cuDNN: Efficient Primitives for Deep Learning
read the original abstract
We present a library of efficient implementations of deep learning primitives. Deep learning workloads are computationally intensive, and optimizing their kernels is difficult and time-consuming. As parallel architectures evolve, kernels must be reoptimized, which makes maintaining codebases difficult over time. Similar issues have long been addressed in the HPC community by libraries such as the Basic Linear Algebra Subroutines (BLAS). However, there is no analogous library for deep learning. Without such a library, researchers implementing deep learning workloads on parallel processors must create and optimize their own implementations of the main computational kernels, and this work must be repeated as new parallel processors emerge. To address this problem, we have created a library similar in intent to BLAS, with optimized routines for deep learning workloads. Our implementation contains routines for GPUs, although similarly to the BLAS library, these routines could be implemented for other platforms. The library is easy to integrate into existing frameworks, and provides optimized performance and memory usage. For example, integrating cuDNN into Caffe, a popular framework for convolutional networks, improves performance by 36% on a standard model while also reducing memory consumption.
This paper has not been read by Pith yet.
Forward citations
Cited by 18 Pith papers
-
Prism: Symbolic Superoptimization of Tensor Programs
Prism is the first symbolic superoptimizer for tensor programs that uses sGraph for compact representation of program families, two-level search, e-graph equivalence checking, and auto-tuning to achieve up to 2.2x spe...
-
Learning 1-Bit LiDAR-based Localization with Auxiliary Objective
BiLoc is the first binary neural network framework for 6-DoF LiDAR pose estimation that uses an auxiliary objective to adaptively regulate information retention and achieve SOTA among BNNs on large outdoor datasets.
-
Toward Compiler World Models: Learning Latent Dynamics for Efficient Tensor Program Search
A latent dynamics model for schedule trajectories in TVM AutoScheduler finds programs with 1.37x better GPU latency than Ansor using the same 64 trials and matches 10K-trial Ansor with 10x fewer measurements.
-
ATLAAS: Automatic Tensor-Level Abstraction of Accelerator Semantics
ATLAAS automatically converts RTL-extracted bit-level accelerator semantics into tensor-level ISA specs via an 8-pass MLIR pipeline, enabling automated compiler backend generation for designs like Gemmini and VTA.
-
PyTorch: An Imperative Style, High-Performance Deep Learning Library
PyTorch demonstrates compatibility of imperative Pythonic usability with high performance and accelerator support through its runtime architecture and benchmark results.
-
The Indirect Convolution Algorithm
The Indirect Convolution algorithm avoids im2col by using an indirection buffer, reducing memory overhead proportionally to input channels and outperforming GEMM-based methods by up to 62% for convolutions requiring t...
-
CuLifter: Lifting GPU Binaries to Typed IR
CuLifter recovers types from untyped GPU register files via constraint propagation to lift 99.98% of 24,437 functions across 919 cubins to valid LLVM IR.
-
Evaluating CUDA Tile for AI Workloads on Hopper and Blackwell GPUs
CuTile delivers high performance on select AI workloads and GPUs but varies significantly by architecture and is less portable than Triton across tested platforms.
-
Evaluating CUDA Tile for AI Workloads on Hopper and Blackwell GPUs
CuTile achieves up to 2.5x FlashAttention-2 throughput on B200 with 60 lines of Python but shows significant cross-architecture portability gaps, reaching only 53% of FlashAttention-2 on RTX PRO 6000.
-
TCL: Enabling Fast and Efficient Cross-Hardware Tensor Program Optimization via Continual Learning
TCL delivers 16.8x faster tuning on CPU and 12.48x on GPU with modestly lower inference latency by combining RDU active sampling, a lightweight Mamba cost model, and cross-platform continual knowledge distillation.
-
TileLoom: Automatic Dataflow Planning for Tile-Based Languages on Spatial Dataflow Accelerators
TileLoom compiles tile-based languages to spatial dataflow hardware by distributing tiles across cores and optimizing data reuse via on-chip networks, delivering vendor-comparable performance on Tenstorrent systems.
-
ShinkaEvolve: Towards Open-Ended And Sample-Efficient Program Evolution
ShinkaEvolve improves sample efficiency in LLM-driven program evolution via parent sampling, code novelty rejection-sampling, and bandit LLM ensemble selection, achieving new SOTA circle packing with 150 samples and g...
-
A Unified Optimization Approach for CNN Model Inference on Integrated GPUs
A unified IR plus ML-based scheduling for CNN inference on multi-vendor integrated GPUs matches or exceeds vendor libraries (up to 1.62x) on image models while supporting more models.
-
Sakana Fugu Technical Report
Sakana Fugu trains LLM orchestrators using fine-tuning, evolutionary algorithms, and RL to build query-adaptive multi-agent scaffolds, claiming SOTA results on benchmarks including SWE-Bench Pro and GPQA-Diamond.
-
MOSAIC: Efficient Mixture-of-Agent Scheduling via Adaptive Aggregation and Inference Concurrency
MOSAIC uses an Integer Linear Program scheduler for expert placement and prompt assignment plus adaptive aggregation to achieve 1.7-2.3x end-to-end speedup on 4-GPU MoA workloads while keeping accuracy within 0.1pp.
-
CUDA Kernel Optimization and Counter-Free Performance Analysis for Depthwise Convolution in Cloud Environments
Warp-tiled CUDA kernel for depthwise convolution delivers 3.26x runtime reduction versus naive baseline and 1.29x end-to-end training speedup using counter-free analysis in cloud settings.
-
Evaluating Cross-Architecture Performance Modeling of Distributed ML Workloads Using StableHLO
StableHLO serves as a viable unified representation for cross-architecture performance modeling of distributed ML workloads, preserving relative trends while exposing fidelity trade-offs.
-
GPU Parallelization Strategies for Forward and Backward Propagation in Shallow Neural Networks: A CUDA-Based Comparative Study
Empirical comparison shows stacked shared-memory tiling, weight pre-transposition, and kernel fusion yield 1.41x speedup for shallow NN propagation on Tesla T4 versus baseline CUDA.
discussion (0)
Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.