FP4 (NVFP4) support on Jetson Orin Nano — is it planned or architecturally excluded?

Robotics & Edge Computing Jetson Systems Jetson Orin NX
,
1

Hi,

I’m running inference on a Jetson Orin Nano Dev Kit Super and wanted to
check whether FP4 (NVFP4) precision is supported or planned for this platform.

Device details:

  • Board: Jetson Orin Nano Engineering Reference Developer Kit Super
  • JetPack: 6.2 (L4T R36.4.4)
  • CUDA: 12.6
  • TensorRT: 10.3.0
  • cuDNN: 9.3.0
  • Compute capability: 8.7 (Ampere)

When I check for FP4 support in the TensorRT Python API:

import tensorrt as trt
logger = trt.Logger(trt.Logger.WARNING)
builder = trt.Builder(logger)
config = builder.create_builder_config()
print(builder.platform_has_fast_fp16)   # True
print(builder.platform_has_fast_int8)   # True
config.get_flag(trt.BuilderFlag.FP4)    # AttributeError: no attribute 'FP4'

The FP4 builder flag does not exist in TensorRT 10.3 on this device.

My understanding is that FP4/NVFP4 requires Ada Lovelace (sm_89) or
Blackwell (sm_100+), and Orin is sm_87 (Ampere).

Questions:

  1. Is FP4 architecturally excluded on Jetson Orin (sm_87)?
  2. Is FP4 support planned in a future JetPack/TensorRT release for Jetson?
  3. Is INT8 the recommended lowest-bit precision for Jetson Orin today?

Currently using FP16 TensorRT engines which work well — just confirming
there is no lower-precision option before finalizing our deployment.

Thanks

4

Hi,

Orin doesn’t support FP4 (NVFP4).
This is the hardware limitation, you will need the Blackwell series (ex. Thor) to get the support.

Thanks.

5

Hi,

Thank you for the clarification.