Roofline: an insightful visual performance model for multicore architectures

By Samuel Williams, Andrew Waterman, David A. Patterson. 2009

1,226 citations recorded by the source corpus.

Subfields: Computer Networks and Communications, Hardware and Architecture, ML Systems & Hardware

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Computer Networks and Communications

In Hardware and Architecture

In ML Systems & Hardware

Citations in the corpus

GQA: Training Generalized Multi-Query Transformer Models from Multi-Head Checkpoints (2023, background): “This work is focused on achieving a better tradeoff between decoder quality and inference time through reducing the memory bandwidth overhead (Williams et al., 2009) from loading keys and values.”
Efficiently Scaling Transformer Inference (2022, neutral): “We compare our MFU metric to the commonly used rooﬂine metric (Williams et al., 2009).”
Once for All: Train One Network and Specialize it for Efficient Deployment (2019, positive): “Removing the memory bottleneck results in higher utilization and GOPS/s by 70%-90%, pushing the operation point to the upper-right in the rooﬂine model (Williams et al., 2009), as shown in Figure 13.”
SqueezeNext: Hardware-Aware Neural Network Design (2018, background): “The reason for this is the inefﬁciency of depthwise-separable convolution in terms of hardware performance, which is due to its poor arithmetic intensity (ratio of compute to memory operations) [24].”
MobileNetV4 - Universal Models for the Mobile Ecosystem (2024, neutral): “To this end, we use the Rooﬂine Model [49] which estimates the performance of a given workload and predicts whether it is memory-bottlenecked or compute-bottlenecked.”
Big Little Transformer Decoder (2023, background): “It is worth noting that this non-autoregressive approach would require the same amount of FLOPs as a regressive approach that predicts 𝑦 1 through 𝑦 𝑛 + 1 sequentially; however, it is much faster on hardware due to its token-level parallelism and increased arithmetic intensity [69].”
Triton: an intermediate language and compiler for tiled neural network computations (2019, neutral): “Performance of various implementations of C = AB T vs. Roofline [46] Model (NVIDIA GeForce GTX1070), where A ∈ R 1760 × 1760 and B ∈ R N × 1760 as N modulates arithmetic intensity.”
Designing Effective Sparse Expert Models (2022, background): “The compute-tomemory ratio (operational intensity) can serve as an estimate of the efficiency of different operations (Williams et al., 2009; Shazeer, 2019).”
FINN-R (2018, methodology): “or [%] WFP AFP (Single Precision) W(<8) A(<8) (Reduced Precision) W1 A1 (Binary) Fig. 2. Accuracy over time for ImageNet Classification. We use a roofline model, as introduced by Williams etal. [62], to exemplify the computational benefits of QNNs. Fig.1 shows the peak performance of an AWSF1 ”
AI and Memory Wall (2024, neutral): “This is due to the higher memory operations and lower arithmetic intensity of matrix-vector operations inherent in the auto-regressive inference of GPT (see Figure 3c).”
Atom: Low-bit Quantization for Efficient and Accurate LLM Serving (2023, methodology): “We use the Roofline model (Williams et al., 2009) to evaluate the effect of quantization approaches.”
Interactive Machine Learning via a GPU-accelerated Toolkit (2017, methodology): “BIDMach uses rooﬂine design [36] to optimize computational kernels toward hardware limits.”
Scientiﬁc Benchmarking of Parallel Computing Systems Twelve ways to tell the masses when reporting performance results (2017, background): “It is advisable to limit the set of features to the most significant ones, typical features are memory bandwidth and floating point rates as popularized by the roofline model (k = 2) [60].”