Add RTX 4090 baseline support #22

Activity

1. khanhnd61-vr commented on Oct 2, 2026

   

   khanhnd61-vr

   on Oct 2, 2026

   Collaborator

   More actions

   Thanks for the request, @Zctoylm0927, and sorry for the slow reply.

   We don't have an RTX 4090 at the moment. We're trying to find one to test on, and we'll post the numbers here when we do, so we'll keep this issue open until then.

   In the meantime, the benchmarks now go well beyond the RTX 5090: latency and memory for all 13 vla.cpp policies on 15 devices, from Jetson boards to desktop GPUs. You can browse them on VLA Hub: https://khanhnd61-vr.github.io/vla-hub/

   Search a policy or a GPU to see its latency, actions per second and peak VRAM on every device, next to vla.simd on CPUs. The closest cards to a 4090 there are the RTX 3090 and the RTX 5090: https://khanhnd61-vr.github.io/vla-hub/?q=rtx

   If you have access to a 4090, we'd love your numbers. The result form on VLA Hub walks through the whole protocol step by step, from the build to reading peak VRAM. The core of it is:

   cmake -B build -DCMAKE_BUILD_TYPE=Release -DGGML_CUDA=ON -DCMAKE_CUDA_ARCHITECTURES=89
   cmake --build build -j
   ./build/vla-bench -hf vrfai/smolvla-libero-gguf --images 2 --size 512 --warmup 3 --reps 20

   Then submit your results here. We'll review them and add the RTX 4090 to the hub, with credit to you.