Summary

This report benchmarks GPU options for deploying Scope's real-time video diffusion inference pipelines. We evaluate performance, memory fit (OOM risk), and cost trade-offs across multiple resolutions and four pipelines: reward-forcing, longlive, streamdiffusionv2, and krea-realtime-video.

Key Takeaways

• H200 SXM is the fastest in every measured case where it runs, typically delivering about 5% - 15% higher FPS than H100 SXM and materially higher FPS than RTX 5090.
• H100 SXM is the best “default”: it supports high resolutions reliably (large VRAM), performs close to H200 SXM, and avoids the 5090’s earlier OOM failures.
• RTX 5090 offers strong value when workloads fit in 32 GB VRAM, but it hits memory limits at higher resolutions and on memory-heavy pipelines (notably krea-realtime-video).
• Provider choice can matter at low resolutions (underutilized GPU): RTX 5090 on TensorDock was 10–30% faster than on RunPod in several low-res tests, while results converge at high resolutions where GPUs saturate.
• Based on current results, reward-forcing is the strongest performing pipeline based on FPS: it offers the highest-FPS pipeline across GPUs and resolutions in this benchmark set.

Next Steps to Optimize Performance

• Accelerate inference: evaluate TensorRT, torch.compile, and quantization (FP16/INT8 where feasible) to reduce latency and boost throughput.
• Lower-res + upscale: run diffusion at a smaller base resolution, then apply an upscaler to reach the target output size—balancing quality against cost.
• Increase GPU packing: at low resolutions, GPU utilization often falls below 90% with free VRAM remaining. Test running multiple concurrent streams per GPU to maximize throughput per dollar.

Which GPU should you choose?

Use this as the default selection logic; detailed evidence appears in later sections.

Benchmark

Methodology

Pipelines tested: