Coding AIMIT
NVIDIA Nemotron 3 Ultra
nvidia/NVIDIA-Nemotron-3-Ultra-550B-A55B-BF16
Parameters
550B
Context
1024K
Downloads
497K
Likes
316
Architecture
moe_transformer
GPU Requirements
VRAM / RAM / disk estimates by quantization and usage scenario
| Quantization | Minimum | Recommended | Production |
|---|---|---|---|
| FP1616bit | 1300.9 GB RAM 1951.3 · Disk 1637.2 | 1375.5 GB RAM 2063.3 · Disk 1663.2 | 1531.4 GB RAM 2297.0 · Disk 1702.2 |
| FP88bit | 652.9 GB RAM 979.4 · Disk 838.1 | 698.1 GB RAM 1047.2 · Disk 864.1 | 824.5 GB RAM 1236.7 · Disk 903.1 |
| INT88bit | 652.9 GB RAM 979.4 · Disk 838.1 | 698.1 GB RAM 1047.2 · Disk 864.1 | 824.5 GB RAM 1236.7 · Disk 903.1 |
| AWQ4bit | 341.1 GB RAM 511.6 · Disk 453.5 | 372.1 GB RAM 558.2 · Disk 479.5 | 484.3 GB RAM 726.4 · Disk 518.5 |
| GPTQ4bit | 345.1 GB RAM 517.7 · Disk 458.5 | 376.4 GB RAM 564.5 · Disk 484.5 | 488.7 GB RAM 733.1 · Disk 523.5 |
| GGUF4bit | 353.2 GB RAM 529.9 · Disk 468.5 | 384.8 GB RAM 577.2 · Disk 494.5 | 497.5 GB RAM 746.3 · Disk 533.5 |
GPU Recommendations
Best ValueBEST
8× H200 SXM
1128 GB VRAM
$20,440/mo
≈ 57.8 tok/s
Effective production throughput ≈ 40.5 tok/s
score 0.3077
Cheapest
8× RTX A6000
384 GB VRAM
$1,678/mo
≈ 9.2 tok/s
Effective production throughput ≈ 6.4 tok/s
score 0.4893
Performance
4× H200 SXM
564 GB VRAM
$10,220/mo
≈ 43.4 tok/s
Effective production throughput ≈ 30.4 tok/s
score 0.4453
Min Complexity
4× H200 SXM
564 GB VRAM
$10,220/mo
≈ 43.4 tok/s
Effective production throughput ≈ 30.4 tok/s
score 0.4453
💬 Community — real deployment experience
0 Articles · 0 Benchmarks