Rangharajan Venkatesan

Charbel Sakr, Steve Dai, Rangharajan Venkatesan et al.

Data clipping is crucial in reducing noise in quantization operations and improving the achievable accuracy of quantization-aware training (QAT). Current practices rely on heuristics to set clipping threshold scalars and cannot be shown to be optimal. We propose Optimally Clipped Tensors And Vectors (OCTAV), a recursive algorithm to determine MSE-optimal clipping scalars. Derived from the fast Newton-Raphson method, OCTAV finds optimal clipping scalars on the fly, for every tensor, at every iteration of the QAT routine. Thus, the QAT algorithm is formulated with provably minimum quantization noise at each step. In addition, we reveal limitations in common gradient estimation techniques in QAT and propose magnitude-aware differentiation as a remedy to further improve accuracy. Experimentally, OCTAV-enabled QAT achieves state-of-the-art accuracy on multiple tasks. These include training-from-scratch and retraining ResNets and MobileNets on ImageNet, and Squad fine-tuning using BERT models, where OCTAV-enabled QAT consistently preserves accuracy at low precision (4-to-6-bits). Our results require no modifications to the baseline training recipe, except for the insertion of quantization operations where appropriate.

Felix Abecassis, Anjulie Agrusa, Dong Ahn et al. · nvidia

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.

Kavya Sreedhar, Jason Clemons, Rangharajan Venkatesan et al.

State-of-the-art deep learning models for computer vision tasks are based on the transformer architecture and often deployed in real-time applications. In this scenario, the resources available for every inference can vary, so it is useful to be able to dynamically adapt execution to trade accuracy for efficiency. To create dynamic models, we leverage the resilience of vision transformers to pruning and switch between different scaled versions of a model. Surprisingly, we find that most FLOPs are generated by convolutions, not attention. These relative FLOP counts are not a good predictor of GPU performance since GPUs have special optimizations for convolutions. Some models are fairly resilient and their model execution can be adapted without retraining, while all models achieve better accuracy with retraining alternative execution paths. These insights mean that we can leverage CNN accelerators and these alternative execution paths to enable efficient and dynamic vision transformer inference. Our analysis shows that leveraging this type of dynamic execution can lead to saving 28\% of energy with a 1.4\% accuracy drop for SegFormer (63 GFLOPs), with no additional training, and 53\% of energy for ResNet-50 (4 GFLOPs) with a 3.3\% accuracy drop by switching between pretrained Once-For-All models.

Coleman Hooper, Charbel Sakr, Ben Keller et al.

Quantization is a powerful tool to improve large language model (LLM) inference efficiency by utilizing more energy-efficient low-precision datapaths and reducing memory footprint. However, accurately quantizing LLM weights and activations to low precision is challenging without degrading model accuracy. We propose fine-grained mixed precision (FGMP) quantization, a post-training mixed-precision quantization hardware-software co-design methodology that maintains accuracy while quantizing the majority of weights and activations to reduced precision. Our work makes the following contributions: 1) We develop a policy that uses the perturbation in each value, weighted by the Fisher information, to select which weight and activation blocks to keep in higher precision. This approach preserves accuracy by identifying which weight and activation blocks need to be retained in higher precision to minimize the perturbation in the model loss. 2) We also propose a sensitivity-weighted clipping approach for fine-grained quantization which helps retain accuracy for blocks that are quantized to low precision. 3) We then propose hardware augmentations to leverage the efficiency benefits of FGMP quantization. Our hardware implementation encompasses i) datapath support for FGMP at block granularity, and ii) a mixed-precision activation quantization unit to assign activation blocks to high or low precision on the fly with minimal runtime and energy overhead. Our design, prototyped using NVFP4 (an FP4 format with microscaling) as the low-precision datatype and FP8 as the high-precision datatype, facilitates efficient FGMP quantization, attaining <1% perplexity degradation on Wikitext-103 for the Llama-2-7B model relative to an all-FP8 baseline design while consuming 14% less energy during inference and requiring 30% less weight memory.

Zichen Fan, Steve Dai, Rangharajan Venkatesan et al.

Diffusion models have gained significant popularity in image generation tasks. However, generating high-quality content remains notably slow because it requires running model inference over many time steps. To accelerate these models, we propose to aggressively quantize both weights and activations, while simultaneously promoting significant activation sparsity. We further observe that the stated sparsity pattern varies among different channels and evolves across time steps. To support this quantization and sparsity scheme, we present a novel diffusion model accelerator featuring a heterogeneous mixed-precision dense-sparse architecture, channel-last address mapping, and a time-step-aware sparsity detector for efficient handling of the sparsity pattern. Our 4-bit quantization technique demonstrates superior generation quality compared to existing 4-bit methods. Our custom accelerator achieves 6.91x speed-up and 51.5% energy reduction compared to traditional dense accelerators.

