# RF-DETR - Latency Results Do Not Match Paper

**URL:** https://discuss.roboflow.com/t/rf-detr-latency-results-do-not-match-paper/11633
**Category:** 🤝  Community Help
**Tags:** export
**Created:** [December 21, 2025, 11:26am UTC](https://discuss.roboflow.com/t/rf-detr-latency-results-do-not-match-paper/11633 "2025-12-21T11:26:41Z")
**Posts on this page:** 3
**Page:** 1

<div class="post-metadata">

### Author: ![Omer\_Taub](https://yyz1.discourse-cdn.com/flex029/user_avatar/discuss.roboflow.com/omer_taub/32/6460_2.png) [@Omer\_Taub](https://discuss.roboflow.com/u/Omer_Taub)
#### Post date: [December 21, 2025, 11:26am UTC](https://discuss.roboflow.com/t/rf-detr-latency-results-do-not-match-paper/11633/1 "2025-12-21T11:26:41Z")

</div>

Hello,

I am benchmarking the model reported in the paper and am seeing a noticeable mismatch between the latency numbers reported in the paper and the latency I measure in practice.

**Setup details:**

- Batch size: 1

- GPU: NVIDIA RTX PRO 4500 (Blackwell)

- Inference focused (no training, no data loading overhead)

Despite matching the batch size and using a modern high-end GPU, the measured latency is consistently higher than what is reported in the paper. I want to confirm:

1. Whether the paper’s latency numbers were measured with any specific assumptions (e.g., mixed precision, TensorRT, specific CUDA/cuDNN versions, or warmup strategy).

2. Whether preprocessing/postprocessing was excluded from the reported latency.

3. If Flash Attention, fused kernels, or other backend-specific optimizations were explicitly enabled.

4. Whether the reported numbers reflect end-to-end latency or pure model forward time.

Any clarification on the exact benchmarking methodology used in the paper would be very helpful, as I am trying to reproduce the results as closely as possible.

Thank you!

---

<div class="post-metadata">

### Author: ![isaacrob](https://yyz1.discourse-cdn.com/flex029/user_avatar/discuss.roboflow.com/isaacrob/32/5638_2.png) [@isaacrob](https://discuss.roboflow.com/u/isaacrob)
#### Post date: [December 21, 2025, 2:15pm UTC](https://discuss.roboflow.com/t/rf-detr-latency-results-do-not-match-paper/11633/2 "2025-12-21T14:15:00Z")

</div>

The [paper](https://arxiv.org/pdf/2511.09554) says the specific conditions under which latency is measured, see section 4 and the appendix on CUDA graphs. As mentioned in the paper, we use TensorRT 10.4 and CUDA 12.4 on a T4 GPU with FP16 and CUDA graphs enabled. We measure model forward time, which for a DETR is the end to end time.

As is also mentioned in the paper, we take a close look at power throttling as a source of inconsistencies in latency measurements in different papers and propose a solution. To facilitate research, we linked to an [open source repo](https://github.com/roboflow/single_artifact_benchmarking) containing code to exactly reproduce our claimed latencies and accuracies.

---

<div class="post-metadata">

### Author: ![system](https://canada1.discourse-cdn.com/flex029/uploads/roboflow1/original/1X/bb6edbfa56d0cf2ee603f061e9e09103502cc8cc.png) [@system](https://discuss.roboflow.com/u/system)
#### Post date: [January 11, 2026, 2:15pm UTC](https://discuss.roboflow.com/t/rf-detr-latency-results-do-not-match-paper/11633/3 "2026-01-11T14:15:43Z")

</div>

This topic was automatically closed 21 days after the last reply. New replies are no longer allowed.