90a	175	567	0	14	19	18	16	-1	-1	39	4

Latency / sched-class descriptor (72 B) — sm_8x family (first 16 classes)

p7_self equals class_id in every record (self-reference). p1_tput = throughput class, p5_maxstall = max-stall cycles, p11 always 0. pipeA/pipeB are per-pipe eligibility byte vectors (0xFF byte = pipe N/A).

idx	class_id	pipeA_hex	pipeB_hex	p0_flags	p1_tput	p2	p3	p4	p5_maxstall	p6	p7_self	p8	p9	p10
0	2	0303ffffffffffff	0000ffffffff0000	0	4	0	1	0	7	3	2	1	2	3
1	3	0303ffffffffffff	0000ffffffff0000	0	4	33	2	0	7	3	3	1	2	3
2	4	04040202ffffffff	0000ffffffff0000	0	4	1	1	3	7	3	4	1	2	3
3	5	04040202ffffffff	0000ffffffff0000	268435456	4	1	1	3	7	3	5	1	2	3
4	6	04040202ffffffff	0000ffffffff0000	268435456	4	1	1	3	7	3	6	1	2	3
5	7	04040202ffff0101	0000ffffffff0000	0	4	2	3	3	7	3	7	1	2	3
6	8	04040202ffff0101	0000ffffffff0000	0	4	2	3	3	7	3	8	1	2	3
7	9	04040202ffffffff	0000ffffffff0000	0	4	1	1	3	7	3	9	1	2	3
8	10	0404ffffffff0101	0000ffffffff0000	0	4	4	1	3	7	3	10	1	2	1
9	11	0404ffffffff0101	0000ffffffff0000	276824064	4	1	1	3	7	3	11	1	2	1
10	12	0404ffffffff0101	0000ffffffff0000	268439552	4	1	1	3	7	3	12	1	2	1
11	15	0000ffffffffffff	0000ffffffff0000	0	0	39	4	19	3	2	15	0	1	1
12	16	1212ffffffffffff	00000303ffff0000	0	4	13	1	14	6	1	16	3	1	1
13	17	1212ffffffffffff	00000303ffff0000	0	4	13	1	14	6	1	17	3	1	1
14	18	1313ffffffffffff	00000303ffff0000	0	4	13	1	14	6	1	18	3	1	1
15	19	1111ffffffffffff	00000303ffff0000	0	4	13	1	14	6	1	19	3	1	1

Scoreboard config — sm_100 (≤6 triplets) vs sm_90 (single triplet)

Each config is up to 6 {sb_id, threshold, mask} triplets. mask = -1 = unconditional wait; small masks (2/4/8/32) = pipe-specific. threshold 56 dominates. sm_100 (Blackwell) uses up to 6 scoreboards per config (async dependency-barrier model); sm_90 uses one triplet with a pipe mask.

sm_100 (first 5 configs)

idx	triplet_count	sb_id	threshold	mask
0	2	0	56	-1
0	2	2	56	-1
1	3	0	56	-1
1	3	2	56	-1
1	3	5	56	-1
2	4	0	56	-1
2	4	2	56	-1
2	4	5	56	-1
2	4	15	56	-1
3	6	0	56	-1
3	6	2	56	-1
3	6	12	56	-1
3	6	15	56	-1
3	6	23	56	-1
3	6	34	56	-1
4	3	0	56	-1
4	3	2	56	-1
4	3	15	56	-1

sm_90 (first 10 configs)

idx	triplet_count	sb_id	threshold	mask
0	1	0	56	8
1	1	2	56	2
2	1	5	56	4
3	1	6	56	8
4	1	12	8	2
5	1	12	9	1
6	1	12	12	4
7	1	12	13	2
8	1	13	56	8
9	1	15	56	4