RTX 3090 with tensor-rate unlock up to 6.6x faster than M5 Max on MiniMax H3 video benchmarks
Ok_Bird4483 · reddit · 2026-11-09
Identical MiniMax H3 ComfyUI workflows, same seeds, 11 rows: a 24GB RTX 3090 beat both the 128GB ASUS RTX Spark and the M5 Max on every row despite streaming models over PCIe x4.
Key numbers
- Stock 3090: 1.25–1.84x faster per step than the Spark, 3.5–4.9x faster than the M5 Max.
- With the tensor-rate unlock (bf16 matmul raised from 75 to 124 TFLOPS): 1.85–2.46x vs Spark, 4.9–6.6x vs M5 Max; 1.34–1.49x real-world speedup over stock, biggest on heavy workflows.
- Power: 3090 draws 320–350W vs 65W (Spark) and 22W (Mac), but finishes each step 1.5–6.5x sooner, so energy per step is only 2.2–3.4x higher.
- All 22 runs on the 3090 passed output validation with no memory-saving flags, same ComfyUI config as the Spark.
- The M5 Max produced flat-black video on both full-reference ref2va rows (exceeding the 34k-token MPS limit), so no valid results there.
Full tables for seconds-per-step, wall-clock per clip, and GPU power are included.
More from Infra
- Agentic AI is orchestration of model calls; datacenter limits may push edge models — DavidLinthicum · 2026-11-09
- "Bro got distilled": meme pairs Jensen Huang photo with Huawei's Nvidia chip challenge — StephanSturges · 2026-11-09
- SpaceX plans $24 billion AI supercomputer in Starbase, Texas — beffjezos · 2026-11-09
- Megaspeed allegedly imported $4.6B of Nvidia GPUs, $2B landing in Malaysia DCs — bytebot · 2026-11-09
- GLM-130B Training Cost About $600K: Two Months on 96 A100s — ccerrato147 · 2026-11-09
- $16K CAD 4x R9700 local inference server: 190 tok/s per user, 8 users at 128K context benchmarked — Anarchaotic · 2026-11-09