Skip to content
KernelIndex
Search⌘K

Audio relative position attention

19 eligible runs
audio

Audio attention with relative position embeddings, local causal masking with configurable context windows, attention logit soft-capping, and block-based processing. Processes audio in chunks with overlapping context windows using sinusoidal position encoding.

Fastest reported
82.7µsmean · 91% of SOL
RecursiveLicense unknown · unknown

Reported evidence · last observed 2026-07-10. Reported by source; not independently reproduced. No public source. License unknown.

No public sourceRun detail →

Current records

Workloadsuite of 14 cases · mean latencyhead_dim = 64 · num_heads = 8 · chunk_size = 32 · hidden_size = 512 · max_forward = 0 · context_size = 160 · max_backward = 127 · num_timescales = 25614 cases

Not measured on H100 for this workload. Challenges →

Source-native comparison · GPU NVIDIA B200 · Workload suite of 14 cases · mean latency · Protocol SOL-ExecBench evaluation stack v1.1 · mean · 12 results · last observed 2026-08-17Record history →
#
Implementation
Latency
vs #1
Trust
Observed
1
82.7µs91% SOL
1.00×
Reported · license unknown · no source
2026-07-10

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

no sourcelicense unknownno install recipeRun detail →
2
87.6µs90% SOL
1.06×
Reported · license unknown · no source
2026-08-04

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

no sourcelicense unknownno install recipeRun detail →
3
102.8µs88% SOL
1.24×
Reported · license unknown · no source
2026-07-19

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

no sourcelicense unknownno install recipeRun detail →
4
113.9µs87% SOL
1.38×
Reported · license unknown · no source
2026-07-10

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

no sourcelicense unknownno install recipeRun detail →
5
114.6µs87% SOL
1.39×
Reported · license unknown · no source
2026-07-19

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

no sourcelicense unknownno install recipeRun detail →
6
119.7µs87% SOL
1.45×
Reported · license unknown · no source
2026-08-12

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

no sourcelicense unknownno install recipeRun detail →
7
131.8µs86% SOL
1.59×
Reported · license unknown · no source
2026-07-11

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

no sourcelicense unknownno install recipeRun detail →
8
136.1µs84% SOL
1.65×
Reported · license unknown · no source
2026-07-17

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

no sourcelicense unknownno install recipeRun detail →
9
156.9µs83% SOL
1.90×
Reported · license unknown · no source
2026-08-17

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

no sourcelicense unknownno install recipeRun detail →
10
188.2µs81% SOL
2.28×
Reported · license unknown · no source
2026-07-14

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

no sourcelicense unknownno install recipeRun detail →
11
332.1µs69% SOL
4.02×
Reported · license unknown · no source
2026-07-10

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

no sourcelicense unknownno install recipeRun detail →
12
541.9µs58% SOL
6.56×
Reported · license unknown · no source
2026-07-10

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

no sourcelicense unknownno install recipeRun detail →

Implementations

Implementation
Runtime
Best latency
Evidence
Availability
unknown
104.8µs
1.37×
Reported
license unknown · no source
unknown
114.6µs
1.49×
Reported
license unknown · no source
unknown
119.7µs
1.56×
Reported
license unknown · no source
unknown
131.8µs
1.72×
Reported
license unknown · no source
unknown
136.1µs
1.77×
Reported
license unknown · no source
unknown
456.1µs
5.94×
Reported
license unknown · no source
unknown
102.8µs
1.34×
Reported
license unknown · no source
unknown
113.9µs
1.48×
Reported
license unknown · no source
Show all 19 implementations ›
unknown
156.9µs
2.05×
Reported
license unknown · no source
unknown
188.2µs
2.45×
Reported
license unknown · no source
unknown
266.0µs
3.47×
Reported
license unknown · no source
unknown
317.7µs
4.14×
Reported
license unknown · no source
unknown
332.1µs
4.33×
Reported
license unknown · no source
unknown
541.9µs
7.06×
Reported
license unknown · no source
unknown
639.9µs
8.34×
Reported
license unknown · no source
unknown
816.1µs
10.64×
Reported
license unknown · no source
unknown
76.7µs
1.00×
Reported
license unknown · no source
unknown
82.7µs
1.08×
Reported
license unknown · no source
unknown
87.6µs
1.14×
Reported
license unknown · no source

Semantics

Inputs and outputs
hidden_statesfp32 [batch_size, seq_len, hidden_size]
maskbool [batch_size, seq_len]
q_proj_weightfp32 [hidden_size, hidden_size]
k_proj_weightfp32 [hidden_size, hidden_size]
v_proj_weightfp32 [hidden_size, hidden_size]
pos_proj_weightfp32 [hidden_size, hidden_size]
per_dim_scalefp32 [head_dim]
inv_timescalesfp32 [num_timescales]
attention_logits_soft_capfp32 scalar
context_vectorsfp32 [batch_size, seq_len, num_heads, head_dim]
Axes and behavior
seq_lenvariable
head_dimconstant = 64
num_headsconstant = 8
batch_sizevariable
chunk_sizeconstant = 32
hidden_sizeconstant = 512
max_forwardconstant = 0
context_sizeconstant = 160
max_backwardconstant = 127
num_timescalesconstant = 256
determinismunspecified
constraintsNo mutation or aliasing
Identity
alias038_audio_relative_position_attentionmodelgemma-3n-e2bsha25612572dcdcc91…
Sources: NVIDIA SOL-ExecBench (2026-08-17)last observed 2026-08-17How records are decidedJSON