We introduce Nemotron 3 Ultra, a 550 billion total and 55 billion active parameter Mixture-of-Experts Hybrid Mamba-Attention language model. We pre-trained Nemotron 3 Ultra on 20 trillion text tokens, then extended the context length to 1M tokens, and post-trained using Supervised Fine Tuning (SFT), Reinforcement Learning (RL), and Multi-teacher On-Policy Distillation (MOPD). Nemotron 3 Ultra is our most capable model yet, employing multiple key technologies - LatentMoE, Multi Token Prediction (MTP), NVFP4 pre-training, multi-environment RLVR, MOPD, and reasoning budget control. Nemotron 3 Ultra achieves up to ~6x higher inference throughput as compared to state-of-the-art publicly available LLMs while attaining on-par accuracy. The state-of-the-art accuracy, high inference throughput, and 1M token context length make Nemotron 3 Ultra ideal for long-running autonomous agentic tasks. We open-source the base, post-trained, and quantized checkpoints, along with the training data and recipe on HuggingFace.

We describe the pre-training, post-training, and quantization of Nemotron 3 Super, a 120 billion (active 12 billion) parameter hybrid Mamba-Attention Mixture-of-Experts model. Nemotron 3 Super is the first model in the Nemotron 3 family to 1) be pre-trained in NVFP4, 2) leverage LatentMoE, a new Mixture-of-Experts architecture that optimizes for both accuracy per FLOP and accuracy per parameter, and 3) include MTP layers for inference acceleration through native speculative decoding. We pre-trained Nemotron 3 Super on 25 trillion tokens followed by post-training using supervised fine tuning (SFT) and reinforcement learning (RL). The final model supports up to 1M context length and achieves comparable accuracy on common benchmarks, while also achieving up to 2.2x and 7.5x higher inference throughput compared to GPT-OSS-120B and Qwen3.5-122B, respectively. Nemotron 3 Super datasets, along with the base, post-trained, and quantized checkpoints, are open-sourced on HuggingFace.

As inference-time scaling becomes critical for enhanced reasoning capabilities, it is increasingly becoming important to build models that are efficient to infer. We introduce Nemotron-H, a family of 8B and 56B/47B hybrid Mamba-Transformer models designed to reduce inference cost for a given accuracy level. To achieve this goal, we replace the majority of self-attention layers in the common Transformer model architecture with Mamba layers that perform constant computation and require constant memory per generated token. We show that Nemotron-H models offer either better or on-par accuracy compared to other similarly-sized state-of-the-art open-sourced Transformer models (e.g., Qwen-2.5-7B/72B and Llama-3.1-8B/70B), while being up to 3× faster at inference. To further increase inference speed and reduce the memory required at inference time, we created Nemotron-H-47B-Base from the 56B model using a new compression via pruning and distillation technique called MiniPuzzle. Nemotron-H-47B-Base achieves similar accuracy to the 56B model, but is 20

Mixture of Experts (MoE) models enhance neural network scalability by dynamically selecting relevant experts per input token, enabling larger model sizes while maintaining manageable computation costs. However, efficient training of large-scale MoE models across thousands of GPUs presents significant challenges due to limitations in existing parallelism strategies. We introduce an end-to-end training framework for large-scale MoE models that utilizes five-dimensional hybrid parallelism: Tensor Parallelism, Expert Parallelism, Context Parallelism, Data Parallelism, and Pipeline Parallelism. Central to our approach is MoE Parallel Folding, a novel strategy that decouples the parallelization of attention and MoE layers in Transformer models, allowing each layer type to adopt optimal parallel configurations. Additionally, we develop a flexible token-level dispatcher that supports both token-dropping and token-dropless MoE training across all five dimensions of parallelism. This dispatcher accommodates dynamic tensor shapes and coordinates different parallelism schemes for Attention and MoE layers, facilitating complex parallelism implementations. Our experiments demonstrate significant improvements in training efficiency and scalability. We achieve up to 49.3

