Skip to content
KernelIndex
Search⌘K

Gated DeltaNet chunk_fwd_o

15 eligible runs
gated-deltanet

The chunk_fwd_o output computation of Gated DeltaNet (arXiv:2412.06464): per 64-timestep chunk, combine the inter-chunk contribution q @ h * exp(g) with the causally masked intra-chunk attention over v_new, scaled by k^(-0.5). All tensors are fp32; t must be a multiple of 64.

Fastest reported
5.86µsgeomean
submission 555616fluudgate · License unknown · python

Reported evidence · last observed 2026-03-15. Reported by source; not independently reproduced. License unknown.

Current records

Not measured on H100 for this workload. Challenges →

Source-native comparison · GPU NVIDIA B200 · Workload suite of 3 cases · geomean score · Protocol reference-kernels eval.py · geomean · 15 results · last observed 2026-03-15Record history →
#
Implementation
Latency
vs #1
Trust
Observed
1
5.86µs
1.00×
Reported · license unknown · source
2026-03-15stale

Measured exactly what you asked. Reported by source; not independently reproduced.

source mirroredlicense unknownno install recipeView source →Run detail →
2
18.0µs
3.07×
Reported · license unknown · source
2026-03-14stale

Measured exactly what you asked. Reported by source; not independently reproduced.

source mirroredlicense unknownno install recipeView source →Run detail →
3
18.8µs
3.20×
Reported · license unknown · source
2026-03-14stale

Measured exactly what you asked. Reported by source; not independently reproduced.

source mirroredlicense unknownno install recipeView source →Run detail →
4
18.8µs
3.21×
Reported · license unknown · source
2026-03-14stale

Measured exactly what you asked. Reported by source; not independently reproduced.

source mirroredlicense unknownno install recipeView source →Run detail →
5
19.0µs
3.24×
Reported · license unknown · source
2026-03-15stale

Measured exactly what you asked. Reported by source; not independently reproduced.

source mirroredlicense unknownno install recipeView source →Run detail →
6
submission 555463pradeep03071
20.1µs
3.44×
Reported · license unknown · source
2026-03-15stale

Measured exactly what you asked. Reported by source; not independently reproduced.

source mirroredlicense unknownno install recipeView source →Run detail →
7
submission 553929bloomberg9383
20.3µs
3.47×
Reported · license unknown · source
2026-03-14stale

Measured exactly what you asked. Reported by source; not independently reproduced.

source mirroredlicense unknownno install recipeView source →Run detail →
8
submission 553224the.defenestrator
21.4µs
3.65×
Reported · license unknown · source
2026-03-14stale

Measured exactly what you asked. Reported by source; not independently reproduced.

source mirroredlicense unknownno install recipeView source →Run detail →
9
21.8µs
3.73×
Reported · license unknown · source
2026-03-14stale

Measured exactly what you asked. Reported by source; not independently reproduced.

source mirroredlicense unknownno install recipeView source →Run detail →
10
21.9µs
3.74×
Reported · license unknown · source
2026-03-14stale

Measured exactly what you asked. Reported by source; not independently reproduced.

source mirroredlicense unknownno install recipeView source →Run detail →
11
submission 552582bloomberg9383
21.9µs
3.74×
Reported · license unknown · source
2026-03-14stale

Measured exactly what you asked. Reported by source; not independently reproduced.

source mirroredlicense unknownno install recipeView source →Run detail →
12
submission 555438happy_sloth_
22.7µs
3.88×
Reported · license unknown · source
2026-03-15stale

Measured exactly what you asked. Reported by source; not independently reproduced.

source mirroredlicense unknownno install recipeView source →Run detail →
13
23.2µs
3.96×
Reported · license unknown · source
2026-03-15stale

Measured exactly what you asked. Reported by source; not independently reproduced.

source mirroredlicense unknownno install recipeView source →Run detail →
14
submission 555436suvasis_29047
50.4µs
8.60×
Reported · license unknown · source
2026-03-15stale

Measured exactly what you asked. Reported by source; not independently reproduced.

source mirroredlicense unknownno install recipeView source →Run detail →
15
60.7µs
10.35×
Reported · license unknown · source
2026-03-14stale

Measured exactly what you asked. Reported by source; not independently reproduced.

source mirroredlicense unknownno install recipeView source →Run detail →

Implementations

Implementation
Runtime
Best latency
Evidence
Availability
python
5.86µs
1.00×
Reported
license unknown · source
python
18.0µs
3.07×
Reported
license unknown · source
python
18.8µs
3.20×
Reported
license unknown · source
python
19.0µs
3.24×
Reported
license unknown · source
submission 555463pradeep03071
python
20.1µs
3.44×
Reported
license unknown · source
submission 553929bloomberg9383
python
20.3µs
3.47×
Reported
license unknown · source
submission 553224the.defenestrator
python
21.4µs
3.65×
Reported
license unknown · source
python
21.8µs
3.73×
Reported
license unknown · source
Show all 15 implementations ›
python
21.9µs
3.74×
Reported
license unknown · source
submission 555438happy_sloth_
python
22.7µs
3.88×
Reported
license unknown · source
python
23.2µs
3.96×
Reported
license unknown · source
submission 555436suvasis_29047
python
50.4µs
8.60×
Reported
license unknown · source
python
60.7µs
10.35×
Reported
license unknown · source
python
18.8µs
3.21×
Reported
license unknown · source
submission 552582bloomberg9383
python
21.9µs
3.74×
Reported
license unknown · source

Semantics

Inputs and outputs
qfp32 [b, t, h, k]
kfp32 [b, t, h, k]
v_newfp32 [b, t, h, v]
hfp32 [b, nt, h, k, v]
gfp32 [b, t, h]
outputfp32 [b, t, h, v]
Axes and behavior
bvariable
hvariable
kvariable
tvariable
vvariable
btconstant = 64
ntnt = t // 64
seedvariable
determinismunspecified
constraintsNo mutation or aliasing
Identity
aliasgated deltanet chunk_fwd_oaliasgated_deltanet_chunk_fwd_osha256f57dca102547…
Sources: GPU MODE KernelBot (2026-03-15) · June 9 Researcher Reciprocity License v1.0last observed 2026-03-15How records are decidedJSON