PC Requirements for Real-Time Computer Vision on a Capture Feed

Why frame time matters more than FPS

A capture card delivers frames at a fixed rate. Your PC has until the next frame arrives to finish every step: copying the frame, converting and resizing it, running the model, post-processing the results (for example merging overlapping detection boxes) and drawing or sending the output. That time is the frame budget.

If the pipeline takes longer than the budget, frames either queue up, so the result lags further behind the picture, or get dropped. Averages hide this. A pipeline that averages 10 ms but spikes to 25 ms every few seconds will stutter at 60 fps, so look at the worst 1 % of frame times, not the mean. The input latency guide shows how processing time adds to capture and display delay.

GPU: NVIDIA, AMD and Intel

The GPU runs the model, so it decides most of your frame time. On Windows, vision software often uses ONNX Runtime, which runs a model through different execution providers. NVIDIA cards can use CUDA or TensorRT. AMD and Intel cards typically use DirectML, which Microsoft supports on any DirectX 12 GPU, including NVIDIA Kepler (GTX 600 series), AMD GCN (Radeon HD 7000 series) and Intel Haswell integrated graphics or newer. DirectML is in sustained engineering: still supported, with new feature work moving to Windows ML. The ONNX Runtime and DirectML guide explains the providers in detail.

TensorRT is often the fastest option on NVIDIA, with one catch: it builds an engine for your exact GPU on the first run, which can take minutes. ONNX Runtime’s documentation shows session creation dropping from 384 seconds to 9 seconds once that engine is cached.

Vendors quote AI performance differently. NVIDIA lists 614 AI TOPS for the RTX 5060, 988 for the RTX 5070 and 1,801 for the RTX 5080; AMD lists 64 AI accelerators on the Radeon RX 9060 XT and 128 on the RX 9070 XT; Intel’s Arc B580 has XMX engines. These peak figures are useful within one product line but do not convert directly into milliseconds for your model. Integrated graphics can run small models at low frame rates, but they share memory bandwidth and power with the CPU.

How much VRAM you need

The model file is rarely the problem: small detection models are often only a few to a few dozen megabytes. VRAM fills up with everything around it: intermediate results for each layer, runtime workspace, cached TensorRT engines, frame buffers, OBS encoding and the Windows desktop itself.

• 8 GB: comfortable for one or two small models at inputs around 640×640 on a 1080p feed.
• 12 GB: room for larger models, bigger inputs or recording on the same GPU.
• 16 GB: several models at once, high input resolutions, or streaming and vision on one card.

Compute also grows with input size: a 1280×1280 input has four times the pixels of 640×640, and the model’s work grows roughly with it. Check actual use in Task Manager under Performance > GPU > Dedicated GPU memory while everything runs.

Entry, recommended and high-end PCs

Previous-generation cards with similar performance and VRAM fit the same tiers. IRONHIDE’s own modules, such as IRONAIM inside IRONICS, follow the same logic: the capture card delivers frames and the GPU runs the model, so treat the tiers as a starting point and measure.

CPU, RAM and USB

The CPU handles frame copies, color conversion and resizing when they are not done on the GPU, decoding if a card sends compressed MJPEG, Python or app logic, and everything else running, such as OBS or a browser. Elgato’s minimums just for capture software give a floor: a 6th-generation Core i5 or Ryzen 5 3600 for the 4K X, an 8th-generation Core i5 or Ryzen for the 4K S. Elgato also warns against U- and M-suffix laptop chips and Pentium or Celeron processors.

Capturing at 1080p instead of 4K cuts both the USB load and the resize work on every frame, and usually costs nothing in accuracy, because the model only sees a downscaled frame anyway. Where your software offers it, doing the resize and color conversion on the GPU frees further CPU time.

For RAM, Elgato asks for 8 GB running in dual channel, which means two or more sticks. With vision software, capture and a browser open, 16 GB is a comfortable minimum, and 32 GB leaves room for recording or streaming.

USB is easy to overlook. Plug the capture card into a USB 3 port directly on the motherboard, and into a 10 Gbps port for 10 Gbps cards. Two USB capture devices need two separate USB 3 root hubs; many laptops have only one. In Device Manager, choose View > Devices by connection to see which ports share a hub. The capture card guide covers bandwidth per resolution.

Windows version and settings

Use Windows 11. Windows 10 support ended on October 14, 2025, so it no longer gets security updates, and some current hardware such as the Elgato 4K S already requires Windows 11. Then check these settings:

1. Install the latest GPU driver from NVIDIA, AMD or Intel.
2. Set Settings > System > Power & battery > Power mode to Best performance. On a laptop, stay plugged in.
3. On PCs with integrated and dedicated graphics, open Settings > System > Display > Graphics, pick your vision app, select Options and choose High performance.
4. Under Graphics, open the default or advanced graphics settings and test Hardware-accelerated GPU scheduling on and off. It needs a supported GPU and a WDDM 2.7 or newer driver, was introduced as an opt-in feature, and takes effect after a restart.
5. For UVC capture cards, turn on Settings > Privacy & security > Camera > Let desktop apps access your camera.
6. Close recorders, overlays and browser tabs you don’t need while testing.

How to measure your pipeline

• Task Manager > Performance > GPU shows each engine, such as 3D, Copy, Video Decode and Compute, plus dedicated GPU memory. In the Details tab you can add GPU and memory columns per process.
• nvidia-smi in a terminal reports utilization, memory use and clocks on NVIDIA cards.
• Per-stage timestamps in your own code, for example time.perf_counter() in Python around capture, preprocessing, inference and output, show where the milliseconds go.
• ONNX Runtime profiling (enable_profiling = True in the session options) writes a JSON trace with per-operator latency that opens in chrome://tracing or Perfetto.

Run each test for at least five minutes with your real capture feed, note the average and the worst 1 %, and watch GPU clocks and temperature for throttling. If the GPU sits near 100 % and frames are late, use a smaller model or input, FP16 or a faster provider. If the GPU is mostly idle and frames are still late, the bottleneck is the CPU, preprocessing or waiting between stages.

Troubleshooting slow or uneven frame times

Change one thing at a time and repeat the same five-minute test, so you know which change helped. Once the pipeline fits its budget, keep some headroom: a pipeline that needs 15 of its 16.7 ms at 60 fps has almost no margin left when a game update, a browser video or a recording starts on the same PC. The goal is for even the worst 1 % of frames to stay clearly inside the budget.

Frequently asked questions

Sources

• Microsoft Learn — Introduction to DirectML
• ONNX Runtime — Execution providers
• ONNX Runtime — DirectML execution provider
• ONNX Runtime — TensorRT execution provider
• ONNX Runtime — Profiling tools
• NVIDIA — GeForce RTX 5050
• NVIDIA — GeForce RTX 5060 family
• NVIDIA — GeForce RTX 5070 family
• NVIDIA — GeForce RTX 5080
• AMD — Radeon RX 9060 XT specifications
• AMD — Radeon RX 9070 XT specifications
• Intel — Arc B580 Graphics specifications
• Elgato — Game Capture 4K X system requirements
• Elgato — Game Capture 4K S system requirements
• Elgato — Two USB 3.0 devices need two root hubs
• Microsoft DirectX Blog — Hardware-accelerated GPU scheduling
• Microsoft DirectX Blog — GPUs in the Task Manager
• Microsoft Support — Windows 10 support has ended
• Microsoft Support — Change the power mode for your Windows PC
• Microsoft Support — Optimizations for windowed games (graphics settings per app)
• Microsoft Support — Manage app permissions for a camera in Windows