Supported hardware

Hardware targets

Target	Hardware	Memory	Precision
orin-nano	Jetson Orin Nano	8 GB	fp16
orin	Jetson Orin	32 GB	fp16
orin-64	Jetson Orin 64	64 GB	fp16
thor	Jetson Thor	128 GB	fp8
desktop	RTX / A100	40 GB	fp16

reflex inspect targets lists current profiles.

Supported GPU architectures

Architecture	Compute	Status	Notes
Ampere (RTX 30-series, A10G, A100)	sm_8.0–8.6	Supported	Tested on Modal A10G + A100, RTX 4090
Ada Lovelace (RTX 40-series, L4)	sm_8.9	Supported
Hopper (H100, H200)	sm_9.0	Supported
Jetson Orin (Orin Nano / NX / AGX)	sm_8.7	Supported	JetPack 5.x or 6.x
Jetson Thor	sm_10.x	Untested	Same Blackwell silicon as desktop, but ORT-bundled CUDA EP needs Blackwell support
Blackwell desktop (RTX 5090, RTX PRO 6000, B200, GB200)	sm_10.0	Not yet supported	ORT’s bundled cuBLAS/cuDNN don’t ship sm_100 kernels
Older NVIDIA (Turing RTX 20, GTX 16)	sm_7.5	Best-effort	Should work but not in CI matrix
Pre-Tensor-Core (Maxwell Jetson Nano 4 GB, GTX 9-series)	sm_5.x	Not supported	NVIDIA EOL’d this hardware at JetPack 4.6 (Python 3.6)

Blackwell users

The bootstrap installer accepts your hardware and the package installs cleanly, but reflex go will segfault at server startup. The real fix requires ORT to ship Blackwell-aware bundled binaries. Workarounds:

• reflex chat works — chat-only mode, no GPU needed
• reflex doctor and reflex models list work
• /act and TRT-engine inference need a non-Blackwell GPU temporarily

A Blackwell-specific runtime path via TensorRT-LLM is planned — see the ADR queue.

Performance

reflex-vla[serve,gpu] (v0.7+) uses ONNX Runtime’s TensorRT execution provider out of the box. Measured on Modal A10G (Ampere, sm_8.6) on 2026-04-29 against SmolVLA monolithic (5 warmup + 20 measured forward passes, batch=1):

Provider	Mean latency	p95
CUDAExecutionProvider (ORT-CUDA fallback)	108.11 ms	108.68 ms
TensorrtExecutionProvider (default in v0.7+)	19.49 ms	19.71 ms

5.55× faster. The win comes from TensorRT’s FP16 kernels + engine fusion. Older releases silently fell back to ORT-CUDA on most installs because libnvinfer.so.10 and CUDA libs weren’t on LD_LIBRARY_PATH — v0.7’s [serve,gpu] extras pull tensorrt>=10 and reflex patches LD_LIBRARY_PATH automatically at import.

Verify the win

reflex doctor

Look for green ✓ on:

• TensorRT runtime (libnvinfer.so.10) — loadable
• CUDA cuBLAS (libcublas.so.12) — loadable
• CUDA cuDNN (libcudnn.so.9) — loadable
• ORT-TRT EP active — session created with TRT EP in active providers

If any are red ✗, the remediation hint says exactly which pip install to run. The most common cause is using [serve,gpu-min] or an older release that didn’t pull tensorrt automatically.