# Attention Mechanism Latency Benchmark on Synthetic sequences Source: [Attention Mechanism Latency Benchmark on Synthetic sequences]() Collection index: [llms.txt]() Data updated: 2026-02-26T05:48:29.200Z Evaluation of the latency performance of the Attention Mechanism when tested on synthetic sequence data, reporting metrics like latency in milliseconds. A research result applies to its stated dataset, split, metric and evaluation configuration. Compare like-for-like settings, preserve the metric's unit and direction, and distinguish paper-reported results from independent reproduction. The best result recorded here is not a claim of exhaustive, current global coverage. | Scope | Recorded value | | --- | --- | | Task | Attention Mechanism Latency Benchmark | | Dataset (including any stated split) | Synthetic sequences | | Metric | Latency (ms) | | Unit | TIME | | Direction | Lower is better | | Results loaded | 16 of 16 | | Coverage | All results currently returned for this benchmark | Evaluation settings may differ between results. The original papers and recorded configurations are needed to establish comparability. A shared metric name alone does not establish an identical protocol. ## Metrics - [Latency (ms)](): Lower is better; unit: TIME. - [Speedup](): Higher is better; unit: SCALAR. - [Sparsity](): Higher is better; unit: PERCENT. ## Recorded results | Method | Latency (ms) | Result date | Report paper | | --- | --- | --- | --- | | Sparse Triton Kernel | 0.03 | Not recorded | 69896a2fc97a0b3e3b5dc60c | | Sparse Triton Kernel | 0.03 | Not recorded | 69896a2fc97a0b3e3b5dc60c | | PyTorch SDPA (Flash Attention) | 0.04 | Not recorded | 69896a2fc97a0b3e3b5dc60c | | Sparse Triton Kernel | 0.08 | Not recorded | 69896a2fc97a0b3e3b5dc60c | | Sparse Triton Kernel | 0.19 | Not recorded | 69896a2fc97a0b3e3b5dc60c | | PyTorch SDPA (Flash Attention) | 0.3 | Not recorded | 69896a2fc97a0b3e3b5dc60c | | Sparse Triton Kernel | 0.5 | Not recorded | 69896a2fc97a0b3e3b5dc60c | | PyTorch SDPA (Flash Attention) | 1 | Not recorded | 69896a2fc97a0b3e3b5dc60c | | Sparse Triton Kernel | 1.05 | Not recorded | 69896a2fc97a0b3e3b5dc60c | | Sparse Triton Kernel | 2.78 | Not recorded | 69896a2fc97a0b3e3b5dc60c | | PyTorch SDPA (Flash Attention) | 3.27 | Not recorded | 69896a2fc97a0b3e3b5dc60c | | Sparse Triton Kernel | 3.94 | Not recorded | 69896a2fc97a0b3e3b5dc60c | | PyTorch SDPA (Flash Attention) | 10.91 | Not recorded | 69896a2fc97a0b3e3b5dc60c | | PyTorch SDPA (Flash Attention) | 41.41 | Not recorded | 69896a2fc97a0b3e3b5dc60c | | PyTorch SDPA (Flash Attention) | 156.25 | Not recorded | 69896a2fc97a0b3e3b5dc60c | | PyTorch SDPA (Flash Attention) | 627.58 | Not recorded | 69896a2fc97a0b3e3b5dc60c | ### Sparse Triton Kernel Latency (ms): 0.03. Result ID: 6998e92f5698665ec44d9d55. [Reporting paper]() [Method origin paper]() Recorded evaluation configuration: {"Sequence Length":1024,"Hardware":"NVIDIA RTX 4090"}. This record does not assert independent reproduction. ### Sparse Triton Kernel Latency (ms): 0.03. Result ID: 6998e92f5698665ec44d9d57. [Reporting paper]() [Method origin paper]() Recorded evaluation configuration: {"Sequence Length":2048,"Hardware":"NVIDIA RTX 4090"}. This record does not assert independent reproduction. ### PyTorch SDPA (Flash Attention) Latency (ms): 0.04. Result ID: 6998e92f5698665ec44d9d54. [Reporting paper]() [Method origin paper]() Recorded evaluation configuration: {"Sequence Length":1024,"Hardware":"NVIDIA RTX 4090"}. This record does not assert independent reproduction. ### Sparse Triton Kernel Latency (ms): 0.08. Result