We introduce the Nemotron 3 family of models - Nano, Super, and Ultra. These models deliver strong agentic, reasoning, and conversational capabilities. The Nemotron 3 family uses a Mixture-of-Experts hybrid Mamba-Transformer architecture to provide best-in-class throughput and context lengths of up to 1M tokens. Super and Ultra models are trained with NVFP4 and incorporate LatentMoE, a novel approach that improves model quality. The two larger models also include MTP layers for faster text generation. All Nemotron 3 models are post-trained using multi-environment reinforcement learning enabling reasoning, multi-step tool use, and support granular reasoning budget control. Nano, the smallest model, outperforms comparable models in accuracy while remaining extremely cost-efficient for inference. Super is optimized for collaborative agents and high-volume workloads such as IT ticket automation. Ultra, the largest model, provides state-of-the-art accuracy and reasoning performance. Nano is released together with its technical report and this white paper, while Super and Ultra will follow in the coming months. We will openly release the model weights, pre- and post-training software, recipes, and all data for which we hold redistribution rights.

Large Language Models (LLMs) today are powerful problem solvers across many domains, and they continue to get stronger as they scale in model size, training set size, and training set quality, as shown by extensive research and experimentation across the industry. Training a frontier model today requires on the order of tens to hundreds of yottaflops, which is a massive investment of time, compute, and energy. Improving pretraining efficiency is therefore essential to enable the next generation of even more capable LLMs. While 8-bit floating point (FP8) training is now widely adopted, transitioning to even narrower precision, such as 4-bit floating point (FP4), could unlock additional improvements in computational speed and resource utilization. However, quantization at this level poses challenges to training stability, convergence, and implementation, notably for large-scale models trained on long token horizons. In this study, we introduce a novel approach for stable and accurate training of large language models (LLMs) using the NVFP4 format. Our method integrates Random Hadamard transforms (RHT) to bound block-level outliers, employs a two-dimensional quantization scheme for consistent representations across both the forward and backward passes, utilizes stochastic rounding for unbiased gradient estimation, and incorporates selective high-precision layers. We validate our approach by training a 12-billion-parameter model on 10 trillion tokens – the longest publicly documented training run in 4-bit precision to date. Our results show that the model trained with our NVFP4-based pretraining technique achieves training loss and downstream task accuracies comparable to an FP8 baseline. These findings highlight that NVFP4, when combined with our training approach, represents a major step forward in narrow-precision LLM training algorithms.

Training large transformer models is one of the most important computational challenges of modern AI. In this paper, we show how to significantly accelerate training of large transformer models by reducing activation recomputation. Activation recomputation is commonly used to work around memory capacity constraints. Rather than storing activations for backpropagation, they are traditionally recomputed, which saves memory but adds redundant compute. In this work, we show most of this redundant compute is unnecessary because we can reduce memory consumption sufficiently without it. We present two novel yet very simple techniques: sequence parallelism and selective activation recomputation. In conjunction with tensor parallelism, these techniques almost eliminate the need to recompute activations. We evaluate our approach on language models up to one trillion parameters in scale and show that our method reduces activation memory by 5x, while reducing execution time overhead from activation recomputation by over 90%. For example, when training a 530B parameter GPT-3 style model on 2240 NVIDIA A100 GPUs, we achieve a Model Flops Utilization of 54.2%, which is 29% faster than the 42.1% we achieve using recomputation. Our implementation will be available in both Megatron-LM and NeMo-Megatron.

Temporal SIMT (TSIMT) has been suggested as an alternative to conventional (spatial) SIMT for improving GPU performance on branch-intensive code. Although TSIMT has been briefly mentioned before, it was not evaluated. We present a complete design and evaluation of TSIMT GPUs, along with the inclusion of scalarization and a combination of temporal and spatial SIMT, named Spatiotemporal SIMT (STSIMT). Simulations show that TSIMT alone results in a performance reduction, but a combination of scalarization and STSIMT yields a mean performance enhancement of 19.6% and improves the energy-delay product by 26.2% compared to SIMT.

