Running Nemotron 3 Ultra on B200 SXM

Nemotron 3 Ultra has 550 billion parameters. B200 SXM offers 1,440 GB of HBM3e across 8 GPUs memory. This page puts the two side by side.

By QDNA · Published on 27 July 2026 · Updated on 15 September 2026

Short answer. Nemotron 3 Ultra fits comfortably on B200 SXM. In FP16, its weights take about 1,265 GB of the 1,440 GB available, leaving 175 GB for the attention cache and concurrency.

How much memory does Nemotron 3 Ultra need?

Weights take the parameter count multiplied by the format size, plus a 15% runtime margin for activations and buffers. In FP16 the calculation reads: 550 billion parameters × 2 bytes (FP16) = 1,100.0 GB of weights; × 1.15 runtime margin = 1,265.0 GB. The attention cache sits on top: it grows with context length and with the number of concurrent requests, so it is sized case by case.

Format	Weights in memory	On this machine	What the format costs
FP16	1,265 GB	tight	full precision, the quality reference
FP8	632 GB	fits	negligible loss on most tasks
NVFP4	316 GB	fits	Blackwell format, native FP4 compute
Q4	379 GB	fits	block quantisation, broadly supported

Model sheet: 550 billion parameters, 55 billion active per token, 262,144-token context, OpenMDW 1.1 licence. 550 billion announced; the BF16 repository holds 560 billion elements (1,121 GB). The card claims “up to 1M” context, config.json declares 262,144.

Nemotron 3 Ultra activates 55 billion parameters per token out of 550 billion. Active parameters govern speed, not footprint: every expert stays resident in memory. Confusing the two under-sizes the machine by an order of magnitude.

What is left to serve requests?

Nemotron 3 Ultra fits comfortably on B200 SXM. In FP16, its weights take about 1,265 GB of the 1,440 GB available, leaving 175 GB for the attention cache and concurrency. The remaining margin decides how many concurrent requests and how much context the machine sustains. By QDNA convention a machine “fits” when the weights take at most 75% of its memory; beyond that it “barely fits” and forces a limit on context or on concurrency. The attention cache is not calculated on this page. These models’ architectures (compressed latent, linear or sparse attention, Mamba layers) share no per-token formula; it is measured on the machine, with the target context and concurrency [TO BE MEASURED].

B200 SXM: price on quotation. NVIDIA publishes no price; Aivres quoted about $340,000 excl. VAT for an HGX B200 server (Arc Compute, updated 30 July 2026, flagged as outdated by the author). The machine targets the GE segment.

Where else can Nemotron 3 Ultra run?

See the Nemotron 3 Ultra and B200 SXM fact sheets.

How much memory does Nemotron 3 Ultra need?

In FP16, weights take about 1,265 GB including the runtime margin. The attention cache sits on top and depends on context.

Does Nemotron 3 Ultra fit on B200 SXM?

Nemotron 3 Ultra fits comfortably on B200 SXM. In FP16, its weights take about 1,265 GB of the 1,440 GB available, leaving 175 GB for the attention cache and concurrency.

Method and sources

Memory footprints are calculated, not measured: a reading on real hardware may differ depending on the engine and the exact weight format. Every figure below carries its source and the date of its last check.

Nemotron 3 Ultra: 550 billion parameters, 55 billion active, 262,144-token context, OpenMDW 1.1 licence. Hugging Face repository (config.json, API, model card), checked on 2 September 2026. 550 billion announced; the BF16 repository holds 560 billion elements (1,121 GB). The card claims “up to 1M” context, config.json declares 262,144.
B200 SXM: 1,440 GB of memory; HGX B200, 8 Blackwell GPUs, 1.4 TB total memory, 144 PFLOPS FP4 with sparsity (72 dense), 1.8 TB/s NVLink per GPU. Manufacturer sheet checked on 2 September 2026.
B200 SXM price: on quotation. NVIDIA publishes no price; Aivres quoted about $340,000 excl. VAT for an HGX B200 server (Arc Compute, updated 30 July 2026, flagged as outdated by the author).
Bytes per parameter: FP16 2 (16 bits per parameter, hence 2 bytes); FP8 1 (8 bits per parameter (E4M3), hence 1 byte, block scales not counted); NVFP4 0.5 (4 bits per parameter, hence 0.5 byte; 4-bit published weights weigh 0.54 to 0.56 byte per parameter with their scales (DeepSeek V4 Pro 865 GB for 1,599 billion, Kimi K3 1,561 GB for 2,780 billion)); Q4 0.6 (taken here as llama.cpp Q4_K_M, 0.6 byte per parameter including scales; Q4_K_S weighs 0.56, AWQ and GPTQ 0.55). Source: connaissance/faits.yaml, quantifications family, checked on 2026-08-31.
Runtime margin × 1.15: QDNA operating assumption, not measured: 15% above the weights for activations, buffers and fragmentation.
“Fits” threshold at 75% of memory: QDNA assumption, not measured, which keeps the remaining quarter for the attention cache and concurrency.