Decomposing Runtime, Kernel, and Quantization Speedups via a Matched FP16 Intermediate: A Hardware-Conditioned Case Study on Four NVIDIA RTX A5000 GPUs

• URL: http://arxiv.org/abs/2607.11368v1
• Date: Mon, 13 Jul 2026 10:31:54 GMT
• Title: Decomposing Runtime, Kernel, and Quantization Speedups via a Matched FP16 Intermediate: A Hardware-Conditioned Case Study on Four NVIDIA RTX A5000 GPUs
• Abstract summary: Quantization extends concurrent users roughly four times past a reproducible half precision memory cliff.<n>A matched intermediate stack keeps the faster runtime without the quantized kernel splits the full speedup into a runtime part and a kernel and quantization part.<n>Differences in sampling mode and prompt pool between the two stacks are documented as threats to validity.
• Score: 0.614481021961242
• License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/
• Abstract: Reported serving speedups from quantized kernels typically bundle the weight format, the kernel, and the inference runtime into one number. We present an attribution study on four NVIDIA RTX A5000 GPUs, 24 GiB each, on a single host with NVLink-bridged pairs. A matched intermediate stack that keeps the faster runtime without the quantized kernel splits the full speedup into a runtime part and a kernel and quantization part. Under matched greedy decoding the full stack reaches $2.58\times$ end to end, with the runtime change accounting for about two thirds of that gain on a logarithmic scale; across three similar model families the kernel and quantization part moves by at most 1.5%. Sharding one instance across all four cards falls well below doubling: a profiler trace attributes about 80% of the per token shortfall to coordination, and an NVLink versus PCIe control on the same hardware shows similar realized bandwidth on both links, pointing away from link bandwidth as the cause. Whether to run one sharded instance or several independent ones depends on the workload and the model, with the ranking reversing on the larger model: the smaller model splits between sharding and multiple instances by workload, while the larger model favors two paired instances on every workload. Quantization extends sustainable concurrent users roughly four times past a reproducible half precision memory cliff. Differences in sampling mode and prompt pool between the two stacks are documented as threats to validity.

Related papers

• mKernel: Fast Multi-GPU, Multi-Node Fused Kernels [37.86216303017995]
  mKernel is a library of multi-GPU, multi-node fused kernels that overlap computation, intra-node NVLink communication, and inter-node RDMA at tile granularity.<n>It achieves speedups of up to 1.72x on GEMM+AllReduce and $1.88times$ on Ring Attention.
  arXiv  Detail & Related papers  (2026-09-11T22:50:04Z)
• FoldPipe: Bounded Remote Streaming of Native Molecular Shards with Asynchronous Prefetch [0.0]
  FoldPipe is a lightweight Python orchestration layer for PyTorch and PyTorch Geometric data.<n>It retrieves one shard ahead in a background thread while the consumer trains on the current shard, keeping the number of live shard payloads bounded.<n>We evaluate a SchNet energy-and-force workload on MD17 aspirin using 20 paired, order-alternating benchmark passes on a Tesla T4.
  arXiv  Detail & Related papers  (2026-08-27T12:15:16Z)
• Pre-Compiled Pipeline Shards for Distributed LLM Inference on Intel AI PC Fleets [0.0]
  We show that a handful of AIPCs can serve models beyond the capability of any single one.<n>A four-node deployment of Lunar Lake AI PCs on Intel Tiber Cloud serves a single user at interactive speed.
  arXiv  Detail & Related papers  (2026-08-19T17:33:28Z)
• LongLive-2.0: An NVFP4 Parallel Infrastructure for Long Video Generation [84.86814271505109]
  LongLive-2.0 is an NVFP4-based parallel infrastructure throughout the full training and inference workflow of long video generation.<n>For training, we introduce sequence-parallel autoregressive (AR) training, instantiated as Balanced SP.<n>Experiments show up to 2.15x speedup in training, and 1.84x in inference.
  arXiv  Detail & Related papers  (2026-05-18T17:57:03Z)
• COREY: Entropy-Guided Runtime Chunk Scheduling for Selective Scan Kernels [11.316541559874864]
  Prototype scheduler uses activation entropy estimated via fixed-width histograms as a runtime signal for chunk-size selection.<n>COREY is positioned as a Concept and Feasibility contribution.<n>This work does not yet include a fully integrated end-to-end run connecting Tier 2a and Tier 2b.
  arXiv  Detail & Related papers  (2026-04-12T12:07:48Z)
• Scaling DoRA: High-Rank Adaptation via Factored Norms and Fused Kernels [83.99688944263843]
  Weight-De Low-Rank Adaptation (DoRA) extends LoRA by decoupling weight magnitude from direction.<n>At d_in = 8192 and rank r = 384, a single module's norm requires about 512 MB of transient working memory in bf16.<n>A factored norm decomposes the squared norm into base, cross, and Gram terms computable through O(d_out r + r2) intermediates, eliminating the dense product.
  arXiv  Detail & Related papers  (2026-03-23T17:57:24Z)
• ParallelKittens: Systematic and Practical Simplification of Multi-GPU AI Kernels [40.94392896555992]
  Existing systems mitigate this through compute-communication overlap but often fail to meet theoretical bandwidth across workloads and new accelerators.<n>Instead of operator-specific techniques, we ask whether a small set of simple, reusable principles can guide the optimal optimal performance of workloads.<n>PKKittens (PK) kernels achieves up to $2.33 times times parallel workloads.
  arXiv  Detail & Related papers  (2025-11-17T21:48:33Z)
• MARLIN: Mixed-Precision Auto-Regressive Parallel Inference on Large Language Models [58.3342517278868]
  This paper describes the design of Mixed-precision AutoRegressive LINear kernels. It shows that batchsizes up to 16-32 can be supported with close to maximum ($4times$) quantization speedup. MarLIN accomplishes this via a combination of techniques, such as asynchronous memory access, complex task scheduling and pipelining.
  arXiv  Detail & Related papers  (2024-08-21T16:10:41Z)
• AsyncDiff: Parallelizing Diffusion Models by Asynchronous Denoising [49.785626309848276]
  AsyncDiff is a universal and plug-and-play acceleration scheme that enables model parallelism across multiple devices. For the Stable Diffusion v2.1, AsyncDiff achieves a 2.7x speedup with negligible degradation and a 4.0x speedup with only a slight reduction of 0.38 in CLIP Score. Our experiments also demonstrate that AsyncDiff can be readily applied to video diffusion models with encouraging performances.
  arXiv  Detail & Related papers  (2024-06-11T03:09:37Z)
• Accelerating Training and Inference of Graph Neural Networks with Fast Sampling and Pipelining [58.10436813430554]
  Mini-batch training of graph neural networks (GNNs) requires a lot of computation and data movement. We argue in favor of performing mini-batch training with neighborhood sampling in a distributed multi-GPU environment. We present a sequence of improvements to mitigate these bottlenecks, including a performance-engineered neighborhood sampler. We also conduct an empirical analysis that supports the use of sampling for inference, showing that test accuracies are not materially compromised.
  arXiv  Detail & Related papers  (2021-10-16T02:41:35Z)
• Adaptive Elastic Training for Sparse Deep Learning on Heterogeneous Multi-GPU Servers [65.60007071024629]
  We show that Adaptive SGD outperforms four state-of-the-art solutions in time-to-accuracy. We show experimentally that Adaptive SGD outperforms four state-of-the-art solutions in time-to-accuracy.
  arXiv  Detail & Related papers  (2021-10-13T20:58:15Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.