Skip to content
Updated minutes ago
NVIDIA

Nemotron Ultra 253B

NVIDIA · dense · 253B parameters · 131,072 context

Quality
86.0

Parameters

253B

Context Window

128K tokens

Architecture

Dense

Best GPU

B200 NVL (pair)

Cheapest API

$6.00/M

Quality Score

86/100

Intelligence Brief

Nemotron Ultra 253B is a 253B parameter DENSE model from NVIDIA, featuring Grouped Query Attention (GQA) with 96 layers and 12,288 hidden dimensions. With a 131,072 token context window, it supports tools, structured output, code, math, multilingual, reasoning. On standardized benchmarks, it achieves MMLU 90, HumanEval 76, GSM8K 94. The most cost-effective API deployment is via nvidia at $6.00/M output tokens. For self-hosted inference, B200 NVL (pair) delivers optimal throughput at $9965/month.

Architecture Details

TypeDENSE
Total Parameters253B
Active Parameters253B
Layers96
Hidden Dimension12,288
Attention Heads96
KV Heads8
Head Dimension128
Vocab Size128,256

Memory Requirements

BF16 Weights

506.0 GB

FP8 Weights

253.0 GB

INT4 Weights

126.5 GB

KV-Cache per Token393216 bytes
Activation Estimate5.00 GB

Fits on (multi-GPU with Tensor Parallelism)

Multi-GPU configurations use Tensor Parallelism (TP) to split model layers across GPUs. Requires NVLink or NVSwitch interconnect for optimal performance.

GPU Compatibility Matrix

Nemotron Ultra 253B is compatible with 8% of GPU configurations across 41 GPUs at 3 precision levels.

BF16 (Full)
FP8 (Half)
INT4 (Quarter)
Blackwell(7 GPUs)
B200 NVL (pair)360GB
B300288GB
B100 SXM192GB
GB200 NVL72 (per GPU)192GB
Hopper(7 GPUs)
H100 NVL 94GB (per GPU pair)188GB
H200 SXM141GB
H2096GB
GH20096GB
Ada Lovelace(11 GPUs)
L40S48GB
L4048GB
RTX 6000 Ada48GB
L2048GB
Ampere(16 GPUs)
A100 80GB SXM80GB
A100 80GB PCIe80GB
A1664GB
RTX A600048GB
Legend:No fitVery tightTightModerateGoodExcellent

GPU Recommendations

B200 NVL (pair)optimal

FP8 · 1 GPU · tensorrt-llm

93/100

score

Throughput

280.0 tok/s

Latency (ITL)

3.6ms

Est. TTFT

1ms

Cost/Month

$9965

Cost/M Tokens

$13.54

Use this config →
H20optimal

FP8 · 4 GPUs · tensorrt-llm

90/100

score

Throughput

280.0 tok/s

Latency (ITL)

3.6ms

Est. TTFT

1ms

Cost/Month

$3758

Cost/M Tokens

$5.11

Use this config →
B200 SXMoptimal

FP8 · 2 GPUs · tensorrt-llm

88/100

score

Throughput

280.0 tok/s

Latency (ITL)

3.6ms

Est. TTFT

1ms

Cost/Month

$8522

Cost/M Tokens

$11.58

Use this config →

Deployment Options

API

API Deployment

nvidia

$6.00/M

output tokens

Self-Hosted

Single GPU

B200 NVL (pair)

$9965/mo

Min VRAM: 253 GB

Scale

Multi-GPU

H20 x4

280.0 tok/s

TP· $3758/mo

API Pricing Comparison

ProviderInput $/MOutput $/MBadges
nvidia$2.00$6.00
Cheapest

Cost Analysis

ProviderInput $/MOutput $/M~Monthly Cost
nvidiaBest Value$2.00$6.00$40

Cost per 1,000 Requests

Short (500 tok)

$2.20

via nvidia

Medium (2K tok)

$8.80

via nvidia

Long (8K tok)

$28.00

via nvidia

Performance Estimates

Throughput by GPU

B200 NVL (pair)
280.0 tok/s
H20
280.0 tok/s
B200 SXM
280.0 tok/s

VRAM Breakdown (B200 NVL (pair), FP8)

Weights
Weights 253.0 GBKV-Cache 3.2 GBActivations 40.0 GBOverhead 12.7 GB

Precision Impact

bf16

506.0 GB

weights/GPU

fp8

253.0 GB

weights/GPU

~280.0 tok/s

int4

126.5 GB

weights/GPU

Quality Benchmarks

Top 10%
95th percentile across all models
MMLU
90.0
Top 10% (92th pctile)
HumanEval
76.0
Top 10% (94th pctile)
GSM8K
94.0
Above Average (84th pctile)

Capabilities

Features

Tool Use Vision Code Math Reasoning Multilingual Structured Output

Supported Frameworks

vllmtensorrt-llm

Supported Precisions

BF16 (default)FP8INT4

Where to Deploy Nemotron Ultra 253B

Similar Models

Nemotron 340B

340B params · dense

Quality: 85

from $4.20/M

Similar specsCompare →
Cheaper, Smaller modelCompare →
Lower quality, CheaperCompare →

DeepSeek V2.5

236B params · moe

Quality: 78

from $0.28/M

Lower quality, CheaperCompare →

Qwen 3 235B

235B params · moe

Quality: 83

from $3.00/M

Frequently Asked Questions

How much VRAM does Nemotron Ultra 253B need for inference?

Nemotron Ultra 253B requires approximately 506.0 GB of VRAM at BF16 precision, 253.0 GB at FP8, or 126.5 GB at INT4 quantization. Additional VRAM is needed for KV-cache (393216 bytes per token) and activations (~5.00 GB).

What is the best GPU for Nemotron Ultra 253B?

The top recommended GPU for Nemotron Ultra 253B is the B200 NVL (pair) using FP8 precision. It achieves approximately 280.0 tokens/sec at an estimated cost of $9965/month ($13.54/M tokens). Score: 93/100.

How much does Nemotron Ultra 253B inference cost?

Nemotron Ultra 253B API inference starts from $2.00/M input tokens and $6.00/M output tokens. Self-hosted inference costs depend on your GPU configuration — use our ROI calculator for a detailed breakdown.