Modern GPUs provide massive processing power (arithmetic throughput) as well as memory throughput. Presently, while it appears to be well understood how performance can be improved by increasing throughput, it is less clear what the effects of micro-architectural latencies are on the performance of throughput-oriented GPU architectures. In fact, little is publicly known about the values, behavior, and performance impact of microarchitecture latency components in modern GPUs. This work attempts to fill that gap by analyzing both the idle (static) as well as loaded (dynamic) latency behavior of GPU microarchitectural components. Our results show that GPUs are not as effective in latency hiding as commonly thought and based on that, we argue that latency should also be a GPU design consideration besides throughput.

GPUs are much more power-efficient devices compared to CPUs, but due to several performance bottlenecks, the performance per watt of GPUs is often much lower than what could be achieved theoretically. To sustain and continue high performance computing growth, new architectural and application techniques are required to create power-efficient computing systems. To find such techniques, however, it is necessary to study the power consumption at a detailed level and understand the bottlenecks which cause low performance. Therefore, in this paper, we study GPU power consumption at component level and investigate the bottlenecks that cause low performance and low energy efficiency. We divide the low performance kernels into low occupancy and full occupancy categories. For the low occupancy category, we study if increasing the occupancy helps in increasing performance and energy efficiency. For the full occupancy category, we investigate if these kernels are limited by memory bandwidth, coalescing efficiency, or SIMD utilization.

DRAM memory stores its contents in leaky cells that require periodic refresh to prevent data loss. The refresh operation does not only degrade system performance, but also consumes significant amounts of energy in mobile systems. Relaxed DRAM refresh has been proposed as one possible building block of approximate computing. Multiple authors have suggested techniques where programmers can specify which data is critical and can not tolerate any bit errors and which data can be stored approximately. However, in these approaches all bits in the approximate area are treated as equally important. We show that this produces suboptimal results and higher energy savings or better quality can be achieved, if a more fine-grained approach is used. Our proposal is able to save more refresh power and enables a more effective storage of non-critical data by utilizing a non-uniform refresh of multiple DRAM chips and a permutation of the bits to the DRAM chips. In our proposal bits of high importance are stored in a high quality storage bits and bits of low importance are stored in low quality storage bits. The proposed technique works with commodity DRAMs.

Modern GPUs are true power houses in every meaning of the word: While they offer general-purpose (GPGPU) compute performance an order of magnitude higher than that of conventional CPUs, they have also been rapidly approaching the infamous “power wall”, as a single chip sometimes consumes more than 300W. Thus, the design space of GPGPU microarchitecture has been extended by another dimension: power. While GPU researchers have previously relied on cycle-accurate simulators for estimating performance during design cycles, there are no simulation tools that include power as well. To mitigate this issue, we introduce the GPUSimPow power estimation framework for GPGPUs consisting of both analytical and empirical models for regular and irregular hardware components. To validate this framework, we build a custom measurement setup to obtain power numbers from real graphics cards. An evaluation on a set of well-known benchmarks reveals an average relative error of 11.7% between simulated and hardware power for GT240 and an average relative error of 10.8% for GTX580. The simulator has been made available to the public [1].

In this paper, we evaluate the performance and usability of the parallel programming model OpenMP Superscalar (OmpSs), apply it to 10 different benchmarks and compare its performance with corresponding POSIX threads implementations.

In the past years, research and industry have introduced several parallel programming models to simplify the development of parallel applications. A popular class among these models are task-based programming models which proclaim ease-of-use, portability, and high performance. A novel model in this class, OpenMP Superscalar, combines advanced features such as automated runtime dependency resolution, while maintaining simple pragma-based programming for C/C++. OpenMP Superscalar has proven to be effective in leveraging parallelism in HPC workloads. Embedded and consumer applications, however, are currently still mainly parallelized using traditional thread-based programming models. In this work, we investigate how effective OpenMP Superscalar is for embedded and consumer applications in terms of usability and performance. To determine the usability of OmpSs, we show in detail how to implement complex parallelization strategies such as ones used in parallel H.264 decoding. To evaluate the performance we created a collection of ten embedded and consumer benchmarks parallelized in both OmpSs and Pthreads.