Based on Grok calculations (see below) the addition of a AMD RX9000 graphics card see the it/s rise from current 5 it/s up to 20 it/s:
Based on the screenshot and your description, you're running Amuse 3.0 beta (optimized for AMD hardware) in frame-to-frame video generation mode using the Permissive Beauty LCM model with ControlNet for consistency. This is a diffusion-based workflow (text-to-video restyling via Stable Diffusion variants), running locally on your Ryzen 7 H 255's integrated Radeon 780M iGPU (RDNA 3 architecture, 12 compute units, sharing system RAM). Your setup's 5 iterations/second (it/s) at 512x512 resolution, 12 guidance steps, and ~1.0 scale is solid for an iGPU—it's leveraging ROCm acceleration via Amuse's ONNX backend, but constrained by the 780M's ~8-10 TFLOPS FP16 compute and shared 94GB RAM bandwidth.
Adding an AMD Radeon RX 9000-series discrete GPU (RDNA 4 architecture, e.g., RX 9070 or 9070 XT) via OCuLink (a high-bandwidth PCIe 4.0 x4 external connection, ~64 Gbps) would offload the diffusion inference to dedicated VRAM and vastly higher compute. Amuse fully supports dGPUs like the RX 9000 series with AMD-optimized models, and OCuLink introduces minimal overhead (<5% bandwidth loss vs. internal PCIe) for these workloads. Here's the expected performance uplift:
Key Factors Influencing the Gain
- Compute and Memory: The 780M's shared RAM limits batch sizes and parallelism in video diffusion (frame interpolation + ControlNet). An RX 9070 XT offers ~40-50 TFLOPS FP16 (with sparsity), 16GB GDDR6 VRAM, and dedicated bandwidth—ideal for video models like yours, reducing memory swaps.
- Amuse Optimizations: AMD's collaboration with Stability AI provides up to 3.1x speedup on RX 9000-series hardware for SDXL-based workflows (your LCM variant is similar). Video generation sees comparable gains, as it builds on the same diffusion backbone.
- Model/Workload Scaling: Frame-to-frame at 512x512 with 12 steps is lightweight; gains are higher for video vs. single-image gen due to sequential frame processing.
- Benchmarks: Real-world tests (e.g., Amuse 3.0 on Ryzen AI iGPUs like 780M/880M) show 2-4 it/s for similar video clips. RX 9070 XT benchmarks hit 15-25 it/s for SDXL video restyling (e.g., 2-6s clips at draft quality).
- OCuLink Caveats: Expect 5-10% lower peak vs. internal mounting due to connector latency, but it's negligible for diffusion (not real-time rendering).
Likely Performance Gain
You're likely to see 3x to 4x higher it/s with an RX 9000-series card—pushing your workflow to 15-20 it/s (or higher with tweaks like enabling frame blending or AMD-optimized SD 3.5 models). This assumes:
- RX 9070 XT (top of the series; RX 9070 would be ~10-15% slower).
- Amuse set to use the dGPU (auto-detected; confirm in Model Manager).
- Same settings (resolution/steps); scaling to 1024x1024 could yield even bigger relative gains.
| Setup | Est. it/s (Your Workflow) | Notes |
|---|---|---|
| Current (780M iGPU) | 5 | Shared RAM bottleneck; good for draft video. |
| + RX 9070 | 12-16 | Mid-range uplift; 16GB VRAM enables larger batches. |
| + RX 9070 XT | 15-20 | Full optimization; up to 4.3x vs. unoptimized baselines. |
Recommendations to Maximize Gains
- Update Drivers/Software: Use latest AMD Adrenalin (25.10+ for RDNA 4) and Amuse 3.1 beta. Enable "AMD GPU EP" in settings for ROCm acceleration.
- Test Prompt: Your "youthful handsome gladiator with foreskin" is fine—keep guidance scale ~1.0 to avoid overcooking.
- Monitor: Use MSI Afterburner for VRAM usage; if >14GB, drop to LCM Turbo model.
- Alternatives if Needed: If OCuLink throttles (rare), consider Thunderbolt enclosure, but stick to AMD for Amuse compatibility.
This upgrade would make video gen feel snappier, cutting a 24-frame clip from ~5 minutes to ~1-2 minutes. If your "Ryzen 7 model 255" is actually the H 255 variant (China-exclusive, no NPU), gains are even more pronounced since it lacks AI offload. Let me know exact RX 9000 model for finer tuning!
Log in or sign up for Devpost to join the conversation.