Efficiency benchmark · one RTX 4080 SUPER · updated 2026-09-24 11:24 UTC

What each model costs

6 datasets
10 models
single GPU
published configs

The same measurements efficiency-focused recommender papers publish — trainable parameters, peak GPU memory, training time per epoch, and inference time per batch — taken on one idle GPU with each model in its exact published main-table configuration. Timings are medians over repeated, synchronized steps; the last column is the real wall-clock of the published three-seed runs.

Takeaway

The notepad's inference cost sits between the light recurrent models and the attention stack, with the smallest inference memory of any adaptive model — its per-user state is bounded, so serving memory stays flat while attention models grow with history. Training is mid-pack: the write loop makes epochs slower than GRU4Rec's but comparable to the state-space and contrastive baselines. The fastest trainer everywhere is GRU4Rec; the slowest is Mamba4Rec's unfused scan.

MIND-large · batch 1024

Model	Trainable params	Train s/epoch	Peak train GB	Infer ms/batch	Peak infer GB	Total train min†
Popularity	0	·	·	0.04	0.004	3.7
ItemKNN	0	·	·	942.39	0.003	15.7
BPRMF	48.03M	13.06	0.921	0.11	0.409	12.1
GRU4Rec	25K	8.96	0.447	2.32	0.078	10.6
SASRec	103K	21.72	0.621	4.18	0.181	11.2
DuoRec	103K	49.71	1.653	3.97	0.693	14.9
LRURec	101K	87.05	1.265	4.72	0.242	22.9
Mamba4Rec	66K	100.54	1.718	6.65	0.282	18.7
HSTU	62K	28.83	0.663	3.54	0.273	9.5
FWPRec (Ours)	83K	50.41	1.509	10.86	0.191	16.1

EB-NeRD · batch 1024

Model	Trainable params	Train s/epoch	Peak train GB	Infer ms/batch	Peak infer GB	Total train min†
Popularity	0	·	·	0.05	0.003	0.9
ItemKNN	0	·	·	1015.77	0.003	4.4
BPRMF	1.20M	0.27	0.071	0.09	0.039	1.5
GRU4Rec	25K	1.06	0.08	1.76	0.074	1.3
SASRec	103K	2.28	0.599	4.05	0.159	1.4
DuoRec	103K	5.2	1.631	3.91	0.671	1.2
LRURec	101K	8.9	1.243	4.56	0.22	1.7
Mamba4Rec	66K	10.09	1.696	6.48	0.26	1.7
HSTU	62K	2.86	0.642	3.41	0.251	1.3
FWPRec (Ours)	83K	5.15	1.487	10.78	0.169	2.2

Amazon Video Games · batch 256

Model	Trainable params	Train s/epoch	Peak train GB	Infer ms/batch	Peak infer GB	Total train min†
Popularity	0	·	·	0.03	0.002	0.6
ItemKNN	0	·	·	431.62	0.002	1.2
BPRMF	6.06M	1.18	0.121	0.08	0.065	1.2
GRU4Rec	25K	3.94	0.092	1.59	0.028	1.3
SASRec	106K	6.08	0.332	2.33	0.124	1.5
DuoRec	106K	14.52	0.874	2.21	0.398	2.6
LRURec	101K	24.24	0.653	4.98	0.123	4.1
Mamba4Rec	66K	32.52	0.859	8.36	0.138	6.1
HSTU	65K	10.97	0.41	2.41	0.173	2.3
FWPRec (Ours)	83K	14.48	0.399	3.34	0.069	2.6

MovieLens-1M · batch 256

Model	Trainable params	Train s/epoch	Peak train GB	Infer ms/batch	Peak infer GB	Total train min†
Popularity	0	·	·	0.03	0.001	1.2
ItemKNN	0	·	·	524.48	0.001	4.4
BPRMF	387K	2.03	0.053	0.07	0.017	2.9
GRU4Rec	25K	8.71	0.053	2.58	0.024	3.2
SASRec	103K	8.85	0.164	1.48	0.053	4.0
DuoRec	103K	21.58	0.42	1.38	0.184	4.7
LRURec	101K	31.11	0.328	3.57	0.069	7.0
Mamba4Rec	66K	44.25	0.438	4.17	0.078	8.2
HSTU	62K	10.53	0.174	1.55	0.076	3.5
FWPRec (Ours)	83K	32.03	0.39	3.49	0.06	7.2

KuaiRand-Pure (IDs) · batch 512

Model	Trainable params	Train s/epoch	Peak train GB	Infer ms/batch	Peak infer GB	Total train min†
Popularity	0	·	·	0.04	0.003	2.5
ItemKNN	0	·	·	567.19	0.003	2.8
BPRMF	2.22M	1.31	0.053	0.12	0.03	3.9
GRU4Rec	516K	5.03	0.069	2.4	0.047	4.9
SASRec	597K	11.99	0.64	4.06	0.221	5.5
DuoRec	597K	27.09	1.721	4.06	0.765	6.2
LRURec	592K	71.68	1.279	5.68	0.22	19.4
Mamba4Rec	557K	68.73	1.699	7.69	0.249	20.0
HSTU	556K	20.57	0.798	4.64	0.32	8.6
FWPRec (Ours)	567K	13.13	0.763	3.35	0.102	6.0

KuaiRec · batch 256

Model	Trainable params	Train s/epoch	Peak train GB	Infer ms/batch	Peak infer GB	Total train min†
Popularity	0	·	·	0.04	0.002	4.8
ItemKNN	0	·	·	619.25	0.002	27.1
BPRMF	90K	5.23	0.049	0.13	0.015	6.5
GRU4Rec	25K	27.08	0.057	2.89	0.029	7.8
SASRec	106K	30.89	0.326	2.54	0.119	9.0
DuoRec	106K	73.37	0.869	2.27	0.393	14.0
LRURec	101K	142.0	0.648	4.95	0.118	26.0
Mamba4Rec	66K	248.29	0.854	8.33	0.132	35.7
HSTU	65K	46.4	0.405	2.33	0.168	9.3
FWPRec (Ours)	83K	96.52	0.393	3.3	0.064	12.4

Fine print

† mean wall-clock of the three published final seeds, including per-epoch validation and early stopping — so it reflects what the runs actually cost, not epochs × epoch-time. Trainable parameters exclude the frozen Qwen table (identical for every model on content columns); on KuaiRand all item tables are learned, so counts include them. Popularity and ItemKNN involve no gradient training; their fit happens at construction. Batch sizes are each column's published training batch size.