TFLOPS(FP16) about DLA (Deep Learning Accelerator) on Jetson Orin NX

[Hiromitsu.Matsuura]

I think TensorRT supports running networks in either INT8 or FP16 on DLA.

And I can get the max performance of INT8 from the following Technical Brief.

This enables up to 105 INT8 Sparse TOPs total on Jetson AGX Orin DLAs.

nvidia-jetson-agx-orin-technical-brief.pdf

However, I cannot find the the max performance of FP16 from the following Technical Brief.

Could anyone tell me the max performance of FP16?

Thanks.

[AastaLLL]

Hi,

Please check the topic below:

How many 16FP Tops for agx-orin 32G

It’s described in the datasheet,but there’s no answer I need. and also see, DLA-v2 is slower than DLA-v1 - #8 by AastaLLL “Orin’s DLA has more int8 dense TOPs but fewer fp16 TOPs.” I want to know what the actual data of FP16 TOPs should be， Thank you for your answer. AI Performance JAO 64GB: Up to 275 Sparse TOPS (INT8) JAO 32GB: Up to 200 Sparse TOPs (INT8) Ampere GPU JAO 64GB: 2 GPC | 8 TPC | 2048 NVIDIA® CUDA® cores | 64 Tensor cores Ray-Tracing cores | 170 Sparse TOPS | Maximum…

Thanks.

[Hiromitsu.Matsuura]

Hi, AastaLLL,

Thank you for your infomation.
I overlooked the topic.

From the topic, I understood that NVIDIA cannot share the exact number here.

Regards,
hiro