ID: 6998e92f5698665ec44d9d59. [Reporting paper]() [Method origin paper]() Recorded evaluation configuration: {"Sequence Length":4096,"Hardware":"NVIDIA RTX 4090"}. This record does not assert independent reproduction. ### Sparse Triton Kernel Latency (ms): 0.19. Result ID: 6998e92f5698665ec44d9d5b. [Reporting paper]() [Method origin paper]() Recorded evaluation configuration: {"Sequence Length":8192,"Hardware":"NVIDIA RTX 4090"}. This record does not assert independent reproduction. ### PyTorch SDPA (Flash Attention) Latency (ms): 0.3. Result ID: 6998e92f5698665ec44d9d56. [Reporting paper]() [Method origin paper]() Recorded evaluation configuration: {"Sequence Length":2048,"Hardware":"NVIDIA RTX 4090"}. This record does not assert independent reproduction. ### Sparse Triton Kernel Latency (ms): 0.5. Result ID: 6998e92f5698665ec44d9d5d. [Reporting paper]() [Method origin paper]() Recorded evaluation configuration: {"Sequence Length":16384,"Hardware":"NVIDIA RTX 4090"}. This record does not assert independent reproduction. ### PyTorch SDPA (Flash Attention) Latency (ms): 1. Result ID: 6998e92f5698665ec44d9d58. [Reporting paper]() [Method origin paper]() Recorded evaluation configuration: {"Sequence Length":4096,"Hardware":"NVIDIA RTX 4090"}. This record does not assert independent reproduction. ### Sparse Triton Kernel Latency (ms): 1.05. Result ID: 6998e92f5698665ec44d9d5f. [Reporting paper]() [Method origin paper]() Recorded evaluation configuration: {"Sequence Length":32768,"Hardware":"NVIDIA RTX 4090"}. This record does not assert independent reproduction. ### Sparse Triton Kernel Latency (ms): 2.78. Result ID: 6998e92f5698665ec44d9d61. [Reporting paper]() [Method origin paper]() Recorded evaluation configuration: {"Sequence Length":65536,"Hardware":"NVIDIA RTX 4090"}. This record does not assert independent reproduction. ### PyTorch SDPA (Flash Attention) Latency (ms): 3.27. Result ID: 6998e92f5698665ec44d9d5a. [Reporting paper]() [Method origin paper]() Recorded evaluation configuration: {"Sequence Length":8192,"Hardware":"NVIDIA RTX 4090"}. This record does not assert independent reproduction. ### Sparse Triton Kernel Latency (ms): 3.94. Result ID: 6998e92f5698665ec44d9d63. [Reporting paper]() [Method origin paper]() Recorded evaluation configuration: {"Sequence Length":131072,"Hardware":"NVIDIA RTX 4090"}. This record does not assert independent reproduction. ### PyTorch SDPA (Flash Attention) Latency (ms): 10.91. Result ID: 6998e92f5698665ec44d9d5c. [Reporting paper]() [Method origin paper]() Recorded evaluation configuration: {"Sequence Length":16384,"Hardware":"NVIDIA RTX 4090"}. This record does not assert independent reproduction. ### PyTorch SDPA (Flash Attention) Latency (ms): 41.41. Result ID: 6998e92f5698665ec44d9d5e. [Reporting paper]() [Method origin paper]() Recorded evaluation configuration: {"Sequence Length":32768,"Hardware":"NVIDIA RTX 4090"}. This record does not assert independent reproduction. ### PyTorch SDPA (Flash Attention) Latency (ms): 156.25. Result ID: 6998e92f5698665ec44d9d60. [Reporting paper]() [Method origin paper]() Recorded evaluation configuration: {"Sequence Length":65536,"Hardware":"NVIDIA RTX 4090"}. This record does not assert independent reproduction. ### PyTorch SDPA (Flash Attention) Latency (ms): 627.58. Result ID: 6998e92f5698665ec44d9d62. [Reporting paper]() [Method origin paper]() Recorded evaluation configuration: {"Sequence Length":131072,"Hardware":"NVIDIA RTX 4090"}. This record does not assert independent reproduction. Page 1 of 1. This is an index page, not the full collection.