Akshat Ramachandran, Marina Neseem, Charbel Sakr et al.

The long-output context generation of large reasoning models enables extended chain of thought (CoT) but also drives rapid growth of the key-value (KV) cache, quickly overwhelming GPU memory. To address this challenge, we propose ThinKV, a thought-adaptive KV cache compression framework. ThinKV is based on the observation that attention sparsity reveals distinct thought types with varying importance within the CoT. It applies a hybrid quantization-eviction strategy, assigning token precision by thought importance and progressively evicting tokens from less critical thoughts as reasoning trajectories evolve. Furthermore, to implement ThinKV, we design a kernel that extends PagedAttention to enable efficient reuse of evicted tokens' memory slots, eliminating compaction overheads. Extensive experiments on DeepSeek-R1-Distill, GPT-OSS, and NVIDIA AceReason across mathematics and coding benchmarks show that ThinKV achieves near-lossless accuracy with less than 5% of the original KV cache, while improving performance with up to 5.8x higher inference throughput over state-of-the-art baselines.

Jiawei Zhao, Steve Dai, Rangharajan Venkatesan et al.

Representing deep neural networks (DNNs) in low-precision is a promising approach to enable efficient acceleration and memory reduction. Previous methods that train DNNs in low-precision typically keep a copy of weights in high-precision during the weight updates. Directly training with low-precision weights leads to accuracy degradation due to complex interactions between the low-precision number systems and the learning algorithms. To address this issue, we develop a co-designed low-precision training framework, termed LNS-Madam, in which we jointly design a logarithmic number system (LNS) and a multiplicative weight update algorithm (Madam). We prove that LNS-Madam results in low quantization error during weight updates, leading to stable performance even if the precision is limited. We further propose a hardware design of LNS-Madam that resolves practical challenges in implementing an efficient datapath for LNS computations. Our implementation effectively reduces energy overhead incurred by LNS-to-integer conversion and partial sum accumulation. Experimental results show that LNS-Madam achieves comparable accuracy to full-precision counterparts with only 8 bits on popular computer vision and natural language tasks. Compared to FP32 and FP8, LNS-Madam reduces the energy consumption by over 90% and 55%, respectively.

Steve Dai, Rangharajan Venkatesan, Haoxing Ren et al.

Quantization enables efficient acceleration of deep neural networks by reducing model memory footprint and exploiting low-cost integer math hardware units. Quantization maps floating-point weights and activations in a trained model to low-bitwidth integer values using scale factors. Excessive quantization, reducing precision too aggressively, results in accuracy degradation. When scale factors are shared at a coarse granularity across many dimensions of each tensor, effective precision of individual elements within the tensor are limited. To reduce quantization-related accuracy loss, we propose using a separate scale factor for each small vector of ($\approx$16-64) elements within a single dimension of a tensor. To achieve an efficient hardware implementation, the per-vector scale factors can be implemented with low-bitwidth integers when calibrated using a two-level quantization scheme. We find that per-vector scaling consistently achieves better inference accuracy at low precision compared to conventional scaling techniques for popular neural networks without requiring retraining. We also modify a deep learning accelerator hardware design to study the area and energy overheads of per-vector scaling support. Our evaluation demonstrates that per-vector scaled quantization with 4-bit weights and activations achieves 37% area saving and 24% energy saving while maintaining over 75% accuracy for ResNet50 on ImageNet. 4-bit weights and 8-bit activations achieve near-full-precision accuracy for both BERT-base and BERT-large on SQuAD while reducing area by 26% compared to an 8-bit baseline.

Angshuman Parashar, Minsoo Rhu, Anurag Mukkara et al.

Convolutional Neural Networks (CNNs) have emerged as a fundamental technology for machine learning. High performance and extreme energy efficiency are critical for deployments of CNNs in a wide range of situations, especially mobile platforms such as autonomous vehicles, cameras, and electronic personal assistants. This paper introduces the Sparse CNN (SCNN) accelerator architecture, which improves performance and energy efficiency by exploiting the zero-valued weights that stem from network pruning during training and zero-valued activations that arise from the common ReLU operator applied during inference. Specifically, SCNN employs a novel dataflow that enables maintaining the sparse weights and activations in a compressed encoding, which eliminates unnecessary data transfers and reduces storage requirements. Furthermore, the SCNN dataflow facilitates efficient delivery of those weights and activations to the multiplier array, where they are extensively reused. In addition, the accumulation of multiplication products are performed in a novel accumulator array. Our results show that on contemporary neural networks, SCNN can improve both performance and energy by a factor of 2.7x and 2.3x, respectively, over a comparably provisioned dense CNN accelerator.