Figure 6c reports backward attention throughput on A100 80GB SXM4; PyTorch is marked OOM at sequence length 16k. The panel covers causal attention with head dimension 64.

Figure 6c reports backward attention throughput on A100 80GB SXM4; PyTorch is marked OOM at sequence length 16k. The panel covers causal attention with head dimension 64. · CiteArk

Measurement results

0/29 assessed

0 supported0 disputed or conflicting29 not assessed

The complete result group stays with its claim. Assessed rows come first; their judgments do not replace the whole-claim conclusion. Expand a row for conditions, sources, and evidence history.

Each row is a measurement with its reported value, latest available observation, and cumulative evidence state.
Metric and conditions	Reported	Latest observation	Evidence judgment
method: PyTorch · sequenceLength: 512	19 TFLOPs/s	—	Not assessed
method: FlashAttention · sequenceLength: 512	58 TFLOPs/s	—	Not assessed
method: xFormers · sequenceLength: 512	46 TFLOPs/s	—	Not assessed
method: FlashAttention Triton · sequenceLength: 512	53 TFLOPs/s	—	Not assessed
method: FlashAttention-2 · sequenceLength: 512	81 TFLOPs/s	—	Not assessed
method: PyTorch · sequenceLength: 1k	21 TFLOPs/s	—	Not assessed
method: FlashAttention · sequenceLength: 1k	70 TFLOPs/s	—	Not assessed
method: xFormers · sequenceLength: 1k	54 TFLOPs/s	—	Not assessed
method: FlashAttention Triton · sequenceLength: 1k	68 TFLOPs/s	—	Not assessed
method: FlashAttention-2 · sequenceLength: 1k	111 TFLOPs/s	—	Not assessed
method: PyTorch · sequenceLength: 2k	24 TFLOPs/s	—	Not assessed
method: FlashAttention · sequenceLength: 2k	76 TFLOPs/s	—	Not assessed

Figure 6c reports backward attention throughput on A100 80GB SXM4; PyTorch is marked OOM at sequence length 16k. The panel covers causal attention with head dimension 64.

Reported and observed measurements

Assessments (0)

Runs (0)

Measurement results