RTX 5090 vs AMD vs Intel: Local Video Generation GPU Benchmarks
prompt_seeker · reddit · 2026-08-06
A developer benchmarked several mainstream GPUs running the MiniMax H3 I2VA video generation model (544x800, 5s duration).
Key Results:
- RTX 5090 (Power limited to 400W): 55.6s sampling, 1m total.
- RTX 3090: 2m 22s sampling.
- AMD R9700: 3m 03s. While slower than the 3090, it suggests improving price-to-performance for upcoming AMD cards.
- RX 7900XT: >5m sampling, bottlenecked partly by PCIe 4.0 x4 and slow VAE decoding.
- RTX 3060: 9m total.
- Intel B580: Failed to run stably after 3 hours of troubleshooting.
Technical Notes:
- Enabling Dynamic VRAM improved efficiency even on the RTX 5090.
- Using --disable-pinned-memory helps AMD GPUs.
- Sage Attention performed better on the R9700, while Flash Attention was faster on the RX 7900XT.
More from Infra
- Running DeepSeek V4 Flash on MacBook M5 Pro Hits 17 t/s via SSD Streaming — vogelvogelvogelvogel · 2026-08-06
- AI Model Router Startup Sapiom Raises $35M Series A — darian314 · 2026-08-06
- Discussion: Running llama-server Inference Across Machines via RPC Clustering — _TheWolfOfWalmart_ · 2026-08-06
- Modal Rebuilds Sandbox Platform to Create 1M Concurrent Containers in Under a Minute — dscape · 2026-08-06
- Nebius Tops Endpoint Accuracy for GLM-5.2, Hits ~300 Tokens/s Output — songhan_mit · 2026-08-06
- Gavin Baker on AI Compute: SRAM Accelerators Offer Unbeatable ROI, Disaggregation is Key — IanAndrewsDC · 2026-08-06