HelixWorld world model tested: 70GB VRAM per BF16 GPU run, consumer cards need not apply
cocktailpeanut · x · 2026-10-20
Pinokio creator cocktailpeanut tested the open-source real-time audio-visual world model NoizAI/HelixWorld and shared the hardware requirements:
- VRAM: the README recommends an 80GB GPU; a single-GPU BF16 run uses about 70GB, with the Gemma-3-12B text encoder accounting for part of it.
- Measured performance: 0.77 real-time factor at 768×512, 24 FPS on one H800 80GB; similar results expected on A100/H100 80GB.
- Other requirements: Linux, Python 3.11, CUDA 12.x, plus ffmpeg/ffprobe.
His takeaway: VRAM is the bottleneck, so consumer GPUs won't run it anytime soon unless a low-VRAM option appears.
More from Infra
- SGLang merges megakernel low-latency mode, hitting ~1000 tok/s/user on GB300 NVL72 — ying11231 · 2026-10-21
- The topology of an LLM agent: where the model, loop, and tools run shapes cost and security — davidcrawshaw · 2026-10-21
- Chris Lattner shows how MAX and Mojo extend PyTorch across vendor hardware at PyTorchCon 2026 — PyTorch · 2026-10-21
- General Compute partners with Positron AI for cluster-scale deployment, live Q1 2027 — eptwts · 2026-10-21
- MemTensor's MemT-1.0 makes persistent memory part of model compute: 1.86x faster inference, 66x less storage — _akhaliq · 2026-10-21
- LithosAI sweeps 12 #1 spots on Artificial Analysis, GLM 5.3 Flash at 1,543 tok/s — JiaZhihao · 2026-10-21