Skip to content
KernelIndex
Search⌘K
Sign in

NVIDIA GeForce RTX 3090

sm_86
Records held
162
Runs
324
Operations
48 implementations

Records held on this GPU

Operation · workload
History
Record
Implementation
Since
Fused neighborhood attentionfp32 · [512] · seq 64 · dilation 2
597.0µs
PyTorch
2025-05-27
Fused neighborhood attentionfp32 · [512] · seq 128 · dilation 2
546.1µs
Liger
2025-05-27
Fused neighborhood attentionfp32 · [512] · seq 256 · dilation 2
556.1µs
PyTorch
2025-05-27
Fused neighborhood attentionfp32 · [512] · seq 512 · dilation 2
634.8µs
Liger
2025-05-27
Fused neighborhood attentionfp32 · [512] · seq 1024 · dilation 2
1.47ms
Liger
2025-05-27
Fused neighborhood attentionfp32 · [512] · seq 2048 · dilation 2
4.15ms
Liger
2025-05-27
Fused neighborhood attentionfp32 · [512] · seq 4096 · dilation 2
15.2ms
Liger
2025-05-27
Fused neighborhood attentionfp32 · [1024] · seq 64
2.62ms
Liger
2025-05-27
Fused neighborhood attentionfp32 · [1024] · seq 128
2.64ms
Liger
2025-05-27
Fused neighborhood attentionfp32 · [1024] · seq 256
2.61ms
Liger
2025-05-27
Fused neighborhood attentionfp32 · [1024] · seq 512
2.73ms
Liger
2025-05-27
Fused neighborhood attentionfp32 · [1024] · seq 1024
5.59ms
Liger
2025-05-27
Show all 40 records ›
Fused neighborhood attentionfp32 · [1024] · seq 2048
15.9ms
Liger
2025-05-27
Fused neighborhood attentionfp32 · [1024] · seq 4096
54.3ms
Liger
2025-05-27
Fused neighborhood attentionfp32 · [1024] · seq 64
493.6µs
PyTorch
2025-05-27
Fused neighborhood attentionfp32 · [1024] · seq 128
595.0µs
PyTorch
2025-05-27
Fused neighborhood attentionfp32 · [1024] · seq 256
785.4µs
Liger
2025-05-27
Fused neighborhood attentionfp32 · [1024] · seq 512
1.55ms
Liger
2025-05-27
Fused neighborhood attentionfp32 · [1024] · seq 1024
3.64ms
Liger
2025-05-27
Fused neighborhood attentionfp32 · [1024] · seq 2048
10.2ms
Liger
2025-05-27
Fused neighborhood attentionfp32 · [1024] · seq 4096
35.7ms
Liger
2025-05-27
Fused neighborhood attentionfp32 · [512] · seq 64 · dilation 2
170.0µs
Liger
2025-05-27
Fused neighborhood attentionfp32 · [512] · seq 128 · dilation 2
155.2µs
Liger
2025-05-27
Fused neighborhood attentionfp32 · [512] · seq 256 · dilation 2
170.0µs
Liger
2025-05-27
Fused neighborhood attentionfp32 · [512] · seq 512 · dilation 2
330.8µs
Liger
2025-05-27
Fused neighborhood attentionfp32 · [512] · seq 1024 · dilation 2
855.0µs
Liger
2025-05-27
Fused neighborhood attentionfp32 · [512] · seq 2048 · dilation 2
2.37ms
Liger
2025-05-27
Fused neighborhood attentionfp32 · [512] · seq 4096 · dilation 2
8.25ms
Liger
2025-05-27
Fused neighborhood attentionfp32 · [512] · seq 64 · dilation 2
2.76ms
Liger
2025-05-27
Fused neighborhood attentionfp32 · [512] · seq 128 · dilation 2
2.68ms
Liger
2025-05-27
Fused neighborhood attentionfp32 · [512] · seq 256 · dilation 2
2.67ms
Liger
2025-05-27
Fused neighborhood attentionfp32 · [512] · seq 512 · dilation 2
2.80ms
Liger
2025-05-27
Fused neighborhood attentionfp32 · [512] · seq 1024 · dilation 2
2.88ms
Liger
2025-05-27
Fused neighborhood attentionfp32 · [512] · seq 2048 · dilation 2
6.52ms
Liger
2025-05-27
Fused neighborhood attentionfp32 · [512] · seq 4096 · dilation 2
23.5ms
Liger
2025-05-27
Fused neighborhood attentionfp32 · [512] · seq 128 · dilation 1
555.0µs
PyTorch
2025-05-27
Fused neighborhood attentionfp32 · [512] · seq 256 · dilation 1
547.0µs
PyTorch
2025-05-27
Fused neighborhood attentionfp32 · [512] · seq 2048 · dilation 1
4.15ms
Liger
2025-05-27
Fused neighborhood attentionfp32 · [512] · seq 4096 · dilation 1
15.2ms
Liger
2025-05-27
Fused neighborhood attentionfp32 · [768] · seq 64
237.6µs
Liger
2025-05-27

122 more records · Open in the records ledger →

Coverage by operation family

Family
Operations
Runs
With source
neighborhood-attention
1
168
168
multi-token-attention
1
72
72
softmax
1
48
48
sparsemax
1
36
36

Record activity

No record transitions on this GPU in the trailing year.

Sources: Liger-Kernel benchmarks (2025-05-27) · BSD-2-Clauselast observed 2025-05-27