ROCMFluxBenchmark

The pack's README makes a big claim: 78% faster Flux generation on a Strix Halo APU. This is the node that lets you check claims like that against your own hardware. ROCm Flux Benchmark runs a real sweep - multiple resolutions, multiple CFG values, measuring decode time and peak memory at each step - and hands you four text reports explaining what your GPU does well and where to tune. It's not a magic speedup; it's a truth-finder for your specific setup, which on AMD is exactly what you need because the "optimal" settings are never the ones in the NVIDIA tutorial you're reading.

What it actually runs

Connect a loaded MODEL, VAE, and CLIP (the standard loader outputs) and it starts timing. The input string that matters most is test_resolutions - default 256x320,512x512,1024x1024, comma-separated WxH pairs. test_steps (20) and test_cfg_values (default 1.0,3.5,8.0) control how much work each test does, and test_hipblas (on) toggles the HIP BLAS path checks. There's also generate_report for the full optimization report.

Fair warning from the source and the README both: this node actually runs the decode, so it takes minutes, not seconds. Don't click it and expect a quick answer - run it once, read the reports, save them.

The four outputs are all STRING, so wire each to a Show Text node:

• BENCHMARK_RESULTS - per-resolution average decode time and peak memory.
• PERFORMANCE_CHART - the raw timing breakdown per test.
• OPTIMIZATION_RECOMMENDATIONS - the useful one: it detects whether you're on AMD and gives different advice per platform (fp32 + Euler/Heun + tile 768–1024 on AMD; fp16/bf16 + DPM++ 2M on NVIDIA).
• MEMORY_ANALYSIS - total VRAM, current usage, device name.

The catch in the source

Be honest about what it measures: the benchmark decodes random latents through your VAE at each resolution and CFG. It does not run a full Flux denoising pass with your text conditioning - the "CFG values" here are fed to the VAE-decode test loop, not to actual sampling. So read it as a VAE-decode and memory-characterization benchmark for your GPU rather than an end-to-end Flux speed test. That's still genuinely useful - VAE decode is where a lot of AMD OOM pain lives, and knowing your card's decode ceiling at 1024px is actionable. But the 78% number in the README comes from a full-workflow A/B test, not this node.

Install

One pack, one install: ComfyUI Manager → "ROCm Ninodes," or

cd ComfyUI/custom_nodes
git clone https://github.com/iGavroche/rocm-ninodes.git

Restart, find it under ROCm Ninodes → Benchmark. The pack's dependencies are light (numpy, Pillow, psutil, gguf, safetensors); no extra model downloads - it uses the model, VAE, and CLIP you already have loaded.

The honest take

If you're chasing every last drop of speed on an AMD card, this is a reasonable afternoon's tool: it gives you a repeatable baseline so you can tell whether a settings change actually helped or you just got lucky with the seed. If you're on NVIDIA, skip it - the recommendations it produces are tuned for the ROCm world. And remember the pack's own philosophy: its performance claims are tested on one specific box (GMTek Evo-X2 Strix Halo, 128GB unified RAM), so treat its suggested numbers as a starting point, not gospel. Your mileage is the only number that matters.