Can I Run Nemotron 70B on 48 GB system RAM?

Superseded model. Nemotron 70B has been superseded by Nemotron 3 Super. This page is kept for reference; the newer family is a better starting point. View Nemotron 3 Super →

Written by Jakub Rusinowski · Last updated October 15, 2024

Technically yes, but not recommended

It loads, but it is not worth running — Nemotron 70B Instruct at Q3_K_M fits in 48 GB system RAM's 38.4 GB, yet the memory bandwidth limits it to ~2.1 tok/s (estimated), well below usable interactive speed.

Confidence: high · Recommended quantization: Q3_K_M · Estimated speed: ~2.1 tok/s

48 GB system RAM — what it gives a model

Usable memory for models	38.4 GB
Memory bandwidth	90 GB/s

Nemotron 70B on 48 GB system RAM: memory by quantization

Quant	Memory needed	Fits 38.4 GB?	Max context	Est. speed	Download
F16	144.7 GB	✗ No	—	—	141.2 GB
Q8_0	78.5 GB	✗ No	—	—	75 GB
Q6_K	61.4 GB	✗ No	—	—	57.9 GB
Q5_K_M	53.5 GB	✗ No	—	—	50 GB
Q4_K_M	46.1 GB	✗ No	—	—	42.6 GB
Q3_K_M	33.6 GB	✓ Yes	16K	~2.1 tok/s	30.1 GB
Q2_K	26.7 GB	✓ Yes	32K	~2.7 tok/s	23.2 GB

What to watch out for

At ~2.1 tok/s this loads but is too slow for interactive use — expect roughly 29 seconds per 60 tokens.
Q3_K_M is the only quantization that fits, and it is a heavily degraded one — expect noticeably worse output than the same model at Q4_K_M. A smaller model at Q4 is usually the better trade.
1 larger variant of Nemotron 70B does not fit and would need CPU offload or different hardware.
These figures assume CPU-only inference. Any discrete GPU, even an 8 GB one, will be several times faster for models that fit in its VRAM.

Recommended setup

llama.cpp (CPU build) or Ollama — both run without a GPU

How these numbers are calculated

Memory need = quantized weights + KV cache at 8,192 tokens (f16) + 0.8 GB runtime overhead.
38.4 GB of the 48 GB is treated as usable for model weights (80% — the rest is the OS and running applications).
DDR5-5600 dual channel at 89.6 GB/s peak. CPU decode is assumed to sustain 35% of that peak, because CPU inference is not purely bandwidth-bound — it also spends real time in compute and thread synchronisation. This figure is an assumption, not a fitted constant: no CPU measurement is in the calibration set.
CPU-only inference: no GPU is assumed. A GPU of any size will beat these figures substantially.
Throughput is a memory-bandwidth roofline estimate, not a measurement. It is labelled "estimated" everywhere it appears.
KV cache is computed from this model's published attention configuration.

FAQ