Hailo-8 vs Hailo-8L: YOLOv11s inference latency only improves from ~19 ms to ~13 ms

Abhay_Payadi · September 9, 2026, 8:27am

Hi Hailo team/community,

I’m comparing the inference performance of the same YOLOv11s car detection model on Hailo-8L and Hailo-8.

The model is compiled as a Hailo HEF and I’m running it using hailortcli.

Hailo-8L

Inference FPS: 52.45 FPS
Approx. latency: 19.07 ms

Hailo-8

Using:

hailortcli run ~/hef_models/yolov11s_car_hailo8.hef \
    --measure-latency \
    --measure-overall-latency

I get:

HW Latency:       13.12 ms
Overall Latency:  15.46 ms

The normal streaming run gives approximately:

FPS: 64 FPS

I also tested the Hailo-8 in both power modes:

--power-mode performance
FPS: 64.46

--power-mode ultra_performance
FPS: 63.86

So there is essentially no performance difference between performance and ultra_performance.

My question

Given that Hailo-8 has approximately twice the theoretical TOPS of Hailo-8L, I expected a larger performance improvement.

Instead, I am seeing approximately:

Hailo-8L:  ~19.07 ms
Hailo-8:   ~13.12 ms HW latency

or about a 31% reduction in hardware latency, rather than something close to 2×.

Is this expected for this type/size of YOLOv11s network?

Could the limitation be related to:

Hailo core/cluster utilization
LCU utilization
Memory bandwidth
Network architecture / layer parallelism
Dataflow/compiler mapping
Batch size
Some other hardware resource that doesn’t scale linearly between Hailo-8L and Hailo-8

Is there a recommended way in HailoRT 4.24 to determine how much of the Hailo-8 compute resources are actually being utilized by this HEF?

I have also collected runtime data from the HEF and can provide the profiler output if that would help.

Thanks!

Klaus_Koschinsky · September 9, 2026, 8:52am

The performance difference between running a model on the Hailo-8 and Hailo-8L can vary widely depending on the limiting factors.

For example, a model that fits well within a single context on the Hailo-8 but requires multiple contexts on the Hailo-8L could potentially run 10× faster on the Hailo-8.

On the other hand, a small model may run at roughly the same speed on both devices if the Dataflow Compiler cannot take advantage of the Hailo-8’s additional hardware resources.

Yes, all of them.

The profiler report can give you detailed information about the resources used.

In the profiler report select: Compilation & Runtime Details, then select the context on the left side in the Compiled Model Graph and the information will be shown on the right side in the Context Analytics window. You can switch between a graphical representation and a table with all details using the two icons in the upper right corner of the Context Analytics window.