Recursive
Competition author · sol-execbench/user/Recursive · license unknown
Records held
107as of 2026-10-01 10:44 UTC
Entries
446
Runs
446
Is this you? Claim this project ›
A claim grants attribution and metadata maintenance, never the right to edit evidence. Evidence claims are reviewed by a maintainer.
Current records
Operation · workload
History
Record
Implementation
Hardware
Since
MoE complete layer with shared expert backwardsuite of 16 cases
143.9µs
6.9% faster · was 154.6 µs
B200
2026-06-07
Show all 40 records ›Showing all 40 records ⌄
NVFP4 MoE expert linear with gatingsuite of 16 cases
121.0µs
71.1% faster · was 419.3 µs
B200
2026-06-06
NVFP4 grouped query attention with KV repeatsuite of 16 cases
391.7µs
55.5% faster · was 880.4 µs
B200
2026-06-06
NVFP4 fused decoder layer pre norm attentionsuite of 16 cases
635.4µs
46.1% faster · was 1.18 ms
B200
2026-06-06
NVFP4 attention QKV projection with RoPEsuite of 16 cases
101.9µs
18.7% faster · was 125.3 µs
B200
2026-06-06
NVFP4 MoE routing with topk selectionsuite of 16 cases
36.8µs
16.3% faster · was 43.9 µs
B200
2026-06-06
NVFP4 attention output projection with residualsuite of 16 cases
75.2µs
52.8% faster · was 159.3 µs
B200
2026-06-06
FP8 multi latent attention QKV projectionsuite of 16 cases
456.5µs
28.7% faster · was 640.3 µs
B200
2026-06-06
Edit consistency loss with perceptual weightingsuite of 16 cases
516.1µs
23.1% faster · was 670.9 µs
B200
2026-06-07
Region aware self attention with edit bias backwardsuite of 16 cases
386.4µs
31.4% faster · was 563.0 µs
B200
2026-06-07
QK norm scaled dot product attentionsuite of 16 cases
92.9µs
26.4% faster · was 126.2 µs
B200
2026-06-06
Attention QKV with QK norm single kernel backwardsuite of 16 cases
24.5µs
32.3% faster · was 36.1 µs
B200
2026-06-06
Video spatial attention with RoPE 3dsuite of 16 cases
133.8µs
24.3% faster · was 176.8 µs
B200
2026-06-06
SAM hq vision attention with relative position backwardsuite of 16 cases
429.9µs
37.0% faster · was 682.1 µs
B200
2026-06-07
GELU approximate feedforward backwardsuite of 16 cases
218.8µs
2.7% faster · was 224.9 µs
B200
2026-06-07
Audio encoder conv positional layer stacksuite of 15 cases
15.8ms
6.4% faster · was 16.9 ms
B200
2026-06-07
Vision encoder layer with gated residualssuite of 16 cases
185.8µs
31.1% faster · was 269.8 µs
B200
2026-06-07
Text decoder layer with self attention and MLPsuite of 16 cases
682.0µs
20.1% faster · was 853.5 µs
B200
2026-06-07
MoE training token repeat and expert computationsuite of 16 cases
4.67ms
19.1% faster · was 5.77 ms
B200
2026-06-07
ConvNeXtV2 layer with NHWC persistence backwardsuite of 14 cases
181.2µs
48.4% faster · was 351.0 µs
B200
2026-06-07
67 more records · Open in the records ledger →
Measured entries
Implementation
Operation
Runtime
Best latency
Evidence
Availability
Show all 40 entries ›Showing all 40 entries ⌄
unknown
5.07µs
Reported
license unknown · no source
unknown
5.34µs
Reported
license unknown · no source
unknown
5.41µs
Reported
license unknown · no source
unknown
5.93µs
Reported
license unknown · no source
unknown
6.00µs
Reported
license unknown · no source
unknown
6.30µs
Reported
license unknown · no source
Showing 40 of 446; the full list is in the API dossier below.
Record activity
Nov
Dec
Jan
Feb
Mar
124Apr
110May
93Jun
33Jul
Aug
Sep
Oct
Sources: NVIDIA SOL-ExecBench (2026-07-14)last observed 2026-07-14JSON