Home / GPU ladder / RTX 4090 48GB (modded)
RTX 4090 48GB (modded) for MiniMax H3: Speed, VRAM & Recommended Setup
| Size (16:9) | 5s | 8s | 10s | 15s |
|---|---|---|---|---|
| 416 × 256 0.1MP · flicker-prone | 47 s VRAM 44G | 53 s VRAM 45G | 57 s VRAM 45G | 1 min 9 s VRAM 45G |
| 864 × 480 0.4MP | 1 min 20 s VRAM 46G | 1 min 53 s VRAM 26G | 2 min 21 s VRAM 27G | 3 min 49 s VRAM 29G |
| 960 × 544 0.5MP | 1 min 34 s VRAM 46G | 2 min 22 s VRAM 27G | 3 min 3 s VRAM 28G | 5 min 19 s VRAM 30G |
| 1152 × 640 0.7MP | 2 min 6 s VRAM 26G | 3 min 34 s VRAM 29G | 4 min 52 s VRAM 30G | 9 min 20 s VRAM 33G |
| 1344 × 768 1.0MP · official max | 3 min 3 s VRAM 28G | 5 min 47 s VRAM 31G | 8 min 19 s VRAM 33G | 17 min 36 s VRAM 38G |
| 1376 × 768 1.0MP · above native | 3 min 8 s VRAM 28G | 6 min VRAM 31G | 8 min 39 s VRAM 33G | 18 min 28 s VRAM 38G |
| 1920 × 1088 2.0MP · above native | 8 min 40 s VRAM 33G | 20 min 50 s VRAM 39G | 33 min 21 s VRAM 39G | 1.4 h VRAM 39G |
Times are per whole clip (model load and text encoding included); the first run is slower due to kernel compilation. ⚠ = weight offloading / noticeably slower; ✕ = not recommended.
Reading the numbers
With the recommended combo, the green (no disk spill) ceiling is 1920 × 1088 · 15s at ~1.4 h per clip. Already top-tier: stack 4–8 step distillation with Sage/Kitchen attention; for very long clips try SLA / VSA sparse attention.
Recommended combo (Flagship tier (32GB+ VRAM))
- Main model: minimax_h3_fl2va_pruned_int8_convrot.safetensors（20.97GB）
- Text encoder: qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors（15.69GB）
- Acceleration: LightX2V Turbo 8-step v1.0 · ~2.08× sampling
- Launch flags:
--fast-disk
1344x768 full length 5-15s; 4-8 steps; ~80s end-to-end for a 3s clip (measured on our 5090)
FAQ
Can the RTX 4090 48GB (modded) run MiniMax H3?
Yes. Estimated with 32GB RAM + fast-disk: 1920×1088·15s at ~1.4 h per clip. Configs that exceed VRAM fall back to weight offloading (slower, but they still render).
How fast is H3 on the RTX 4090 48GB (modded)?
About 1.4 h per 1920×1088·15s clip (including model load and encoding). Longer or denser clips scale quadratically in the attention term.
How much RAM do I need?
At least 64GB of system RAM recommended (minimum 48GB plus a 16GB pagefile). With 32GB, always enable --fast-disk or weights spill into the pagefile.
How can I make it faster?
Stack step distillation (e.g. LightX2V Turbo 8-step v1.0), enable SageAttention / Kitchen attention; for long clips try SLA / VSA sparse attention. All of it is switchable in step 2 of the calculator.
→ Run the numbers for your RAM and budget in the calculator (RTX 4090 48GB (modded) pre-filled)
Same tier: RTX 3090 24GB · RTX 4070 Ti Super 16GB · RTX 4080 16GB · RTX 4080 32GB (modded) · RTX 4080 Super 16GB · RTX 4080 Super 32GB (modded) · RTX 4090 24GB · RTX 4090 D 24GB
Related: Hardware guide · Acceleration compared · 8GB VRAM setups