What Your GPU Actually Does (and Doesn't Do) in Traditional Editors
In Premiere Pro or DaVinci Resolve, your GPU handles:
- Playback acceleration: Decoding and displaying video on the timeline
- Effects rendering: Lumetri color, transitions, and GPU-accelerated effects
- Export encoding: NVENC hardware encoding (if enabled) during final render
What your GPU does not do in traditional editors:
- Understand your footage: No AI inference - your GPU doesn't analyze what's in your clips
- Make editing decisions: Clip selection, ordering, and timing are 100% manual
- Search your library: Finding specific clips requires manual scrubbing or filename guessing
In other words, most of your GPU's capability sits idle. The CUDA cores that could run transformer models do nothing while you manually scrub through footage.
CUDA vs. CPU: Why the Gap Is Enormous for AI Inference
AI models like CLIP are built on matrix multiplications — thousands of them per frame. A modern CPU handles these sequentially across 8–16 cores; an RTX 4070 has 5,888 CUDA cores executing those same operations in parallel. In practice, running OpenCLIP ViT-L/14 on a CPU is slow — on the order of seconds per frame — while a modern NVIDIA GPU runs the same inference well over an order of magnitude faster.
This gap matters most during ingest, when every clip in your library needs a semantic embedding. A large library that would crawl on CPU finishes in a fraction of the time on GPU — the difference between a long workflow pause and a quick step. Onset Engine still runs fully on CPU when no GPU is available; it's just slower.
NVENC: Why Hardware Encoding Matters
NVIDIA GPUs include a dedicated hardware encoder (NVENC) that operates independently of the CUDA cores. While your CUDA cores handle AI inference, NVENC can simultaneously encode your final output to H.264 or H.265. This means rendering does not compete with AI processing for GPU resources, and hardware encoding is far faster than CPU-based x264. For 4K output, the gap widens further. See how this powers hardware-encoded video concatenation in practice.
What Cloud AI Services Are Selling You
When you use RunwayML, Descript, or CapCut's AI features, here's what's actually happening:
- Your footage is uploaded to their servers
- Their GPUs (identical to yours) run AI inference on your footage
- They charge you $0.05–$0.50 per minute of processed video
- Results are sent back over the network
You're paying a monthly fee for remote access to the same CUDA architecture sitting under your desk. The only difference is the software layer.
The Shortcut: Full GPU Utilization with Onset Engine
Onset Engine is built ground-up for local GPU execution. Here's what each component of your GPU actually does:
- CUDA cores → AI inference: OpenCLIP ViT-L/14 runs on your CUDA cores via PyTorch. Every clip gets a 768-dimensional semantic embedding. Your GPU understands your footage
- NVENC encoder → hardware rendering: Final video output is encoded by NVIDIA's dedicated hardware encoder - not the CPU - so encoding doesn't bottleneck the render
- VRAM → model hosting: The CLIP model loads once into VRAM (~2GB) and stays resident. Inference runs at GPU memory bandwidth, not disk speed
- Tensor cores → mixed precision: FP16 inference on RTX tensor cores doubles throughput on supported cards
GPU utilization during an Onset Engine ingest: 60–85% GPU, which is what that hardware was designed for. Cloud AI charges you $0.50/min for this. Onset Engine does it for $0/min after a one-time $59.50 purchase.
Hardware Requirements and VRAM Considerations
- Minimum: NVIDIA GTX 1060 6GB (CUDA 6.1) - functional but slow. The 6GB VRAM is tight — the CLIP model occupies ~2GB, leaving 4GB for video decode buffers.
- Recommended: NVIDIA RTX 3060 12GB - fast inference, 7th-gen NVENC, and 12GB VRAM provides ample headroom for large batch ingests without memory pressure.
- Optimal: NVIDIA RTX 4070+ 12GB - 4th-gen tensor cores enable FP16 mixed-precision inference at nearly double the throughput. 8th-gen NVENC supports AV1 hardware encoding.
VRAM is the most important spec for local AI work. The OpenCLIP ViT-L/14 model needs approximately 2GB of VRAM when loaded in FP16 precision. The remaining VRAM handles video frame buffers during inference. Cards with 8GB or more will never hit memory limits during normal editing workflows. More CUDA cores means faster batch processing across large clip libraries, but even a mid-range card runs the full pipeline — the same pipeline that cloud services charge monthly for.
NVIDIA is the recommended and best-supported path: CUDA drives the AI inference and NVENC drives hardware encoding. AMD and Intel GPUs are not supported for AI inference (that runs on CPU on those systems), but AMD's AMF hardware encoder is used for rendering where available. Apple Silicon (MPS) support is experimental.